AGENT SKILL

document

Code-is-truth documentation sync. Detects drift between the codebase and project docs (STATUS.md, CLAUDE.md, README), then applies targeted edits so docs describe what the code actually is. Use this after shipping features, before a release, or when docs feel stale. Secret-scans every doc before writing. Makes small surgical edits, never wholesale rewrites.

License 许可
MIT
最近更新
2026-06-23
Stars
★ 0
Forks
0

01 前言 Front Matter

name
document
description
Code-is-truth documentation sync. Detects drift between the codebase and project docs (STATUS.md, CLAUDE.md, README), then applies targeted edits so docs describe what the code actually is. Use this after shipping features, before a release, or when docs feel stale. Secret-scans every doc before writing. Makes small surgical edits, never wholesale rewrites.

03 SKILL.md 正文

document — Documentation Sync

Scan the code, detect drift from the docs, and patch the docs to match reality.

Usage: /document


Phase 1: Scan the Code (source of truth)

Genericize these probes to the project's actual layout. The goal is a current inventory of routes, modules, data models, and surfaces.

# Recent activity
git log --oneline -10

# Source size / shape («adjust glob to project language»)
find src -type f | wc -l

# Public entry points / routes («PLACEHOLDER: e.g. ls api/  or  routes manifest»)
ls api/ 2>/dev/null

# Modules of interest («PLACEHOLDER: commands/ services/ handlers/ ...»)
ls src/services/ src/commands/ 2>/dev/null

# Data models («PLACEHOLDER: ORM schema, e.g. grep '^model ' prisma/schema.prisma»)
# Skills / commands available
ls .claude/skills/ .claude/commands/ 2>/dev/null

Phase 2: Secret Scan (before any write)

grep -rn -iE "(sk_|pk_|api[_-]?key|secret|password|token|[a-z]+://[^ ]*:[^ ]*@)" \
  README.md CLAUDE.md docs/ .claude/ \
  --include="*.md" 2>/dev/null

If anything real is matched: STOP, alert the user, and remove the secret before writing anything. Do not commit.

Phase 3: Apply Targeted Edits

Update only the parts that drifted.

STATUS.md (path docs/STATUS.md or docs/status/STATUS.md)

  • Last-updated date
  • Completed work / recent PRs
  • Current phase and next steps
  • Known blockers

CLAUDE.md

  • Directory structure (only if it changed)
  • New commands / skills / scripts
  • Architecture patterns and workflows

README.md (if present)

  • Feature list, tech stack, setup steps

Phase 4: Alignment Check

Verify the docs now match the code on countable facts:

Check Doc Code source Verify
Data model count CLAUDE.md «schema file» equal
Command/route list CLAUDE.md «routes / commands dir» equal
Surface count STATUS.md «UI manifest» equal
Skill list CLAUDE.md .claude/skills/ equal

Rules

  1. Code is truth — docs describe what IS, not what should be.
  2. Targeted edits — patch lines; never rewrite whole files.
  3. Single source — each fact lives in one place; others reference it.
  4. Secret scan before every write — no exceptions.
  5. No emojis unless the file already uses them.

04 同仓库更多 Skills

Elad73
Elad73/agentic-base-layer

wrap-up

Session-end safe closeout. Updates STATUS, verifies you are on a feature branch (never main), security-greps the diff for secrets, makes a conventional commit (no AI-tool mentions), then asks for approval before pushing and opening a PR. Use this to properly close a development session. Side-effecting and gated on explicit user approval; run explicitly.

★ 0
Elad73
Elad73/agentic-base-layer

wake-up

Session-start context sync. Reads project STATUS, git state, and runs a quick health check, then emits a short (<15 line) announcement of where things stand and the single best next step. Use this at the beginning of a working session to re-establish context fast. Does NOT re-read CLAUDE.md (already in the system prompt). Read-only and fast.

★ 0
Elad73
Elad73/agentic-base-layer

station

End-of-session composition that chains retrospect → document → cleanup → wrap-up into one closeout pipeline. Use this to fully close a working session — capture learnings, sync docs, prune artifacts, then commit and open a PR — without forgetting a step. Side-effecting; honors each sub-skill's confirmation and approval gates. Run explicitly.

★ 0
Elad73
Elad73/agentic-base-layer

retrospect

Post-session learning capture. Mines the current conversation for reusable patterns, pitfalls, architecture decisions, performance and security insights, then appends them to the project knowledge registry. Use this at the end of a working session, after a bug is fixed, or after a non-obvious decision is made — so the next session starts smarter. Writes structured entries to .claude/knowledge/REGISTRY.md.

★ 0
Elad73
Elad73/agentic-base-layer

pre-ship

Mandatory pre-commit quality gates — type-check, lint, test, build — run in order and halting on the first failure. Use this BEFORE committing any change that touches application code, so failures are caught while context is still fresh instead of by CI. Surfaces the failing tool's verbatim error and refuses to proceed until gates pass.

★ 0
Elad73
Elad73/agentic-base-layer

debug-like-expert

Deep analysis debugging mode for complex issues. Activates methodical investigation protocol with evidence gathering, hypothesis testing, and rigorous verification. Use when standard troubleshooting fails or when issues require systematic root cause analysis.

★ 0