Growing Docs
raihank43/claude-growing-docs
Documentação que cresce com seu código e sobrevive a perda de contexto
Ver no GitHubPlugin Claude Code com comandos slash para scaffolding e manutenção automática de documentação otimizada para IA que persiste após compactação de contexto. Mantém registros de ideias rejeitadas, anti-padrões e gotchas para evitar que o agente repita erros em sessões longas.
Instalação
npx convez-ai raihank43/claude-growing-docs1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Manter memória persistente de decisões rejeitadas em projetos construídos com IA
- Registrar anti-padrões e falhas para evitar repetição de erros
- Documentar gotchas e edge cases que desaparecem com compactação de contexto
- Scaffolding automático de estrutura docs para novos projetos
Veredito

Solução pragmática para um problema real em sessões longas com IA: documentação como memória do agente, focada no que foi rejeitado e por quê. Reduz significantly a repetição de ideias descartadas após compactação de contexto.
README
A tradução para o português deste README ainda está na fila.
growing-docs
Documentation that scaffolds itself, grows with your codebase, and survives context compaction.
A Claude Code plugin with three slash commands for projects you build with an AI agent over long, single-conversation sessions. It sets up a documentation system the agent maintains automatically — so when the conversation gets compacted and the agent loses context, the knowledge is already written down where it can be re-read.
The problem it solves
When you build a project in one long conversation with an AI agent, two things happen:
- Context compaction eventually drops details — gotchas, decisions, conventions, half-finished plans.
- The agent re-invents bad approaches and re-suggests rejected ideas because the reasons they were dropped vanished with the context.
growing-docs fixes this by giving the project a set of documents the agent reads before every task and updates after every change. The docs are the agent's long-term memory.
What makes it different
Most "AI memory" setups record decisions — but an agent can usually re-derive what to do from the code. The expensive thing that vanishes in compaction is why a path was abandoned. That's what makes an agent confidently re-suggest the idea you shot down an hour ago.
So growing-docs treats the negative space as first-class:
- Rejected Ideas (
PLAN.md) — what you considered and decided not to do, and why. - Anti-Patterns (
RULES.md) — what was tried, broke, and shouldn't be repeated — with the failure reason. - Gotchas (every feature doc) — the edge cases and "looks right but isn't" traps.
Recording the dead ends is the difference between docs that describe the project and docs that stop the agent from repeating your mistakes.
What you get
| File | Purpose |
|---|---|
CLAUDE.md | The agent's workflow: read docs → work → update docs → verify → commit. Plus a git convention and an index of everything. |
docs/PLAN.md | Roadmap, feature status table, decisions log, and rejected ideas — with a Project Phase marker (BRAINSTORMING / BUILDING) that tells a fresh session how to behave. |
docs/ARCHITECTURE.md | Tech stack, folder structure, system overview, data flow. |
docs/RULES.md | Coding conventions, naming, and anti-patterns (things that were tried and failed). |
docs/_feature-template.md | Copied per feature; every feature gets its own doc with a Gotchas section. |
README.md | The human-facing project overview. |
The core idea: the agent decides after every request — large or small — which of these docs need updating, and keeps them in sync with the code. Code is always the source of truth; stale docs get fixed.
Commands
/project-init — new project from scratch
Pitch-first. Describe your idea (you don't even need a name — it'll suggest one), and it scaffolds the docs, initializes git, then brainstorms the plan with you, filling in the docs as you talk. It handles the "I started Claude from my home directory" case by telling you exactly how to relocate without losing anything.
/project-adopt — existing codebase
Run it from inside a repo. It can:
- Adopt an un-instrumented project — you choose how deeply to scan (barebones / map existing docs / full scan), so it never burns tokens crawling a huge repo you didn't ask it to.
- Upgrade a project already using growing-docs to the latest templates — without touching your filled-in content.
- Consolidate a fat existing CLAUDE.md — if a project crammed everything (architecture, gotchas, procedures) into one giant CLAUDE.md, adopt offers to move that knowledge into
docs/and slim CLAUDE.md back to a lean workflow + index. Your choice (it can also just merge), and nothing is lost — content is relocated, then verified, before anything is removed.
/checkpoint — mid-session save-and-sync
Run it at a save point: a feature landed, you're taking a break, or the conversation has gotten big and you want to start a fresh one. It reconciles the docs against both the code and the current conversation — capturing decisions, gotchas, and rejected ideas that were only ever discussed — refreshes the staleness markers, and writes a "Current Focus" handoff note. The deliberate, lossless alternative to waiting for auto-compaction: checkpoint, then start a fresh chat that inherits everything through the docs. Commits following the project's own git/release convention.
Install
/plugin marketplace add raihank43/claude-growing-docs
/plugin install growing-docs
Then /project-init and /project-adopt are available in any conversation.
To update later, pull the newest version:
/plugin marketplace update claude-growing-docs
You can also run it from a local clone instead of GitHub:
/plugin marketplace add /absolute/path/to/claude-growing-docs
How it stays consistent
Both commands copy from one shared template source (plugins/growing-docs/project-scaffold/), referenced via ${CLAUDE_PLUGIN_ROOT} so it resolves correctly on any machine. Editing a template once updates what both commands produce.
License
MIT — see LICENSE.