Altman Mode
Tarcroi/altman-mode
Transforma Claude Code em co-founder direto com avaliador brutal de startups
Ver no GitHubSkill que reconfiguração Claude Code para operar como co-founder ambicioso inspirado em princípios de Sam Altman. Inclui CLAUDE.md para reescrever defaults em torno de leverage, velocidade de ship, e pensamento AI-as-product, mais um evaluador de startups que faz web search e scores ideias em 10 dimensões ponderadas.
Instalação
npx convez-ai Tarcroi/altman-mode1 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
- Estruturar projetos de IA com guidelines de shipping rápido e leverage
- Avaliar ideias de startup com pesquisa web e scoring ponderado
- Reescrever comportamento padrão de Claude para contexto startup/infraestrutura
Veredito

Skill opinionado e bem-executado que reposiciona Claude para mentalidade startup. Útil para times que precisam de directness e prototipagem rápida; o avaliador com web search agrega valor real em due diligence de ideias, mas exige que o user já pense em termos de leverage e ship-speed para aproveitar todo o potencial.
README
A tradução para o português deste README ainda está na fila.
Altman-mode for Claude Code
Check out my open-source project Skrun — deploy any AI agent or skill as an API via
POST /run. The open, multi-model runtime for AI agents. Works with any LLM, on any infrastructure.
A CLAUDE.md and a startup evaluator that make Claude Code think like an ambitious co-founder, not a polite intern. Inspired by publicly discussed Sam Altman principles around leverage, shipping speed, AI-as-product thinking, and infrastructure-grade reliability.
The problem
Default Claude Code hedges. It asks five clarifying questions before writing one line of code. It ships 200 lines where 50 would do. It says "interesting approach!" to ideas it should call tarpits.
Fine for tutorials. Wrong mode for serious AI, startup, and infrastructure work.
What you get
- A
CLAUDE.mdthat rewires Claude's defaults around five rules: leverage, shipping speed, AI-as-product, safety-by-design, directness. - The
altman-mode-installskill — a one-liner setup. Tell Claude "install altman-mode in this project" and it drops (or merges) the CLAUDE.md into your project. - The
altman-graderskill — a brutal startup evaluator. You hand it an idea. It parses what's there, runs 5–10 web searches to verify competitors, recent funding, and timing signals, asks one sharp question at a time on what's still missing, then scores it out of 100 across 10 weighted categories — anchored on what the web actually shows.
Install
Step 1 — Install the plugin (once)
From inside Claude Code:
/plugin marketplace add Tarcroi/altman-mode
/plugin install altman-mode@altman-skills
Step 2 — Apply altman-mode to a project
In any project where you want the guidelines, just tell Claude:
install altman-mode in this project
The altman-mode-install skill takes over. It creates a CLAUDE.md if there isn't one, or merges the altman-mode block into your existing one with <!-- BEGIN/END: altman-mode --> markers — your existing rules are preserved. Re-run anytime to update the block in place.
The altman-grader skill triggers automatically on prompts like "score my idea" or "is this a tarpit?" — no setup needed.
Drop the CLAUDE.md into a single project by hand:
curl -o CLAUDE.md https://raw.githubusercontent.com/Tarcroi/altman-mode/main/CLAUDE.md
Append to an existing CLAUDE.md:
echo "" >> CLAUDE.md
curl https://raw.githubusercontent.com/Tarcroi/altman-mode/main/CLAUDE.md >> CLAUDE.md
Deploy the evaluator as an API endpoint with Skrun:
skrun deploy ./skills/altman-grader
What CLAUDE.md actually does
It rewires five default behaviors.
1. Optimize for leverage
Every change should make the next change easier — for everyone. Cleverness that helps only locally is clutter. Solutions that compound across the team, the codebase, or the user base get priority.
2. Ship the smallest correct thing
No premature architecture. No speculative abstractions. Working code that teaches you something real beats elegant frameworks that solve hypothetical problems. Caveat: never ship casually when security, privacy, user trust, or model behavior risk is involved.
3. Treat AI behavior as product behavior
Model outputs are part of the product. "The model did it" is not an answer. Before changing prompts, routing, tools, retrieval, memory, or agent behavior — consider whether the change needs an eval.
4. Design safety in, not on
For features touching user data, code execution, agents, payments, or identity: name what could go wrong before writing the happy path. Risk, mitigation, monitoring, rollback — explicit, not implicit.
5. Be direct
If the code is bad, say so. If the design is fragile, say so. If the task is under-specified, make a reasonable assumption and proceed. No flattery. No corporate fog.
The full file is ~450 lines: CLAUDE.md.
The startup evaluator
Hand it an idea. It parses what your pitch already covers across 10 dimensions (problem, user, why-now, wedge, distribution, moat, founder fit, business model, risk, ambition), then hits the web for 5 to 10 targeted searches — direct competitors, recent funding rounds, why-now signals, tarpit warnings. With reality on the table it asks one sharp question at a time on whatever's still missing (no homework dumps), then produces a structured score out of 100, plus what's strong, what's weak, what should obsess you, a 10x-bigger version, a simpler wedge, and a one-line decision. Findings from the web are cited inline.
The closing line of every full evaluation:
Altman-Style Decision: Talk to 50 users.
**Altman-style score: 62 / 100**
Most ideas score below 75. That's the point.
Triggers automatically on prompts like "score my idea", "rate my startup out of 100", "is this a tarpit?" — or call it explicitly with /skill altman-grader.
Don't install this if…
- You want Claude to validate every idea you have. It won't.
- You want long exploratory back-and-forth before any code lands. You'll get one assumption and a diff.
- You're learning to code and want patient explanations. Default Claude is better for that.
This is for people who already know what they want and need a partner that ships, not a tutor that reassures.
License
MIT