An intelligent recommendation and feedback evolution system and method based on agent skills

Abstract

The application discloses an intelligent recommendation and feedback evolution system and method based on agent skills, and belongs to the technical fields of artificial intelligence, agent execution framework and the like.The application first collects open-source agent skill description files in large scale, completes format verification, meta-information completion and deduplication, and builds a selected skill set;combined with rule extraction and large language model analysis, a structured skill portrait covering environment dependence, quality and verifiability evaluation is generated.For the verifiable skills, verification tasks are reversely constructed, and operation verification is completed relying on a sandbox environment and a real agent.The system adopts a two-stage mode of vector retrieval recall plus agent fine arrangement to realize accurate recommendation, and encapsulates relevant capabilities as general skills to reduce the access threshold of clients.Meanwhile, structured feedback data is collected and stored taking a conversation identifier as a carrier, so as to promote offline iterative optimization of the skill library.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, software engineering, agent execution framework, information retrieval and task evaluation, and in particular to an intelligent recommendation and feedback evolution system and method based on agent skills. Background Technology

[0002] With the rapid iteration and practical application of artificial intelligence technology, intelligent agent products are gradually moving from the laboratory to actual production and daily office scenarios. Client-side intelligent agents, such as code-based intelligent agents, terminal intelligent agents, and office intelligent agents, are becoming increasingly widespread, serving as core carriers for improving work efficiency and automating processes. To achieve capability reuse and standardized expansion, mainstream client-side intelligent agents generally adopt a skill-based encapsulation mechanism. Using standardized skill description files as the core, repetitive workflows, execution constraints, operation templates, and automation logic are condensed into reusable skill units. When performing tasks, the intelligent agent can load the corresponding skills as needed, eliminating the need for repetitive development and inference calculations, significantly reducing execution costs and improving task completion efficiency and stability. With the rapid development of the open-source community, an intelligent agent skill ecosystem is gradually taking shape. Numerous developers and organizations are releasing various open-source skill resources, resulting in an exponential increase in the number of skills. Open standards and sharing mechanisms are gradually improving, providing abundant resource support for the expansion of intelligent agent capabilities.

[0003] Despite the open standards and vast amount of open-source content in the intelligent agent skill ecosystem, a significant technical gap exists between skill resources and actual agent execution in practical engineering applications. Existing technical solutions cannot achieve efficient utilization of skill resources and stable improvement of agent capabilities. Firstly, open-source skills come from diverse sources and vary in quality. Many skills only conform to specifications in terms of filenames, exhibiting problems such as invalid formats, incomplete content, mixed topics, missing dependencies, and providing only knowledge materials without execution guidance. This results in highly noisy candidate sets, which, if used directly for recommendations, can lead to agents selecting incorrect skills and wasting inference budgets. Secondly, existing skills are only described in text form, failing to structurally extract key information such as runtime environment dependencies and execution constraints. This fails to clearly define execution prerequisites such as operating system compatibility, permission requirements, network conditions, and external dependencies, leading to frequent issues of semantically matching skills but incompatible environments and execution failures. Thirdly, existing skill retrieval primarily relies on static keyword searches and vector searches, catering only to human users' browsing or manual selection, without incorporating runtime context information such as installed tools, the current shell environment, and working directory permissions. The existing technology suffers from several drawbacks. First, it lacks a dynamic adaptation mechanism for accurate recommendations, as it is limited by factors such as network type, environmental variables, and sub-task stages. Second, it only analyzes the text of skill documents, lacking a real-world agent verification process. This prevents the acquisition of objective indicators such as success rate, resource consumption, and failure modes of skills in a real sandbox environment. Skill availability relies solely on author descriptions, compromising authenticity and stability. Third, it lacks a feedback loop from agent operation trajectory to the skill library, failing to systematically record feedback signals such as skill usage scenarios, dependency gaps, permission conflicts, and execution results. Consequently, the skill library cannot be continuously optimized, and recommendation quality cannot be iteratively improved with usage. Fourth, the existing skill recommendation system has a high barrier to entry, requiring client agents to modify kernel inference logic or platform backend interfaces. This hinders low-cost reuse across products and platforms, limiting the large-scale promotion of the technology.

[0004] Current technologies can be broadly categorized into three types. The first type is the agent's native skill loading mechanism, which enables the loading and invocation of individual skills, addressing the issue of using single skills. However, it cannot handle the screening, environment modeling, comprehensive recommendation, and feedback feedback of massive amounts of open-source skills. The second type consists of user-oriented retrieval platforms such as skill directories and skill markets. These platforms utilize keyword search, tag filtering, and starred sorting to facilitate manual skill browsing. However, their target audience is human users, not running agents, and their interaction is statically filtered, deeply decoupled from the agent's reasoning process, thus failing to achieve accurate runtime recommendations and environment adaptation. The third type is agent task evaluation frameworks, used to assess the agent's overall task completion capabilities in a controllable container environment. However, these frameworks do not construct verification tasks from the skill library, making it impossible to apply skill verification results to recommendation optimization. The combination of these three types of existing technologies still cannot form a complete technical closed loop encompassing large-scale open-source skill collection, format verification, deduplication, environment parsing, quality assessment, verifiability determination, task construction, real-world verification, runtime recommendation, feedback extraction, and skill library evolution. This makes it difficult to meet the actual needs of client-side agents for efficient, stable, and accurate use of skill resources. Therefore, those skilled in the art urgently need an intelligent recommendation and feedback evolution system that can realize full-process skill processing, real-world operation verification, dynamic and accurate recommendation, and structured feedback closed loop to solve the core technical defects of the existing skill ecosystem. Summary of the Invention

[0005] The technical problem this invention aims to solve is to provide a complete technical solution covering large-scale collection of agent skills, standardized preprocessing, structured profiling, real-world verification, accurate runtime recommendation, structured feedback collection, and offline evolution of the skill library. This solution addresses the existing technical problems of high noise in the candidate set of agent skill ecosystems, poor environmental adaptability, static retrieval detached from the runtime context, lack of real-world verification, no feedback loop, and high client access costs. It provides full-process technical support for skills to be collectable, verifiable, profilable, recommendable, and evolve with feedback, thereby improving the accuracy of agent skill recommendations, execution success rate, and system versatility.

[0006] To achieve the above objectives, the present invention provides an intelligent recommendation and feedback evolution system based on agent skills, comprising:

[0007] The skill acquisition and preprocessing module is used to retrieve agent skill description files from open-source code hosting platforms on a large scale, verify the legality of skill execution formats, complete metadata and remove duplicates in the recommendation side, and generate a curated set of skills.

[0008] The skill profile construction module, connected to the skill acquisition and preprocessing module, is used to generate a structured skill profile for each skill, including environmental dependency analysis, quality assessment, and verifiability assessment, through rule extraction and large language model analysis.

[0009] The task construction and verification module, connected to the skill profile construction module, is used to reverse construct executable verification tasks for skills that have passed the verifiability assessment, run the tasks in a sandbox environment using a combination of real intelligent agents and models, obtain the verification results of skill execution success rate, resource consumption and failure mode, and write them back to the skill profile.

[0010] The recommendation service module, connected to the skill acquisition and preprocessing module, skill profile construction module, and task construction and verification module, is used to perform vector retrieval to recall candidate skills based on the selected skill set, reorganize the candidate skills into a directory structure according to internal classification, perform intelligent agent-style fine ranking through terminal intelligent agent, and output the final recommended skill list and usage guide.

[0011] The client access module, connected to the recommendation service module, is used to encapsulate recommendation and feedback capabilities into general skills that can be installed on the client, enabling recommendation request initiation, skill download, and feedback submission in the client agent reasoning process;

[0012] The feedback service module, connected to the client access module and the skill profile construction module, is used to receive structured feedback data uploaded by the client, associate recommended sessions with feedback data using session identifiers, and store feedback data to form a feedback library, providing data support for skill profile updates and skill library iterations.

[0013] Preferably, the skill acquisition and preprocessing module is specifically used for:

[0014] The system retrieves agent skill description files through the open-source code hosting platform's search interface, and simultaneously collects the corresponding warehouse owner, warehouse name, number of stars, number of forks, skill relative path, and associated resources.

[0015] The official skill verification script is used to perform format validity checks, and only valid skills are retained for subsequent processes.

[0016] All valid skill instances are retained in the original inventory, and deduplication and folding are performed on skills with completely identical text in the recommended retrieval inventory.

[0017] Preferably, the environment dependency parsing in the skill profile construction module specifically includes:

[0018] The identification skills can run on the operating system, have the required write permissions, require high privileges, require network access, depend on external credential systems, depend on specific command-line tools, depend on model context protocol services, and require specific environment variables.

[0019] Preferably, the quality assessment in the skill profile construction module includes three dimensions: content consistency, citation completeness, and task orientation, while the verifiability assessment includes three dimensions: result verifiability, environment reproducibility, and task constructibility.

[0020] Preferably, the verification task constructed by the task construction and verification module is a task with a low completion rate when the corresponding skill is not used, and a significantly improved completion rate after the corresponding skill is used. The verification process is executed in a container sandbox and supports repeated resets.

[0021] Preferably, in the two-stage recommendation mechanism of the recommendation service module, the first stage is a vector retrieval for rapid recall based on a selected skill set, and the second stage is a fine-grained sorting based on an intelligent system organized by a catalog.

[0022] Preferably, the general skills encapsulated in the client access module support installation via shell commands and automatic prompts from the intelligent agent. After installation, a unified entry point for recommended and feedback calls is formed locally.

[0023] Preferably, when the client access module initiates a recommendation request, it rewrites the user's original request into a task description suitable for retrieval, includes client information, and sends the request to the recommendation service module after locally verifying the API key and download directory permissions.

[0024] Preferably, the structured feedback data received by the feedback service module includes a session identifier, an environment tag, and a subtask-level execution summary. The environment tag only reports the names of environment variables and not their values.

[0025] Preferably, the subtask-level execution summary includes a subtask objective description, a summary of factual actions, the name of the actual skill performed, and subtask results and evidence. The subtask results include success, partial success, failure, and undetermined status.

[0026] Another aspect of the present invention provides an intelligent recommendation and feedback evolution method based on agent skills, comprising the following steps:

[0027] Skill acquisition and preprocessing: The agent skill description files are obtained through the code hosting platform interface or file acquisition tools. The format parser is used to perform field integrity verification, syntax validity verification and duplicate content identification on the skill description files. Meta information is supplemented according to skill name, function description, applicable scenarios and dependency conditions to generate a selected skill set.

[0028] Skill profile construction: By using rule analysis tools and large language model analysis tools, environmental dependencies, input and output conditions, applicable task types, completeness of instructions, quality scores and verifiability tags are extracted for each skill in the selected skill set to form a structured skill profile;

[0029] Task Construction and Verification: For skills with verifiable labels, automatically construct verification tasks that include objectives, input data, expected results, and evaluation criteria, and call the sandbox execution tool to start a real intelligent agent to run the skill. Generate verification results based on execution logs, output results, failure reasons, and resource consumption, and write the verification results back to the corresponding skill profile.

[0030] Runtime recommendation: After the client initiates a task request, the user's original request is rewritten into a search task description. Candidate skills are retrieved from the selected skill set through a vector search tool. The intelligent body ranking module then ranks the candidate skills based on semantic relevance, target matching degree, environmental adaptability, and verification results, and outputs a recommended skill list.

[0031] Client access and execution: After receiving the recommended skill list, the client downloads and installs the target skill according to the skill download address, version information and dependency conditions. During the local task execution process, the client records the actual skill name, subtask target, execution status and result evidence.

[0032] Feedback collection and offline iteration: After the task is completed, the client will form structured feedback data with session identifier, environment tag, actual executed skills, sub-task level execution summary, execution result and evidence description and upload it to the server. The server associates recommendation records and feedback records based on session identifier, stores the feedback data in the feedback database, and regularly updates the skill profile, selected skill set and recommendation ranking strategy according to usage rate, success rate, failure mode and environment adaptability.

[0033] Preferably, in the skill profile construction step, environmental dependency analysis uniformly extracts skill operation constraints, quality assessment filters stable and usable skill units, and verifiability assessment determines whether a skill is suitable for an automated verification process.

[0034] Preferably, in the task construction and verification steps, the verification task is used to distinguish between the model's own capabilities and skill enhancement effects, and the verification results include success rate, token consumption, and failure mode.

[0035] Preferably, in the runtime recommendation step, the intelligent system prioritizes browsing highly relevant categories and candidate files, and completes the sorting by comprehensively considering semantic matching, target focus, completeness of instructions, and environmental adaptability.

[0036] Preferably, in the feedback collection and iteration steps, the structured feedback data is a standardized compression of the original running trajectory, retaining only the core information of skill usage, environmental constraints, and result evidence.

[0037] In a preferred embodiment of the present invention, the skill collection and preprocessing module uses an open-source code hosting platform search interface to achieve large-scale collection. It is not limited to a single platform interface and is compatible with multiple platforms, mirror sites, and index services. The official skill verification script can be replaced with an equivalent legality verification tool, and the deduplication strategy can adopt other repeated folding methods. The core goal is to reduce the noise of the recommended candidate set.

[0038] In another preferred embodiment of the present invention, the skill profile construction module may introduce supplementary means such as static script analysis, configuration file analysis, and controlled trial operation analysis. The large language model used is not limited to a specific brand and version, and its core function is to transform unstructured skill materials into standardized structured profiles.

[0039] In another preferred embodiment of the present invention, in the two-stage recommendation mechanism of the recommendation service module, vector retrieval can be replaced by other equivalent semantic recall methods, and intelligent agent-based fine ranking can adopt other terminal-type intelligent agents with file navigation and multi-candidate comparison capabilities. The core of both is to achieve fine ranking under directory organization.

[0040] In another preferred embodiment of the present invention, the general skills of the client access module can be replaced by plugins, extensions, command-line wrappers, proxy components, etc. Feedback data can be reported locally after being summarized or extracted in a structured manner on the server side. The core is to realize a closed-loop access of recommendation call and feedback submission.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) This invention constructs a carefully selected skill set through large-scale collection, format verification, and deduplication on the recommendation side, which greatly reduces the noise of the original skill base and improves the input quality of retrieval and recommendation;

[0043] (2) This invention constructs a three-dimensional structured skill profile based on environment dependence, quality, and verifiability, thereby achieving the standardized extraction of skill execution constraints and solving the problem of incompatibility between skills and operating environment;

[0044] (3) This invention can verify the reverse construction of high-gain tasks by skills and verify them by running real intelligent agents to obtain objective usability indicators, thereby upgrading the recommendation basis from text matching to real running performance.

[0045] (4) The present invention adopts a two-stage recommendation mechanism of vector recall and intelligent body fine ranking, which takes into account both the efficiency of large-scale retrieval and the accuracy of fine ranking.

[0046] (5) This invention encapsulates recommendation and feedback capabilities into general skills, enabling client-side access without modification and at low cost, without requiring modification of the agent kernel;

[0047] (6) This invention collects structured feedback data by using session identifiers as a link to form an operational evidence base, supports offline iterative optimization of the skill base, and constructs a complete technical closed loop in which skills can be collected, verified, profiled, validated, recommended, and fed back for evolution.

[0048] (7) The feedback data of this invention only reports non-sensitive information, protecting user privacy and system security; the system does not rely on Web front-end interaction and can be fully embedded in the client intelligent agent reasoning process to achieve fully automated skill recommendation and use.

[0049] (8) This invention comprehensively solves the core technical defects of the existing intelligent agent skill ecosystem, significantly improves the accuracy of intelligent agent skill recommendation, execution success rate and system universality, and provides solid technical support for the large-scale, standardized and intelligent development of the intelligent agent skill ecosystem. Attached Figure Description

[0050] Figure 1 This is a system structure diagram of the intelligent recommendation and feedback evolution system based on agent skills of the present invention;

[0051] Figure 2 This is a flowchart illustrating the intelligent recommendation and feedback evolution method based on agent skills of the present invention. Detailed Implementation

[0052] The present invention will be further described in detail below with reference to specific embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0053] like Figure 1 and Figure 2 As shown, this embodiment provides an intelligent recommendation and feedback evolution system based on agent skills, including a skill acquisition and preprocessing module, a skill profile construction module, a task construction and verification module, a recommendation service module, a client access module, and a feedback service module; it also provides an intelligent recommendation and feedback evolution method based on agent skills, including six core steps: skill acquisition and preprocessing, skill profile construction, task construction and verification, runtime recommendation, client access, and feedback collection and iteration.

[0054] The skill acquisition and preprocessing module is used to retrieve agent skill description files on a large scale from the open-source code hosting platform. While obtaining skill files through the platform's search interface, it simultaneously collects related resources such as repository owner, repository name, number of stars, number of forks, relative skill path, scripts, templates, configuration files, and documentation to ensure the completeness and traceability of skill sources. An official skill verification script is used to perform format legality checks on candidate skills, filtering out entries with invalid formats, incomplete structures, or that do not conform to skill specifications, retaining only legal skills for subsequent processes. A dual storage mode is constructed: the original inventory and the recommended retrieval inventory. The original inventory fully retains all legal skill instances for traceability and statistical analysis, while the recommended retrieval inventory performs deduplication and folding on skills with completely identical text, retaining only a single representative entry to reduce the interference of duplicate data on retrieval and recommendation. Finally, a low-noise, high-quality, and well-selected skill set suitable for retrieval and recommendation is generated.

[0055] The skill profiling module, connected to the skill acquisition and preprocessing module, employs a dual-path approach combining rule extraction and large language model analysis to generate a unified structured skill profile for each legitimate skill. This profile comprises three core dimensions: environment dependency resolution, quality assessment, and verifiability assessment, serving as a unified intermediate representation for subsequent recommendation, verification, and feedback iterations. Environment dependency resolution identifies and structurally extracts the operational constraints of skills, including the type of operating system that can run, the required write permissions, whether high-level privileges such as sudo are required, whether the core process requires internet access, whether it depends on external systems with credentials, whether it depends on specific command-line tools, whether it depends on Model Context Protocol (MCP) services, and whether specific environment variables are required. This transforms unstructured environment information scattered in documents, scripts, and configurations into standardized fields, resolving compatibility issues before skill execution. Quality assessment revolves around the stability and usability of the skills themselves, encompassing three dimensions: content consistency, reference completeness, and task orientation. Content consistency determines whether a skill revolves around a single, clear goal, without mixing multiple themes. Reference completeness assesses whether the scripts, templates, resources, and dependencies mentioned in the skill documentation are complete and correspond to each other. Task orientation determines whether the skill provides executable operational guidance, rather than loose knowledge materials. Stable and usable skill units are selected through this three-dimensional assessment. Verifiability assessment determines whether a skill is suitable for entering the automated task construction and execution verification process, encompassing three dimensions: result verifiability, environment reproducibility, and task constructibility. Result verifiability refers to the ability to determine task success or failure with low ambiguity; environment reproducibility refers to the ability to build and repeatedly reset the skill's operating environment in a container sandbox; and task constructibility refers to the ability to generate corresponding tasks and verification logic in batches at a reasonable cost. Only skills meeting all these conditions proceed to the subsequent task construction and verification stages.

[0056] The task construction and verification module, connected to the skill profile construction module, constructs specialized verification tasks from the skill text itself for skills that have passed the verifiability assessment. These tasks are not general benchmark tasks, but rather high-gain tasks with low completion rates when the corresponding skill is not used, but significantly improved completion rates when the corresponding skill is used. This effectively distinguishes between the model's own capabilities and the gains brought by the skill. The task construction process extracts executable task instances from the skill objective, input-output relationship, preconditions, execution steps, and success signals, and generates corresponding environment definitions and verification logic. A combination of real client agents and models is used to execute the verification tasks in a container sandbox environment. The sandbox environment supports repeated resets to ensure the objectivity and repeatability of the verification results. During operation, objective engineering indicators such as task verification results, success rate, token consumption, execution time, and common failure modes are recorded in real time. The verification results are written back to the structured profile of the corresponding skill, so that skill recommendations no longer rely solely on text descriptions, but combine real-world performance, greatly improving the actual usability of the recommendation results.

[0057] The recommendation service module is connected to the skill acquisition and preprocessing module, the skill profile construction module, and the task construction and verification module. It adopts a two-stage recommendation mechanism to balance retrieval efficiency and recommendation accuracy. The first stage is vector retrieval for rapid recall. Based on the preprocessed and selected skill set, vector retrieval is performed to quickly filter out candidate skills semantically relevant to the current task from a large candidate set, providing high-quality input for the second stage of fine ranking. The second stage is intelligent agent-based fine ranking. The candidate skills recalled in the first stage are reorganized into a hierarchical local directory structure according to their internal classification, rather than laying out all skills in full in the context window. A terminal-type intelligent agent with strong terminal navigation capabilities reads, browses, and compares the directory tree, prioritizing access to highly relevant categories and candidate files, and completing a detailed ranking within a limited context budget. The fine ranking process comprehensively considers multiple dimensions such as text semantic matching degree, skill target focus degree, completeness of usage instructions, environmental adaptability, clarity of execution boundaries, and real-world verification results. Finally, it outputs a recommended skill list and a matching user guide. The user guide includes descriptions of skill functions, source information, and operational prerequisites, providing complete support for intelligent agent execution.

[0058] The client access module, connected to the recommendation service module, innovates by encapsulating recommendation and feedback capabilities into a universal skill that can be installed on the client. This eliminates the need to modify the client agent's kernel inference logic or the platform's backend interface, enabling low-cost cross-product and cross-platform access. The universal skill supports two installation methods: manual installation via shell commands or automatic prompting from the client agent when a skill requirement is detected. After installation, a unified entry point for recommendation calls, skill downloads, and feedback submissions is established locally. When the agent executes a task, the client access module rewrites the user's original request into an independent, clear, and searchable task description, optionally including the client name and version information. It locally verifies the validity of the API key and write permissions to the download directory. Upon successful verification, it sends a request to the recommendation service module, receives the recommendation results and session identifier, and downloads the recommended skill to a local writable directory for subsequent use by the agent. After task execution, it organizes structured feedback data based on the session identifier and calls the feedback interface to upload the data.

[0059] The feedback service module connects to the client access module and the skill profile building module. Using session identifiers as the core link, it achieves precise correlation between recommended sessions and feedback data, ensuring the traceability and correspondence of feedback data. The received structured feedback data is not the original, lengthy execution trajectory, but a standardized and compressed evidence package containing three core parts: session identifier, environment tag, and subtask-level execution summary. The environment tag describes the dependent environment during skill execution, including operating system type, maximum write range, whether high privileges are required, external dependency level, environment variable name, top-level executable program name, model context protocol identifier, and environment description text. Only environment variable names are reported, not variable values, to avoid leakage of sensitive information such as keys. The subtask-level execution summary contains subtask identifiers, environment tags, and subtask-level execution summaries. The task objective description includes a summary of factual actions, the actual skill name executed, the sub-task result, and evidence description. The actual skill name executed is limited to the set returned in this recommendation session, and only the skills actually used are recorded. The sub-task result includes four states: success, partial success, failure, and uncertain. The uncertain state is used in scenarios where there is a lack of clear verification signals. The feedback service stores the received valid feedback data in the feedback library to form a standardized operational evidence dataset, which serves as the core data support for subsequent skill profile updates, the organization of selected skill sets, the adjustment of recommendation strategies, and the offline evolution of the skill library. In this embodiment, the feedback capability is used for data accumulation and does not perform automatic skill generation or automatic repair. By accumulating evidence first and then iteratively optimizing, the system stability and data validity are ensured.

[0060] In a preferred embodiment of the present invention, the skill collection and preprocessing module uses an open-source code hosting platform search interface to achieve large-scale collection. It is not limited to a single platform interface and is compatible with multiple platforms, mirror sites, and index services. The official skill verification script can be replaced with an equivalent legality verification tool, and the deduplication strategy can adopt other repeated folding methods. The core goal is to reduce the noise of the recommended candidate set.

[0061] In another preferred embodiment of the present invention, the skill profile construction module may introduce supplementary means such as static script analysis, configuration file analysis, and controlled trial operation analysis. The large language model used is not limited to a specific brand and version, and its core function is to transform unstructured skill materials into standardized structured profiles.

[0062] In another preferred embodiment of the present invention, in the two-stage recommendation mechanism of the recommendation service module, vector retrieval can be replaced by other equivalent semantic recall methods, and intelligent agent-based fine ranking can adopt other terminal-type intelligent agents with file navigation and multi-candidate comparison capabilities. The core of both is to achieve fine ranking under directory organization.

[0063] In another preferred embodiment of the present invention, the general skills of the client access module can be replaced by plugins, extensions, command-line wrappers, proxy components, etc. Feedback data can be reported locally after being summarized or extracted in a structured manner on the server side. The core is to realize a closed-loop access of recommendation call and feedback submission.

[0064] The specific content of the present invention will be further illustrated below through specific embodiments:

[0065] Example 1: Skill Base Construction and Profile Formation

[0066] This embodiment is a basic system embodiment, which realizes large-scale collection of intelligent agent skills, rule preprocessing, structured profile generation and data storage, providing high-quality data support for subsequent verification, recommendation and feedback stages.

[0067] In this embodiment, the skill acquisition and preprocessing module initiates a large-scale acquisition process. The acquisition target is the agent skill description files in the open-source code hosting platform. The acquisition interface adopts the platform's official search interface, which is compatible with multiple platform interfaces and mirror index services, ensuring the comprehensiveness of the acquisition scope. During the acquisition process, the system not only acquires the skill description files themselves, but also simultaneously acquires related information, including the repository owner, repository name, number of stars, number of forks, and relative skill paths. At the same time, it captures all related resources under the skill directory, including script files, template files, documentation, configuration files, dependency lists, etc., to ensure the completeness of skill information and provide comprehensive data support for subsequent profile construction and source tracing. After the acquisition is completed, the system enters the format legality verification stage, using the official skill verification script to perform item-by-item verification. The verification rules cover dimensions such as file structure, syntax, content completeness, and execution feasibility. Candidate items with invalid formats, missing structures, or that do not conform to skill specifications are directly filtered out, retaining only the legal skills that pass the verification, eliminating invalid data from the source, and reducing subsequent processing costs.

[0068] After verification, the system adopts a dual-inventory storage mode. The original inventory module stores all legal skill instances completely without any deletions or modifications, preserving the original source, version, and associated resources of the skills for subsequent statistical analysis, source tracing, version comparison, and other scenarios. The recommendation retrieval inventory module performs deduplication processing. The deduplication rule is that the skill description files must be completely identical. Duplicate entries are collapsed, retaining only one representative entry to avoid duplicate skills interfering with retrieval efficiency and recommendation results. The dual-inventory mode balances data integrity and recommendation purity, ensuring the traceability of original data while improving the processing efficiency and result quality of the recommendation process. After the triple processing of collection, verification, and deduplication, the system generates a selected skill set, which serves as the core candidate set for subsequent skill profile construction, task verification, and recommendation services.

[0069] After the selected skill set is generated, the skill profile construction module initiates a dual-path analysis process. The rule extraction path and the large language model analysis path are executed in parallel, outputting a unified structured skill profile. The rule extraction path, based on preset rules and regular expressions, explicitly extracts environmental constraints, dependency information, and execution conditions from skill description files, scripts, configurations, and documents. The large language model analysis path understands, refines, and annotates unstructured text, generating implicit quality labels and verifiability labels. The results of the two paths are then fused to form a three-dimensional skill profile.

[0070] In terms of environment dependency resolution, the system comprehensively identifies and standardizes the recording of skill execution constraints, including compatible operating system types such as Windows, Linux, and macOS; the scope of write permissions required for execution, such as user directories, system directories, and specified paths; whether high privileges such as sudo and administrator are required; whether the core execution process of the skill must be connected to the internet; whether it depends on external systems requiring API keys and account credentials; whether it depends on specific command-line tools; whether it depends on the Model Context Protocol (MCP) service; and whether environment variables need to be specified. All of the above information is converted into structured fields, replacing traditional unstructured text descriptions, enabling the system and agents to quickly read, judge, and match environment compatibility without the need for manual document parsing.

[0071] In terms of quality assessment, the system scores and judges skills from three levels: First, regarding content consistency, it determines whether the skill revolves around a single task objective, without mixing multiple themes or adding irrelevant content, ensuring the skill's function is singular and clear. Second, regarding the completeness of references, it verifies that all scripts, templates, resources, and dependencies mentioned in the skill documentation exist in the skill directory, without missing parts, incorrect paths, or version incompatibility. Third, regarding task orientation, it determines whether the skill provides executable steps, execution logic, and constraints, rather than merely providing knowledge introductions or explanations of principles—content without execution value. Through this three-dimensional quality assessment, the system selects high-quality skills with standardized content, complete resources, and stable execution, while eliminating low-quality entries that are "similar in form but useless in practice."

[0072] In terms of verifiability assessment, the system determines whether a skill is suitable for an automated verification process. Regarding result verifiability, it assesses whether the skill's execution result has clear and unambiguous criteria, such as file generation, successful command execution, and output matching. Regarding environment reproducibility, it determines whether the skill's operating environment can be built in a container sandbox, without dedicated hardware, non-resettable dependencies, or customized system configurations. Regarding task constructibility, it determines whether executable task instances and verification logic can be quickly generated based on the skill text without extensive manual intervention. Only skills that simultaneously meet all three conditions are marked as verifiable skills and enter the subsequent task construction and verification module. Skills that do not meet the conditions retain their profile and recommendation capabilities but are not forcibly included in the automated verification process, ensuring efficient use of system resources.

[0073] Once the skill profile is constructed, the system uniformly stores the skill text, repository metadata, associated resources, structured profile, quality tags, and verifiability tags into a massive skill database, forming a centralized, standardized, searchable, and scalable skill data center. This provides complete data support for task construction and verification, recommendation services, and feedback iteration. This embodiment fully realizes the entire process of skill processing from original open-source resources to standardized and selected candidate sets and structured profiles, without omitting any technical features, laying a solid foundation for the subsequent implementation of system functions.

[0074] Example 2: Recommendation Closed Loop for Client-Side Intelligent Agents

[0075] This embodiment is a core function embodiment of the system, realizing the skill recommendation service, low-threshold client access, and the entire process of runtime skill invocation, achieving deep coupling between the agent reasoning process and the recommendation system.

[0076] In this embodiment, the client access module encapsulates recommendation and feedback capabilities into standardized general skills. These skills are independent of specific clients, platforms, and agent types, and can run on all client agents that support standardized skill mechanisms. Access requires no modification to the client kernel code, no alteration of the backend interface, and no development of dedicated plugins, significantly reducing access costs. The general skills offer two installation and deployment methods: manual installation by the user via shell command, where the client automatically downloads, configures, and enables the general skills; and automatic installation by the agent, where the client agent automatically prompts and completes the installation of the general skills when it detects that a task requires external skill support, without requiring manual user intervention. After installation, a unified functional entry point is generated locally, including five core functions: recommendation request initiation, skill list retrieval, local skill download, feedback data organization, and feedback data upload, seamlessly integrating with the client agent's inference process.

[0077] After a user inputs a task command on the client, the task enters the execution preparation phase, and the client access module initiates the recommendation request process. First, the client agent semantically refines and rewrites the user's original request, transforming the vague, verbose, and irrelevant request into an independent, clear, semantically concise task description suitable for retrieval, improving search matching accuracy. Optionally, the client name and version information can be included for compatibility adaptation by the recommendation service module. Subsequently, the client access module performs local pre-verification, checking the existence, validity, and access permissions of the API key, as well as the existence, writability, and storage space of the local skill download directory. If all verifications pass, a standardized recommendation request is sent to the recommendation service module; if verification fails, a clear prompt is output to avoid invalid requests.

[0078] Upon receiving a recommendation request, the recommendation service module initiates a two-stage recommendation mechanism. The first stage involves rapid vector retrieval. Based on the curated skill set generated by the skill acquisition and preprocessing module, a vector index library is constructed. The task description is converted into a vector representation, and similarity calculation is performed between the vector and the skill vector to quickly retrieve semantically relevant candidate skills. The retrieval scale can be adjusted according to system performance and requirements, improving processing efficiency while ensuring candidate coverage. The second stage involves intelligent agent-based fine sorting. The system reorganizes the candidate skills retrieved in the first stage into a hierarchical local directory structure according to the internal classification information in the skill profile. The directory structure is consistent with the local file system, facilitating navigation, browsing, and reading by the terminal-type intelligent agent. Subsequently, a terminal-type intelligent agent with strong terminal navigation, file reading, and multi-text comparison capabilities is launched to traverse the directory tree, prioritizing skill categories and candidate files with high semantic similarity, high quality scores, and good environmental adaptability. This avoids displaying all skill texts in the context window, reducing context usage and improving processing efficiency.

[0079] During the fine-tuning process, the system comprehensively considers multiple dimensions to rank the skills, including text semantic matching degree, skill objective focus, completeness of usage instructions, environmental adaptability, clarity of execution boundaries, success rate of real-world verification, token consumption, and failure mode occurrence rate, rather than solely relying on semantic similarity. After fine-tuning, the system outputs the final recommendation results, including a list of recommended skills, reasons for recommendation, and a skill usage guide. The guide details the core function of each skill, open-source source information, prerequisite operating environment, permission requirements, dependencies, and execution precautions, providing comprehensive guidance for the safe and stable execution of the agent. Simultaneously, the recommendation service module generates a unique session identifier, which is bound to this recommendation request and returned to the client access module along with the recommendation results for subsequent feedback data association.

[0080] After receiving the recommendation results and session identifier, the client access module stores the session identifier locally, establishes a binding relationship between the recommended session and the local task, and then downloads the skill files from the open-source repository or a designated server to a writable directory on the local machine, completing the local deployment of skills. During subsequent inference execution, the client agent can call the locally downloaded skills as needed according to the requirements of the subtasks, eliminating the need for repeated requests and downloads, thus improving execution efficiency. During task execution, the client access module monitors skill usage, environmental status, and execution results in real time, providing raw information for subsequent feedback data processing.

[0081] This embodiment achieves automated and accurate acquisition of skills during client-side intelligent agent runtime through general skill encapsulation, two-stage recommendation, and local skill download. It eliminates the need for manual searching, filtering, and installation of skills, fully embedding them into the intelligent agent's reasoning process and automating the entire process of intelligent agent skill usage.

[0082] Example 3: Feedback Accumulation and Subsequent Offline Updates

[0083] This embodiment is a system iterative optimization embodiment, which realizes structured feedback data collection, uploading, storage, analysis and offline evolution of the skill library, and builds a complete closed loop from skill use to skill library optimization.

[0084] In this embodiment, after the client agent determines that the task has been completed, the client access module initiates a feedback data processing procedure. Using the session identifier returned by the recommendation service module as the core index, it associates the current recommendation session, downloaded skills, execution process, and operating environment to ensure that the feedback data accurately corresponds to the recommendation record. The feedback data is not a complete upload of the original execution trajectory, but rather a standardized and compressed structured evidence package. Redundant commands, irrelevant logs, and sensitive information are removed, retaining only core information related to skill usage, environment adaptation, and execution results, balancing data availability, transmission efficiency, and privacy security.

[0085] The structured feedback data comprises three core parts. The first part is the session identifier, which uniquely corresponds to the current recommendation session, achieving precise binding between recommendation and feedback. The second part is the environment tag, which describes the client's operating environment during skill execution, including the operating system type, maximum write permission range, whether high permissions are used, external dependency level, names of involved environment variables, name of the top-level executable program, model context protocol identifier, and evidence-based environment description text. The environment tag only reports variable names, not specific variable values, to avoid leakage of sensitive information such as API keys and tokens, ensuring system and user security. The third part is the subtask-level execution summary. Each subtask entry includes an independent and understandable goal description, a summary of factual actions and environment feedback, the name of the skill actually used in the current recommendation session, the subtask execution result, and evidence supporting the result judgment. The subtask execution result is divided into four states: success, partial success, failure, and uncertain. The uncertain state is used in scenarios where the environment lacks clear verification signals and the result cannot be objectively judged, avoiding data distortion caused by forced judgment.

[0086] After the feedback data is processed, the client access module uploads the structured feedback data to the feedback service module through a secure encrypted channel. Upon receiving the data, the feedback service module performs data integrity checks, legality checks, and session identifier validity checks, filtering out invalid, duplicate, and erroneous data, retaining only valid feedback data. After successful verification, the feedback service module stores the feedback data in a feedback database. The feedback database adopts a centralized and standardized storage structure. Each record includes a skill identifier, session identifier, environment tag, subtask summary, execution result, evidence description, timestamp, and client information, forming a traceable, statistically significant, and analyzable dataset of operational evidence.

[0087] In this embodiment, the system does not perform automatic skill generation, automatic skill repair, or automatic skill optimization. The core function of the feedback data is to accumulate real operational evidence, providing data support for subsequent offline analysis and skill library iteration. The system operator can periodically execute the offline analysis process, analyzing content including skill usage rate, success rate under different environments, common failure modes, environment adaptation issues, dependency missing situations, task matching degree, etc., identifying problematic skills, high-quality skills, skills with unclear environment descriptions, skills with missing dependencies, and unstable execution skills. Based on the analysis results, the system can perform three major optimization operations: First, update the skill profile, correcting environment dependencies, quality scores, and verifiability tags; second, adjust the selected skill set, removing high-frequency failure and skills with no practical use value, and supplementing with high-quality skills; third, optimize the recommendation strategy, adjusting recommendation weights, sorting rules, and recall rules to improve recommendation accuracy.

[0088] The offline iteration process combines manual review with system automation to ensure the accuracy and stability of the skill library updates, avoiding unfounded automatic modifications that could reduce skill usability. By first accumulating feedback evidence and then iterating and optimizing offline, the system continuously accumulates real-world operational data while maintaining the stability of the current version, gradually improving the quality of the skill library and the effectiveness of the recommendation service, thus forming a virtuous cycle of continuous evolution of the skill library.

[0089] This embodiment fully implements the entire process of structured feedback acquisition, secure transmission, standardized storage, offline analysis, and skill library iteration. It does not add any unimplemented functions or exaggerate the system's capabilities, providing core support for the long-term stable operation and continuous optimization of the system.

[0090] Example 4: Specific Application Example for Data Collection and Presentation Generation Tasks

[0091] This embodiment uses the task of a client agent performing "collecting publicly available information on two open-source software programs, MemTensor and MemOS, organizing the structured content, generating a PowerPoint presentation, and generating an HTML version for local deployment" as an example to illustrate the recommendation, execution, feedback, and evolution process of the system in this invention in a real-world task. In this embodiment, the client device is a user terminal running the client agent program. The server includes a recommendation service module, a skill library, a feedback service module, a feedback library, and an offline analysis module. The client and server interact with each other through recommendation and feedback interfaces.

[0092] After the user inputs the above task on the client device, the client access module first rewrites the user's original request into a task description suitable for skill retrieval, and collects current operating environment tags, including operating system type, writable directory range, network connection status, whether high privileges are required, and available command-line tools. Subsequently, the client access module sends a recommendation request to the recommendation service module. The recommendation service module performs recall and ranking based on the structured skill profiles in the skill library, combining task semantics, skill purpose, environmental adaptability, quality score, verifiability tags, and historical running feedback to output skill recommendation results matching the task.

[0093] In this embodiment, the recommendation service module outputs three candidate skills. The first skill is used for multi-source public information retrieval and data organization; the second skill is used to generate presentation files based on the organization results; and the third skill is used to generate HTML display pages based on the presentation content and support local deployment. The recommendation service module also generates a skill usage guide, breaking down the task into sub-steps such as data collection, content summarization, presentation generation, HTML page generation, and local service startup. It also explains the intermediate output transfer relationships between each sub-step, for example, saving the data collection results as a structured document, which is then used as input for the presentation generation and HTML generation skills.

[0094] After receiving the recommendation results, the client agent downloads and invokes the corresponding skills in the recommended order. First, the client agent invokes the data collection skill to access public web pages, open-source repository metadata, and documentation pages, generating a traceable data compilation file. Second, it invokes the presentation generation skill to convert the data compilation file into a presentation file containing a title page, product introduction, feature comparison, and conclusion page. Third, it invokes the HTML generation skill to convert the presentation content into a single-file HTML page and starts a local HTTP service, allowing users to access the generated results through a specified local port. During execution, the client access module records the actual skill name invoked, subtask objective, execution action summary, generated file path, service startup status, and result evidence.

[0095] Upon completion of the task, the client access module generates structured feedback data and uploads it to the feedback service module via the feedback interface. The structured feedback data includes at least a recommended session identifier, client environment tags, a list of skills actually used, a sub-task-level execution summary, the execution result of each sub-task, and supporting evidence. The environment tags include the operating system type, write permission range, network connectivity, whether high privileges were used, and the names of available command-line tools, but sensitive values ​​such as API keys and access tokens are not uploaded. The sub-task-level execution summary includes whether data collection was completed, whether a presentation file was generated, whether an HTML page was generated, whether the local service was started, the names of the skills used in each step, and their success or failure status.

[0096] After receiving the structured feedback data, the feedback service module associates and stores the current recommendation record with the execution feedback record based on the recommendation session identifier. The offline analysis module periodically analyzes the call frequency, success rate, failure reasons, environmental adaptability, and intermediate product connection of different skills in this type of task, and updates the skill profile, adjusts the recommendation weight, or marks unstable skills accordingly. Thus, this embodiment completes a closed loop from user task input, automatic skill recommendation, client execution, structured feedback upload to subsequent offline optimization of the skill library, proving that the present invention can realize the discovery, invocation, verification, feedback, and evolution of skills in specific intelligent agent tasks.

[0097] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make many modifications and variations based on the concept, technical solutions, and embodiments of the present invention without creative effort, such as replacing the data collection interface, verification tools, large language models, recommendation algorithms, access methods, etc., all without departing from the core technical ideas of the present invention. Therefore, any equivalent technical solutions obtained by those skilled in the art through logical analysis, reasoning, and a limited number of experiments based on the concept, technical solutions, and embodiments of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An intelligent recommendation and feedback evolution system based on agent skills, characterized in that, include: The skill acquisition and preprocessing module is used to retrieve agent skill description files from open-source code hosting platforms on a large scale, verify the legality of skill execution formats, complete metadata and remove duplicates in the recommendation side, and generate a curated set of skills. The skill profile construction module, connected to the skill acquisition and preprocessing module, is used to generate a structured skill profile for each skill, including environmental dependency analysis, quality assessment, and verifiability assessment, through rule extraction and large language model analysis. The task construction and verification module, connected to the skill profile construction module, is used to reverse construct executable verification tasks for skills that have passed the verifiability assessment, run the tasks in a sandbox environment using a combination of real intelligent agents and models, obtain the verification results of skill execution success rate, resource consumption and failure mode, and write them back to the skill profile. The recommendation service module, connected to the skill acquisition and preprocessing module, skill profile construction module, and task construction and verification module, is used to perform vector retrieval to recall candidate skills based on the selected skill set, reorganize the candidate skills into a directory structure according to internal classification, perform intelligent agent-style fine ranking through terminal intelligent agent, and output the final recommended skill list and usage guide. The client access module, connected to the recommendation service module, is used to encapsulate recommendation and feedback capabilities into general skills that can be installed on the client, enabling recommendation request initiation, skill download, and feedback submission in the client agent reasoning process; The feedback service module, connected to the client access module and the skill profile construction module, is used to receive structured feedback data uploaded by the client, associate recommended sessions with feedback data using session identifiers, and store feedback data to form a feedback library, providing data support for skill profile updates and skill library iterations.

2. The system according to claim 1, characterized in that, The skill acquisition and preprocessing module is specifically used for: The system retrieves agent skill description files through the open-source code hosting platform's search interface, and simultaneously collects the corresponding warehouse owner, warehouse name, number of stars, number of forks, skill relative path, and associated resources. The official skill verification script is used to perform format validity checks, and only valid skills are retained for subsequent processes. All valid skill instances are retained in the original inventory, and deduplication and folding are performed on skills with completely identical text in the recommended retrieval inventory.

3. The system according to claim 1, characterized in that, The environment dependency parsing in the skill profile construction module specifically includes: The identification skills can run on the operating system, have the required write permissions, require high privileges, require network access, depend on external credential systems, depend on specific command-line tools, depend on model context protocol services, and require specific environment variables.

4. The system according to claim 1, characterized in that, The quality assessment in the skill profile construction module includes three dimensions: content consistency, citation completeness, and task orientation. The verifiability assessment includes three dimensions: result verifiability, environment reproducibility, and task constructibility.

5. The system according to claim 1, characterized in that, The verification task constructed by the task construction and verification module is a task with a low completion rate when the corresponding skill is not used, and a significantly improved completion rate after the corresponding skill is used. The verification process is executed in a container sandbox and supports repeated resets.

6. The system according to claim 1, characterized in that, In the two-stage recommendation mechanism of the recommendation service module, the first stage is fast retrieval based on vector retrieval of a selected skill set, and the second stage is fine sorting based on intelligent system organization of a directory.

7. The system according to claim 1, characterized in that, The general skills encapsulated in the client access module can be installed via shell commands or automatically prompted by the agent. After installation, a unified entry point for recommended and feedback calls is formed locally.

8. The system according to claim 1, characterized in that, When the client access module initiates a recommendation request, it rewrites the user's original request into a task description suitable for retrieval, includes client information, verifies the API key and download directory permissions locally, and then sends the request to the recommendation service module.

9. The system according to claim 1, characterized in that, The structured feedback data received by the feedback service module includes session identifier, environment tag, and subtask-level execution summary. The environment tag only reports the names of environment variables and not their values.

10. The system according to claim 9, characterized in that, The subtask-level execution summary includes a description of the subtask objective, a summary of factual actions, the name of the actual skill performed, and subtask results and evidence. The subtask results include success, partial success, failure, and undetermined status.

11. A method for intelligent recommendation and feedback evolution based on agent skills, characterized in that, Includes the following steps: Skill Acquisition and Preprocessing: Acquire agent skill description files at scale from open-source code hosting platforms, perform format validation, metadata completion, and recommendation-side deduplication to generate a curated set of skills; Skill profile construction: Through rule analysis and large language model analysis, a structured skill profile is created to reflect the environment-dependent, quality, and verifiability of skill generation. Task Construction and Verification: For skills that meet the verifiability conditions, reverse construction verification tasks are carried out. Real intelligent agents are used to run the verification tasks in a sandbox and record the verification results. The verification results are then updated to the skill profile. Runtime recommendation: Based on the selected skill set, vector retrieval is performed to recall candidate skills, which are then reorganized into a directory structure and refined through intelligent algorithms to obtain recommendation results, which are then returned to the client; Client-side integration: The recommendation and feedback capabilities are encapsulated as general skills and installed on the client. The client initiates a recommendation request and downloads the recommended skills when the task begins. Feedback collection and iteration: After the client completes the task, it uploads structured feedback data. The server associates recommendations and feedback with session identifiers and stores the feedback data for offline optimization of the skill library.

12. The method according to claim 11, characterized in that, In the skill profile construction process, environment dependency analysis is used to uniformly extract skill operation constraints, quality assessment is used to screen stable and usable skill units, and verifiability assessment is used to determine whether a skill is suitable for an automated verification process.

13. The method according to claim 11, characterized in that, In the task construction and verification steps, the verification task is used to distinguish between the model's own capabilities and skill enhancement effects. The verification results include success rate, token consumption, and failure mode.

14. The method according to claim 11, characterized in that, During runtime recommendation steps, the intelligent system prioritizes browsing highly relevant categories and candidate files, and completes the sorting by comprehensively considering semantic matching, target focus, completeness of instructions, and environment adaptability.

15. The method according to claim 11, characterized in that, In the feedback collection and iteration steps, the structured feedback data is a standardized compression of the original operation trajectory, retaining only the core information of skill usage, environmental constraints, and result evidence.

Images