AgentSkillsCN

integrate

将WorkOS AuthKit与Next.js App Router（13+）集成。需采用服务器端渲染。

SKILL.md
--- frontmatter
name: integrate
description: >-
  Integrate learnings into the full memory system. Links new knowledge to
  existing concepts, updates the right files, keeps memory tidy, and sets up
  future sessions for success. Run after /learn or whenever significant
  knowledge needs to be persisted.
user-invocable: true
allowed-tools: Read, Write, Edit, Glob, Grep, Bash

Integrate

Wire new learnings into the memory system so they compound instead of evaporate.

What To Do

1. Understand What Needs Integrating

First, figure out what knowledge needs to be captured. Sources:

  • •The current conversation — what was just discussed, built, or discovered
  • •Output from /learn — if it was run, use its analysis as the starting point
  • •Uncommitted changes — git -C /home/clawd status --short may reveal work not yet documented
  • •Recent evals/builds — check logs for undocumented results

2. Read ALL Target Files Before Editing

This is critical. Read every file you plan to modify BEFORE making changes:

code
/home/clawd/systems/orchestrator/HEURISTICS.md    — rules (H1-H15, E1-E6)
/home/clawd/LEARN.md                               — techniques & patterns
/home/clawd/MEMORY.md                              — operational state
/home/clawd/memory/YYYY-MM-DD.md                   — today's daily log
/home/dcarmitage/.claude/projects/-home-dcarmitage/memory/MEMORY.md  — auto memory

You need to understand what's already there to:

  • •Avoid duplicating knowledge
  • •Find the right place to add new knowledge
  • •Link new learnings to existing concepts
  • •Consolidate if old entries are now subsumed by new ones

3. Decide Where Each Learning Goes

Not everything goes everywhere. Use this routing table:

Type of LearningWhere It Goes
Reusable rule with evidenceHEURISTICS.md (new H# or update existing)
Technique or how-toLEARN.md section 4 (Techniques & Patterns)
What happened todaymemory/YYYY-MM-DD.md
System state changeMEMORY.md (update relevant section)
Cross-session patternAuto memory (~/.claude/.../MEMORY.md)
Refinement of existing ruleUpdate the existing entry, don't create a new one
Contradiction of existing beliefUpdate or remove the old entry, add the correction

4. Integration Rules

Link, don't island. When adding a new concept, explicitly reference related existing concepts. "This is H14 applied to X" or "This extends the pattern from section 4.2" or "See also: memory/2026-02-05.md". Isolated knowledge is forgotten knowledge.

Consolidate, don't accumulate. If a new learning subsumes or refines an old one, UPDATE the old entry rather than adding a new one alongside it. Memory files that only grow eventually become unreadable.

Preserve numbering. HEURISTICS.md uses H1-H15 and E1-E6. When adding a new heuristic, use the next number. Never renumber existing entries — other files reference them by number.

Keep MEMORY.md under 200 lines. If it's getting long, move detailed content to topic-specific files and link from MEMORY.md. Operational state should be scannable, not a novel.

Keep auto memory focused. The auto memory (~/.claude/.../MEMORY.md) is loaded into every session's system prompt. Only put things there that help EVERY future session — key patterns, gotchas, system locations. Not daily details.

Date your updates. When modifying a file, update the "Last updated" line if one exists.

Follow existing format. Each file has its own conventions. Match them:

  • •HEURISTICS.md: ### H#: Title + bold rule + explanation + evidence + action
  • •LEARN.md: Tables, code blocks, structured sections
  • •Daily memory: Sections from the save-memory template
  • •MEMORY.md: Dense operational reference
  • •Auto memory: Bullet points, brief

5. Execute the Integration

Make all edits. For each file:

  1. •Read it (if not already read)
  2. •Identify the insertion/update point
  3. •Edit precisely — don't rewrite sections you're not changing
  4. •Verify the edit didn't break surrounding content

6. Verify

After all edits:

bash
# Files were actually modified
git -C /home/clawd diff --stat

# MEMORY.md is still under 200 lines
wc -l /home/clawd/MEMORY.md

# No broken markdown (quick check)
head -5 /home/clawd/systems/orchestrator/HEURISTICS.md
head -5 /home/clawd/LEARN.md
head -5 /home/clawd/MEMORY.md

7. Report What Was Done

Tell the user exactly what was integrated and where:

code
| File | What Changed |
|------|-------------|
| HEURISTICS.md | Added H16: ... |
| LEARN.md | Updated Strategy Council section with ... |
| memory/2026-02-05.md | Added experiment results |
| MEMORY.md | Updated heuristic count |
| Auto memory | Added key pattern for ... |

Important

  • •Read before write. Always. Blind edits break things. You MUST understand what's in a file before modifying it.
  • •Don't bloat. If you're adding more than you're consolidating, something is wrong. Good integration often makes files shorter, not longer.
  • •Test your links. If you reference another file or section, verify it exists.
  • •Don't create new files unless necessary. Prefer editing existing files. New files fragment knowledge.
  • •Be surgical. Edit the minimum needed. Don't reformat surrounding text, don't add comments to code you didn't change, don't "improve" nearby sections.
  • •This is not /save-memory. /save-memory is a comprehensive end-of-session dump. /integrate is surgical insertion of specific learnings into the right places. Use /save-memory at session end; use /integrate after any significant discovery.