Repository Learning Capture
Turn important repository discoveries into reusable skills or instructions.
The skill only directs repository-local inspection and creation or updating of skill and instruction files; there is no evidence of malware, credential theft, or covert exfiltration. README, LICENSE, and the verified Microsoft provenance provide traceable attribution. However, it does not require confirmation before overwriting existing files and does not define backups, rollback, least-change scope, or sensitive-data handling, so points are deducted.
The workflow includes triggers, a classification decision, file checks, quality checks, and user notification, making the normal path reasonably clear. But “major learning” is subjective, the commands assume particular directories and grep availability, and failure cases such as missing files, content conflicts, write errors, or validation failures lack diagnostic and recovery guidance. Static review limits the score and provides no execution credit.
The audience and triggers are fairly clear, including “learn!”, significant patterns, pitfalls, and reusable knowledge. The skill fits repository-maintenance work and has no dependency on overseas services. Boundaries for non-fit cases, criteria for judging significance, Chinese-language workflows, and differing agent-directory layouts are under-specified, so points are deducted.
The document provides frontmatter, structured sections, a Skill/Instruction/Learning decision, path conventions, a template example, quality checks, and a completion-notification requirement. It lacks installation or dependency notes, versioning policy, changelog, explicit maintenance ownership, and a clear update path; repository license context does not make the skill's own governance complete, so points are deducted.
The skill can turn session learnings into persistent repository documentation and supplies actionable classification and authoring steps, so its core value is credible. There are no representative generated outputs, automated checks, or tests demonstrating the key path, and the final content still depends on manual judgment and editing. Under static calibration it cannot exceed 7 and is scored below that ceiling.
The skill text, repository license, README, version metadata, and related CI/test files are auditable, and its triggers and file format can be reviewed. There is no dedicated test coverage, execution record, or third-party reproduction evidence for this skill, so it receives a near-maximum static score of 4 rather than full marks.
- Before updating an existing skill or instruction, show the proposed diff and obtain user confirmation, while preserving a reversible backup or version-control record.
- Add handling and diagnostic guidance for missing directories, command failures, duplicate knowledge, naming conflicts, and write failures.
- The criteria for a “major learning” and long-term retention are subjective; add operational thresholds and non-fit examples.
What it does & when to use it
This skill helps capture important repository knowledge discovered during a session as persistent skills or instructions. It covers identifying the learning, checking existing files, choosing the right destination, performing quality checks, and informing the user of the result. It classifies knowledge as an update to an existing instruction, a standalone skill, or an instruction based on scope and reuse. It is suited to ongoing maintenance of development repositories that need to accumulate project knowledge.
Identifies recurring patterns, non-obvious pitfalls, architectural constraints, and reusable domain knowledge; uses ls to inspect .github/skills/, .agents/skills/, and .github/instructions/, and uses grep to search related content; adds small learnings to an existing instruction's ## Learnings section, or creates or updates .github/skills/{name}/SKILL.md and .agents/skills/{name}/SKILL.md; it can also create an instruction file with applyTo; before saving, it checks generality, actionability, examples, duplication, and description clarity, then tells the user where the content was saved and whether existing content was updated or a new file was created.
- Maintainers of large repositories such as VS Code who need to preserve an important architectural constraint for future sessions.
- Agent users who say `learn!` and want the session's learning captured through a defined procedure.
- Contributors who discover a recurring debugging pitfall and need to decide whether it belongs in an existing instruction or a standalone skill.
- Repository maintainers who want a rule to apply automatically by file pattern through an `applyTo` instruction.
Pros & cons
- Defines an end-to-end process from identifying a learning to saving it and informing the user.
- Clearly separates learnings, skills, and instructions to guide destination selection.
- Requires concrete examples and checks for generality, actionability, and duplication.
- Uses plain `SKILL.md`, shell commands, and filesystem operations, supporting generic Agent Skills clients.
- Provides a workflow and file conventions, but no automation scripts or test suite.
- The source does not document standalone installation, client configuration, or failure recovery.
- Classification and content quality depend on the agent's judgment about scope, reuse, and applicability.
How to install
The source does not provide a standalone installation command. The skill is located at src/vs/sessions/skills/update-skills/SKILL.md; the complete skill collection is available from https://github.com/microsoft/vscode. The repository uses the MIT license.
How to use
Load the skill in an Agent Skills-compatible client, then send learn! when an important repository learning is discovered, or explicitly ask to capture a learning. The source does not specify additional client configuration, trigger registration, or a standalone execution command.