BetterThink
RokyYTR2/BetterThink
Força agentes IA a pensar primeiro, terminar uma coisa, depois próxima
Ver no GitHubSkill de disciplina para agentes Claude Code, Codex e Gemini CLI que impõe uma metodologia rigorosa: pensar antes de codificar, terminar uma tarefa por vez com verificação, manter arquivos de memória enxutos e testar UI em três viewports. Transforma hábitos ruins de agentes IA em práticas sistemáticas.
Instalação
npx convez-ai RokyYTR2/BetterThink1 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
- Melhorar qualidade de código gerado por agentes IA
- Reduzir erros por suposições silenciosas e escopo mal definido
- Garantir verificação de UI responsiva antes de finalizar tarefas
- Manter histórico de projeto limpo sem bloat em arquivos de memória
Veredito

Abordagem pragmática que endereça problemas reais de agentes IA: suposições silenciosas, incompletude de tarefas e falta de verificação. A implementação é clara e portável para múltiplas plataformas (Claude, Gemini, Cursor, Copilot), tornando-a útil para equipes que usam agentes em produção.
README
A tradução para o português deste README ainda está na fila.
A skill for Claude Code, Codex, and Gemini CLI that fixes the most common way AI agents fail — not wrong answers, but wrong habits. Asking after coding. Moving on before done. Bloating memory files. Shipping desktop-only UI.
The Problem
AI coding agents are capable. They're just undisciplined.
❌ Default agent behavior
Starts coding immediately. Assumes scope. Moves to the next task before the current one is verified. Updates
AGENTS.mdafter every tiny change. Ships UI without checking mobile.
✅ BetterThink behavior
Asks until 95% confident. One task at a time, verified before moving on.
AGENTS.mdupdated only on meaningful changes. Three viewports checked every time.
Same agent. Better habits.
┌──────────────────────────────────────────┐
│ ASSUMPTIONS BEFORE CODING ████ none │
│ TASKS FINISHED BEFORE NEXT ████ 100% │
│ MEMORY FILE BLOAT ████ none │
│ UI VIEWPORT COVERAGE ████ 3/3 │
└──────────────────────────────────────────┘
What It Does
Six rules. Applied before every coding task.
| Rule | What it fixes |
|---|---|
| Clarify before coding | Agent asks until 95% confident — no silent assumptions on scope, stack, or edge cases |
| One slice at a time | Current task verified (diff read, test run, behavior checked) before the next begins |
| Delegate mechanical work | File search, linting, screenshots → Haiku subagent. Architecture decisions stay on the main model |
| Keep memory files lean | Update AGENTS.md / GEMINI.md / CLAUDE.md only on meaningful changes — new dep, new script, new folder |
| Screenshot-verify UI | Three viewports (375px / 768px / 1440px), every time, until all three look right |
| Supplementary docs on demand | Deep docs written freely, loaded only when the current task needs them |
Install
Claude Code
mkdir -p ~/.claude/skills/betterthink
cp claude/SKILL.md ~/.claude/skills/betterthink/SKILL.md
Codex
mkdir -p ~/.codex/skills/betterthink
cp codex/SKILL.md ~/.codex/skills/betterthink/SKILL.md
Gemini CLI
mkdir -p ~/.gemini/skills/betterthink
cp gemini/SKILL.md ~/.gemini/skills/betterthink/SKILL.md
Cursor
mkdir -p .cursor/rules
cp cursor/SKILL.md .cursor/rules/betterthink.mdc
GitHub Copilot
mkdir -p .github
cp copilot/SKILL.md .github/copilot-instructions.md
The skill activates automatically when the agent detects a coding task. No extra configuration needed.
Supported Agents
| Agent | Folder | Memory file |
|---|---|---|
| Claude Code | claude/ | CLAUDE.md / AGENTS.md |
| OpenAI Codex | codex/ | AGENTS.md |
| Gemini CLI | gemini/ | GEMINI.md / AGENT.md |
| Cursor | cursor/ | .cursor/rules/*.mdc / .cursorrules |
| GitHub Copilot | copilot/ | .github/copilot-instructions.md |
Why
Most AI coding failures aren't a capability problem. They're a discipline problem.
The agent knew how to fix the bug — it just moved on before checking if it worked. It knew you wanted a mobile-friendly layout — it just never opened DevTools at 375px. It knew AGENTS.md was supposed to stay lean — it just logged every minor refactor anyway.
BetterThink is a set of explicit guardrails written directly into the skill prompt. The agent reads them before every task and can't pretend it didn't.
License
MIT