convez.ai
PluginProdutividadeIniciante
ingridtoulotte

Teach2Claude

ingridtoulotte/Teach2Claude

Converta correções repetidas em regras permanentes no CLAUDE.md

Ver no GitHub
BomPluginScore · de 0 a 100

Plugin JavaScript que analisa histórico de sessões do Claude Code, identifica correções repetidas e gera automaticamente regras persistentes em um arquivo CLAUDE.md. Funciona 100% local, sem dependências ou chamadas de rede.

Instalação

$npx convez-ai ingridtoulotte/Teach2Claude

1 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

Estrelas
4
Forks
1
Acompanhando
0
Saúde
100%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Eliminar feedback repetitivo sobre preferências de código (pnpm vs npm, estilo de comentários, etc)
  • Gerar automaticamente um arquivo CLAUDE.md com as regras mais-frequentes da equipe
  • Auditar padrões de correção em histórico de sessões para melhorar prompts

Veredito

Ferramenta bem-pensada que resolve um problema real: Claude esquece instruções repetidas entre sessões. A abordagem local e sem dependências é sólida, ideal para quem já usa Claude Code intensivamente e precisa de memória persistente de preferências de projeto.

README

A tradução para o português deste README ainda está na fila.

Teach2Claude

Stop teaching Claude the same thing over and over.

Turn repeated corrections into permanent CLAUDE.md memory.

You've told Claude Code "use pnpm, not npm" five times this month. It's already written down in your session logs — Claude just never reads them back. Teach2Claude does.

✓ Learns from your history  ·  ✓ Generates persistent rules  ·  ✓ 100% local ✓ No API key  ·  ✓ No cloud  ·  ✓ Zero dependencies

CI license node deps network calls

graph LR
    A["🗂 Claude sessions<br/><code>~/.claude/projects/*.jsonl</code>"] --> B["🔍 Teach2Claude"]
    B --> C["♻️ Repeated corrections"]
    C --> D["📜 Generated rules"]
    D --> E["🧠 CLAUDE.md"]
    E --> F["✅ Future sessions<br/>that already know"]

teach2claude distill — your most-repeated corrections, counted, dated, and drafted as a ready-to-paste CLAUDE.md block.


The hidden cost of forgetting

Claude Code logs every session — every correction, every preference, every "no, do it this way." Then it starts the next session knowing none of it.

Here's what one typical week of that looks like (illustrative example — run distill to see yours):

Correction you typedTimes repeated
Use pnpm, not npm5
Run the tests before committing3
Don't add comments I didn't ask for2
Repeated instructions typed:  10
Rules they boil down to:       3
Repetitions needed after
adding them to CLAUDE.md:      0

The rules you most need in CLAUDE.md are exactly the ones you're too busy repeating to write down. If you've corrected Claude twice for the same thing, that rule already exists in your history. Teach2Claude finds it.

Before → After

BEFORE — you, every single session
┌──────────────────────────────────────────┐
│ Session 1   ❯ use pnpm, not npm          │
│ Session 2   ❯ use pnpm, not npm          │
│ Session 3   ❯ use pnpm, not npm          │
│ Session 4   ❯ use pnpm, not npm          │
│ Session 5   ❯ use pnpm, not npm          │
└──────────────────────────────────────────┘

AFTER — teach2claude, run once
┌──────────────────────────────────────────┐
│ ♻️  Recurring preference detected (5×)    │
│                                          │
│ Rule  →  CLAUDE.md:                      │
│   - use pnpm, not npm                    │
│                                          │
│ Future repetitions:  0                   │
└──────────────────────────────────────────┘

Real output:

$ teach2claude distill

DISTILL — recurring corrections you keep giving Claude

 5×  use pnpm, not npm                                preference  2026-06-09
 3×  always run the tests before committing           rule        2026-06-08
 2×  don't add comments unless I ask                  prohibition 2026-06-02

Suggested CLAUDE.md block (review before adopting):
- use pnpm, not npm
- always run the tests before committing
- don't add comments unless I ask

Paste the block into CLAUDE.md. Done. Those corrections never need typing again.

Why CLAUDE.md exists

  • Claude forgets everything between sessions.
  • Claude reads CLAUDE.md at the start of every session.
  • Teach2Claude bridges the gap: it moves what you keep saying into the one file Claude actually reads.
graph LR
    S1[Session 1] -. forgotten .-> S2[Session 2]
    S2 -. forgotten .-> S3[Session 3]
    M["🧠 CLAUDE.md<br/>(read every session)"] --> S1
    M --> S2
    M --> S3
    T["🔍 Teach2Claude"] -- "your repeated corrections" --> M

How it works

graph LR
    A["✏️ Correction"] --> B["🔎 Detection"]
    B --> C["🧩 Grouping"]
    C --> D["⚗️ Distillation"]
    D --> E["🧠 CLAUDE.md"]
    E --> F["✅ Future sessions"]
    F -. "fewer corrections needed" .-> A
  1. Scan — streams every session transcript on your machine (~/.claude/projects/*.jsonl).
  2. Detect — pattern-matches user messages that are corrections, prohibitions, or preferences (English and French).
  3. Group — normalizes and dedupes, so "use pnpm not npm" and "pnpm, not npm!" count as one rule said twice.
  4. Distill — anything repeated enough becomes a drafted rule, ranked by repetition count and recency.
  5. Reuse — paste the block into CLAUDE.md. Optionally pipe through Claude itself for polish: teach2claude distill --prompt | claude -p (no API key — it's your existing claude CLI).

What it usually finds

Illustrative distribution — varies by user. Run distill to see yours.

Preferences        ████████████░░░░░░░░░░░░░░░░░░  ~40%   "use pnpm, not npm"
Workflow rules     ██████████░░░░░░░░░░░░░░░░░░░░  ~35%   "run tests before committing"
Coding style       █████░░░░░░░░░░░░░░░░░░░░░░░░░  ~15%   "no comments unless asked"
Other              ███░░░░░░░░░░░░░░░░░░░░░░░░░░░  ~10%   project-specific habits

At a glance

100%001
LocalDependenciesNetwork callsPurpose
your data stays on diskpure Node ≥18 stdlibno telemetry, everteach Claude your rules

Why it's different

  • Prompt hacks — "always remember to…" pasted at the top of every session. Decays the moment you forget to paste it.
  • Hand-written memory files — you write the rules you remember needing, not the ones you actually repeat most.
  • AI wrapper apps — your transcripts get uploaded somewhere, an API key gets involved, a subscription appears.

Teach2Claude is none of these. The rules come from evidence — what you actually said, counted and dated — and the output goes into the one file Claude already reads.

Teach2ClaudePrompt hacksManual CLAUDE.mdWrapper apps
Finds rules automatically✅❌❌some
Evidence-based (counts real corrections)✅❌❌❌
Persists across sessions✅❌✅varies
Local-first — nothing leaves your machine✅✅✅❌
Repeatable — same history, same rules✅❌❌❌
Transparent — shows counts, dates, sources✅❌❌❌
No API key, no subscription✅✅✅❌
Dependencies0——many

Who it's for

You are…Teach2Claude gives you…
Solo developeryour personal habits become standing rules — once
Open-source maintainerproject conventions enforced every session, not re-explained per PR
Staff engineerteam standards (test-first, commit style) distilled from how you actually review
AI-heavy workflowdozens of sessions/week — repetition compounds fastest, so does the payoff
Team on Claude Code dailya shared CLAUDE.md grown from evidence instead of meeting notes

Why trust it

  • ✓ Reads local logs only (~/.claude) — read-only
  • ✓ No telemetry — not even update checks
  • ✓ No uploads — zero network calls, verifiable in the source
  • ✓ No cloud processing — everything runs on your machine
  • ✓ Open source — MIT, small dependency-free codebase you can read in one sitting
  • ✓ Reproducible — same history in, same rules out; deterministic pattern matching, not a black box
  • ✓ Transparent — every suggested rule shows its repetition count and last-seen date
  • ✓ You stay in control — it suggests a CLAUDE.md block; you review and paste

Quick start

 ① Install  ──▶  ② Run  ──▶  ③ Review rules  ──▶  ④ Paste into CLAUDE.md  ──▶  ✅ Done

① Install (or skip straight to npx):

npm install -g github:ingridtoulotte/Teach2Claude

② Run:

teach2claude distill                          # see what you keep repeating
teach2claude distill --prompt | claude -p     # let Claude itself polish the rules

③ Review the suggested rules — each one shows how often you said it and when.

④ Paste the block into your CLAUDE.md.

Zero-install one-liner:

npx github:ingridtoulotte/Teach2Claude distill

No config. No API key. No telemetry. It reads ~/.claude and prints answers.

What happens after install

Day 1   ─────────●  Run distill. Your most-repeated corrections surface,
                    counted and dated. Paste the first rules into CLAUDE.md.

Day 7   ─────────●  New repetitions show up as you work. CLAUDE.md grows
                    from evidence, not guesswork.

Day 30  ─────────●  Claude behaves consistently across sessions and
                    projects. Mid-session corrections became one line
                    Claude reads at startup.

Also in the box

teach2claude context                    # audit tokens injected before your first message
teach2claude search "race condition"    # full-text search across every session, every project
teach2claude stats                      # usage, top tools, estimated cost across your history
teach2claude sessions --since 7d        # browse recent sessions; `show <id>` replays one

Every command takes --json — including a CI guard that fails a PR when context startup cost crosses a budget.

teach2claude context deserves a special mention: it statically audits everything injected into your window before you type a word (MCP schemas, CLAUDE.md files, memory indexes) — these routinely eat 20%+ of a 200k window before hello.

teach2claude context — every token spent before you type a word, ranked, with one-line fixes.

Performance

Streaming JSONL parser with a raw-line prefilter before any JSON.parse:

scan:   29.2 MB, 100,000 lines in 170 ms   (≈588,000 lines/s)
search: 100,000 lines in 48 ms             (≈2,000,000 lines/s)

Reproduce: npm run bench.

Privacy

Read-only over your local ~/.claude directory. Zero network calls — no telemetry, no update checks, no API requests. The --prompt flow only prints text; you choose to pipe it into claude.

Docs

Getting started · Command reference · Architecture · FAQ · Roadmap

Contributing

PRs welcome — the codebase is small, dependency-free Node and stays that way. See CONTRIBUTING.md.

License

MIT


If you've ever typed the same instruction twice, Teach2Claude already has something to teach you.

npx github:ingridtoulotte/Teach2Claude distill

Saved you one "wait, I already told Claude this" moment? Star the repo so others find it. ⭐

Ver o original no GitHub

Bom para

Relacionados