Open kit for Claude Code
An expert's method becoming your mentor
Teaches by building, reviews before delivery and only says "it works" after running it. Each rule points to a real passage of what the expert published.
Teaches by building, reviews before delivery and only says "it works" after running it. Each rule points to a real passage of what the expert published.
When you ask an AI to explain something, it tends to over-explain, sound right without being right, assume things without telling you and answer in generic terms. This kit builds a mentor that teaches the way someone you admire would, with proof for every step.
Everything the expert published (videos, texts, repositories, posts) becomes an interlinked wiki that the AI consults without getting lost in the haystack.
Each conduct rule carries a passage copied from the archive and the exact place it came from. A script rejects invented citations.
The mentor follows a fixed step-by-step, and an automatic gate won't let the turn end if code was written and not executed.
Each phase has a command that answers "finished?" with exit 0 or with the reason for failure. Nothing depends on the AI saying it's done.
raw/ is read-only, with a manifest and a hash for each item. The wiki has pages for sources, topics, principles and methods, plus index, hot and log.
active requires 2 different sources; bank waits for the 2nd; inference is flagged to the user. The comparison ignores case, accents and punctuation.
Ready → Smallest version → Prediction → Real output → Broken version → Report with the rule that guided each step.
The scripts use only the Python standard library. No paid API.
For the generator, the collectors and the validators.
python3 --version git --version
Where the agent, the skills and the gate run.
claude --versionyt-dlp for video subtitles and a local transcriber (Whisper, faster-whisper…) for videos without subtitles. pytest only for testing the kit.
yt-dlp --versionReal kit commands. Replace prof-redes with your mentor's short name.
The example is a fictional expert with 3 sources and 7 rules, passing all validators.
git clone https://github.com/inematds/mentor-especialista.git cd mentor-especialista python3 exemplo/tools/stats.py --raiz exemplo python3 exemplo/tools/validar_citacoes.py --raiz exemplo
Generates the folder with the archive, the wiki, the rules, the scripts, the agent, the 6 skills and the execution gate.
python3 novo-mentor.py prof-redes \ --nome "Expert's name" \ --dominio "teach neural networks by building from scratch" \ --destino ~/mentores
The sources go in ESCOPO.md. The collection targets, subtitle languages and your transcriber go in mentor.config.json (see docs/ADAPTAR.md (in Portuguese)).
cd ~/mentores/mentor-prof-redes # in mentor.config.json, for example: "transcrever_cmd": "your-transcriber --url {url} --out {saida}"
In Claude Code, inside the mentor folder. One subagent per source, in parallel. Anything that fails goes to raw/RELATORIO-FALHAS.md.
/prof-redes-coletar # done when: python3 tools/stats.py # → exit 0 (targets met, raw intact)
The wiki links sources, topics, principles and methods in both directions. The rules come out with a passage copied from the archive.
/prof-redes-compilar python3 tools/validar_links.py # → 0 broken links /prof-redes-regras python3 tools/validar_citacoes.py # → 0 missing
ensina builds along with you and checks the answer externally. revisa predicts where it breaks, reproduces it and proves the lean version side by side.
/prof-redes-ensina teach me to build a BPE tokenizer from scratch /prof-redes-revisa scripts/coleta.py python3 tools/validar_resposta.py testes/respostas/<arquivo>.md
A new link becomes a source page, updates the affected pages and can promote a rule from the bank to active.
/prof-redes-ingere https://exemplo.com/post-novo python3 tools/stats.py && python3 tools/validar_links.py && python3 tools/validar_citacoes.py
Build and teach; review a script that "works" but breaks without UTF-8; ingest a new source. Results go in testes/aceitacao/RESULTADOS.md.
PYTHONIOENCODING=cp1252 python3 testes/aceitacao/script_emoji.py # → UnicodeEncodeError: the mentor must predict this before running
Real outputs from the example that ships with the kit, and the gate in action. Outputs shown as the kit prints them (in Portuguese).
tipo itens palavras meta blog 1 123 ≥1 itens / ≥100 pal. posts 1 73 ≥1 itens / ≥50 pal. video 1 145 ≥1 itens / ≥100 pal. TOTAL 3 341 OK: acervo dentro das metas
7 regras (4 ativas), 12 citações encontradas, 0 faltando
OK
# with an invented citation:
7 regras (4 ativas), 11 citações encontradas, 1 faltando
REPROVADO{"decision": "block",
"reason": "Você escreveu código e não rodou: calc.py.
Rode cada um (ou os testes) e mostre a saída real
antes de dizer que funciona."}
Listing or opening the file (ls, cat), git commit or py_compile do not count as execution.
| passo | rodou | saiu | regra |
|------------------|--------------------|--------------------|--------|
| menor versão | contar('a b a') | {'a': 2, 'b': 1} | R2 |
| previsão × saída | o mesmo | bateu | R3, R4 |
| versão quebrada | 'a b'.split(' ') | ['a', '', 'b'] | R5 |The full documentation is in the repository: solution plan, training plan and adaptation guide.