A Claude Code plugin (also installable manually in Codex / OpenCode / etc.) that consolidates every scattered descriptive markdown file in a project (README.md, AGENTS.md, HANDOFF.md, docs/*) into a single canonical CLAUDE.md and reconciles it against the current code.
After one run, the next session — yours, a teammate's, or a fresh agent's — has exactly one entry point.
Working with AI coding agents tends to produce documentation sprawl: a HANDOFF.md from one session, an AGENTS.md from another, a half-stale README.md, a docs/ folder with three-week-old design notes. Each session's "context dump" creates a new file, and over time the project loses its single source of truth.
wrap does the opposite. Each invocation:
- Reads every descriptive
.mdin the project - Scans the actual code (entry points, scripts, routes, configs)
- Folds it all into one
CLAUDE.mdwith a fixed structure - Deletes the originals so there's nothing to drift back out of sync
- Pins a
PROGRESSblock at the top answering: where are we now, what's the next concrete action
This repo is a Claude Code plugin marketplace. From inside Claude Code:
/plugin marketplace add recomby-ai/wrap-skill
/plugin install wrap@wrap-skill
That's it — no git clone, no manual file copying. To update later, run /plugin marketplace update wrap-skill.
Manual install (other agents / no plugin support)
If you're using an agent that supports SKILL.md but not Claude Code's plugin system, copy the file manually:
mkdir -p ~/.claude/skills/wrap && curl -fsSL https://raw.githubusercontent.com/recomby-ai/wrap-skill/main/plugins/wrap/skills/wrap/SKILL.md -o ~/.claude/skills/wrap/SKILL.mdAdjust the destination path for Codex / OpenCode / Cursor / Gemini CLI as needed.
Just say one of these in any project:
wrap/wrap up//wraptidy docs/sync docs/update docs/archivethis phase is done- Chinese:
收尾/同步文档/整理文档/整理一下/更新文档/存档/梳理一下/这阶段做完了
The agent will:
- Inventory every
.mdand cross-check it against the real code - Write a fresh
CLAUDE.mdwithPROGRESS / Architecture / Run / API / Known Limits / Handoffsections - Delete the merged-in source files
- Print a short change summary
CLAUDE.md follows a fixed skeleton:
# <project>
<one-line summary>
## PROGRESS
**Current**: ...
**Next**: ...
**Known outstanding**: ...
## Architecture
## Run
## API / Interfaces (skipped if project has no external interfaces)
## Known Limits
## Handoff
Soft cap of 250 lines. If a project genuinely needs more (multi-service runbooks, large credential tables), the skill applies for a documented exemption rather than padding or hand-waving.
- One file. Zero is too few; two is too many.
- PROGRESS is the soul. Other sections survive a stale week; PROGRESS dies in a day. Rewrite it every run.
- Brevity beats completeness. Decision history and design rationale go in the agent's memory system, not in
CLAUDE.md. - Code is the source of truth. When
CLAUDE.mdconflicts with the code, the code wins — always.
Tested with Claude Code. Should work in any agent that supports the SKILL.md format (Codex, OpenCode, Cursor, Gemini CLI, etc.).
The skill only touches files inside the project directory. It does not interact with any agent's memory system — those are decoupled by design.
MIT — see LICENSE.