update-agent-learnings（更新 Agent 学习）

Non-negotiable rules:

1. Only record learnings that belong in agent memory.
2. Keep one central learnings source of truth; do not invent parallel central files.
3. Sync approved learnings into every relevant agent surface that exists for the supported CLIs.
4. Do not propagate "Claude Code Only" learnings into engineer/reviewer agent prompts.
5. Get explicit user confirmation before modifying the learnings file or any agent file.

Update Agent Learnings

Inputs

• $request: Optional learning candidate, scope hint, agent name, or reminder about what the session revealed

Goal

Add one validated agent learning to the central learnings store and sync it into the matching agent files by:

• confirming the learning belongs in agent memory
• classifying the scope correctly
• updating one canonical learnings file
• regenerating the relevant ## Learnings sections
• syncing those sections into all agent surfaces that exist

Step 0: Confirm the learning belongs here

This skill is only for durable learnings that should shape agent behavior.

Valid examples:

• global coding-agent rules such as scope control, testing, or iteration
• agent-specific rules for one technology or role
• "Claude Code Only" learnings that belong in the central learnings store but should not be pushed into subagent prompts

Invalid examples:

• skill-design rules
• main CLAUDE.md workflow rules
• one-off implementation notes
• direct requests to rewrite an agent prompt right now

Load references/learning-scope.md for routing and scope classification.

If the learning does not belong in agent memory, stop and say where it should go instead.

Success criteria: The learning clearly belongs in the agent learnings system.

Step 1: Extract one concrete learning

Review the session and identify the smallest useful rule.

Classify it as one of:

• Global
• Claude Code Only
• Agent-Specific

Rules:

• write it in imperative mood
• prefer one precise learning over a vague bundle
• only mark it Global if it truly applies across coding agents
• use Claude Code Only for meta-work about skills, orchestration, configs, or project setup

Success criteria: You have one actionable learning candidate with a correct scope.

Step 2: Resolve the central learnings file and agent sync targets

Locate the canonical central learnings file.

Path policy:

• if one central agent learnings file already exists, use it
• if both .agents/learnings/agent-learnings.md and .claude/learnings/agent-learnings.md exist, pick one canonical source and do not maintain both by hand
• in this .agents-first repo, prefer .agents/learnings/agent-learnings.md
• if the repo only has .claude/learnings/agent-learnings.md, use that instead

Then discover agent sync targets:

• sync into .agents/agents/*/AGENT.md when that tree exists
• sync into .claude/agents/* when that tree exists
• treat both trees as live CLI surfaces when both are present

Load:

• references/learning-scope.md for scope and duplicate handling
• references/agent-learnings-template.md only if the canonical learnings file does not exist yet
• references/agent-sync-contract.md for Learnings-section generation and placement

Success criteria: The canonical learnings file and all sync target trees are known.

Step 3: Confirm with the user

Before editing anything, present:

• scope classification
• final wording
• canonical learnings file
• sync targets that will be touched

Use AskUserQuestion if confirmation or wording refinement is needed.

Do not write until the user explicitly approves the update.

Success criteria: The user has approved the learning and the sync surface.

Step 4: Update the central learnings file

Apply the minimal correct edit:

• preserve file structure
• insert the learning in the correct section
• avoid deleting unrelated content
• update the "Last updated" marker only if the file already uses one

Rules:

• if the canonical learnings file is missing, create it from references/agent-learnings-template.md
• if the section is missing, create the smallest compatible section rather than restructuring the whole file
• keep formatting consistent with the existing document

Success criteria: The central learnings file contains the approved learning exactly once.

Step 5: Regenerate and sync Learnings sections into agent files

Use the central learnings file to build the ## Learnings section for each target agent file.

Sync rules:

• global learnings go to all coding/reviewer agent files
• agent-specific learnings go only to the matching agent files
• "Claude Code Only" learnings stay in the central learnings file and are not pushed into subagent prompts
• if an agent file already has a ## Learnings section, replace that section cleanly
• if it does not, insert the section in the location defined by references/agent-sync-contract.md

Important:

• when both .agents and .claude agent trees exist, update both surfaces
• do not assume filename parity; resolve the actual paths present
• do not rewrite unrelated prompt sections while syncing learnings

Success criteria: Every relevant agent file in every present CLI tree has the correct synced Learnings section.

Step 6: Verify and report

Verify:

• the learning exists once in the central learnings file
• sync targets were updated as intended
• agent Learnings sections contain the right global and agent-specific content
• "Claude Code Only" learnings did not leak into subagent prompts

Report:

• scope classification
• canonical learnings file
• sync target trees updated
• final wording
• whether files were created or updated

Success criteria: The user can see exactly what changed centrally and across agent surfaces.

Guardrails

• Do not let the model invoke this skill proactively; it mutates durable learnings and agent prompt files.
• Do not add context: fork; this workflow edits the active repository.
• Do not add paths:; this is a generic maintenance skill.
• Do not keep routing matrices, scorecards, or giant Learnings examples inline in SKILL.md.
• Do not add a learning without explicit user approval.
• Do not maintain two divergent central learnings files.
• Do not skip one CLI tree when both .agents and .claude agent surfaces are present.

When To Load References

• references/learning-scope.md Use for deciding whether the learning belongs in agent memory, choosing the right scope, and handling duplicates.
• references/agent-learnings-template.md Use only when the canonical central learnings file is missing and a minimal compatible file must be created.
• references/agent-sync-contract.md Use for generating the Learnings section and placing it correctly in agent files across the supported CLI trees.

Output Contract

Report:

1. whether the learning was accepted or redirected elsewhere
2. the chosen scope and canonical learnings file
3. the final approved wording
4. which CLI agent trees were updated
5. any duplicate merge or sync-target decisions