AgentSkillsCN

learn

将WorkOS AuthKit与React Router应用集成。支持v6与v7（Framework、Data、Declarative模式）。适用于项目使用react-router、react-router-dom，或提及React Router认证的场景。

SKILL.md
--- frontmatter
name: learn
description: >-
  Analyze what was just learned in this session. Surfaces the what, the why,
  hidden patterns, connections to existing knowledge, and non-obvious insights.
  Does NOT write to files — that's /integrate's job.
user-invocable: true
allowed-tools: Read, Glob, Grep, Bash

Learn

Analyze what just happened and extract the real learnings. Not a summary — a diagnosis.

What To Do

1. Gather What Just Happened

Look at the recent session activity:

bash
# Recent git changes
git -C /home/clawd log --oneline -10

# Modified/new files
git -C /home/clawd status --short

# Recent build activity
ls -t /home/clawd/systems/orchestrator/logs/*.jsonl 2>/dev/null | head -3

# Recent eval results
ls -t /home/clawd/evals/results/*.json 2>/dev/null | head -5

Read any files that were recently created or modified. Understand what work was done.

2. Read Existing Knowledge

Before analyzing, understand what we already know:

code
/home/clawd/systems/orchestrator/HEURISTICS.md  — existing rules
/home/clawd/LEARN.md                             — existing techniques
/home/clawd/MEMORY.md                            — operational state
/home/clawd/memory/YYYY-MM-DD.md                 — today's log

You need this context to spot connections and avoid restating what's already captured.

3. Analyze at Three Levels

Level 1: What happened?

  • •What was built, changed, fixed, or discovered?
  • •Be specific. "Updated the build system" is worthless. "Added H14 Strategy Council heuristic after 5-agent planning experiment proved convergence signals are stronger than single-agent analysis" is useful.

Level 2: Why does it matter?

  • •What problem does this solve?
  • •What capability does this unlock?
  • •What would have gone wrong without this learning?
  • •Who benefits (human, agents, future sessions)?

Level 3: Hidden patterns and connections This is the hard part. Look for:

  • •Rhymes with existing knowledge — Does this reinforce or refine an existing heuristic? Is it a new instance of a known pattern?
  • •Contradictions — Does this contradict something we believed? If so, which belief needs updating?
  • •Generalizations — Does this specific learning apply more broadly? (e.g., "Council of Judges for code quality" generalizing to "Council of Judges for planning")
  • •Gaps revealed — What does this learning expose that we DON'T know yet?
  • •Compound effects — How does this interact with other recent learnings? Do two things combine into something bigger?

4. Present the Analysis

Output a structured analysis. Use this format:

code
## What We Learned
[2-3 bullet points, specific and actionable]

## Why It Matters
[1-2 sentences on the impact — what changes because of this]

## Hidden Patterns
[Connections to existing knowledge, generalizations, contradictions]

## Gaps Revealed
[What we still don't know, what should be investigated]

## Suggested Integration Points
[Which files/sections should be updated — but DON'T update them, that's /integrate]

Important

  • •Don't write to any files. This skill is read-only analysis. Use /integrate to persist learnings.
  • •Don't be shallow. "We learned that X works" is not analysis. Dig into the why and the connections.
  • •Challenge assumptions. If a learning seems obvious, ask why it wasn't already known. That gap is itself a learning.
  • •Name the pattern. If you spot a reusable pattern, give it a name. Named patterns spread; unnamed ones don't.
  • •Be honest about uncertainty. If you're not sure something generalizes, say so. Premature generalization is worse than no generalization.