Open Source / Agent Skill
Socratic
A free skill that makes your AI agent question its own work as it goes. Before it commits to a plan or a fix, it names the assumption most likely to be wrong and tests it against something real, a run, a test, the source, instead of trusting its first guess. Gated and mostly silent, so it stays out of the way on simple work.
npx socratic-methodOne command, every tool. It installs for Claude Code, Codex, Cursor, VS Code, Gemini CLI, Windsurf and the rest in this project at once. Add --global for every project, or --tool cursor to narrow it to one. Full options below.
Heads up: because it thinks a little longer and runs extra checks before it commits, it can use more tokens per session than a bare agent. That is the cost of catching wrong turns early, and it is gated so trivial work stays cheap.
- License
- MIT, free
- Format
- SKILL.md
- Works in
- 11+ agents
What it does
Most agents commit to their first plausible plan and then rationalize it. Telling a model to "double-check itself" barely helps, because a model judging its own reasoning shares its own blind spots. Socratic fixes that with one rule: every doubt has to end in a real check outside the model's own head. The core move is predict-then-run, it commits to an expected result, runs it, and lets the gap between the two redirect the work.
Questions its own work
Before it builds on a plan or a fix, it names the assumption most likely to be wrong.
Tests against reality
It predicts the result, then runs it. The gap redirects the work, instead of a confident guess.
Knows which questions are yours
It checks what can be checked, and asks you only what is genuinely yours to decide.
Install
The fastest path is npx. It copies the skill into the folders your agents actually read, no clone needed, and with no flags it covers all of them. Re-run it any time to update.
npx socratic-methodnpx socratic-method --globalnpx socratic-method --tool cursornpx socratic-method --listTwo directories carry almost the whole ecosystem: .agents/skills/ is the cross-agent path, and Claude Code is the one major tool that reads only its own. So the default writes those two, and Codex, Cursor, VS Code and Copilot, Gemini CLI, Windsurf, Amp and Roo Code all pick the skill up from them. Tools they miss, Kiro, Factory and OpenCode, are added automatically when the installer sees you use them. Run --list to see the exact directory each tool reads.
Prefer not to touch the npm registry? npx github:ergini/socratic-method runs the same installer straight from the repo. Or clone it by hand, it is just files: drop the repo into your tool's skills directory.
In a plain chat UI (Claude.ai, ChatGPT, Gemini) there is no skills folder, so paste SKILL.md into a Project's custom instructions or a saved style.
FAQ
Will it slow the agent down?
No. It is gated and mostly silent. Trivial work gets zero extra checks, normal work gets one quick check on the riskiest assumption, and you see a better result rather than a longer transcript.
Does it use more tokens?
Somewhat, yes. It thinks a little longer and runs a few extra checks before committing to a plan, so a session can cost more than a bare agent. The tradeoff is fewer wrong turns and less rework. Because it is gated, trivial tasks stay cheap, and the extra spend lands on the risky work where it actually pays off.
How is it different from telling the model to double-check itself?
Every doubt has to end in a real external check: a run, a test, the compiler, the source. A model that only doubts itself, with no external signal, is net negative. The core move is predict-then-run: commit an expected result, then run it, and treat the gap as the finding.
Which tools does it work in?
Every tool that reads the SKILL.md format: Claude Code, Codex, Cursor, VS Code and Copilot, Gemini CLI, Windsurf, Amp, Roo Code, Kiro, Factory and OpenCode. One npx socratic-method installs it for all of them at once, so you do not have to pick. In a plain chat UI like Claude.ai, ChatGPT, or Gemini, there is no skills folder, so paste SKILL.md into a Project's custom instructions or a saved style.