Answer Contract — AI skill for Claude Code

answer-contract

**Four rules that stop your AI agent from padding, rambling, and shipping one mediocre answer.**

`Implicit-Need Fill` · `Structured Output` · `3-Option-Plus-Premium` · `Assumption-Circuit-Breaker`

A single-file skill for coding agents (Claude Code, Codex, OpenCode, Pi, and anything that reads `SKILL.md`). Works as a standalone output contract — and composes well with style-level skills like [i-have-adhd](https://github.com/ayghri/i-have-adhd): that skill shapes sentences, this one shapes documents and decisions.

Why

AI output fails in four predictable ways: it answers the question you *typed*, not the one you *meant*; it rambles; it ships one path when you needed a comparison and a recommendation; it sounds confident while guessing — and nobody knows what was assumed or what was never checked.

The four rules

Rule	Fires when	Effect
Implicit-Need Fill	every substantive request	The 2–5 unstated needs behind the request are surfaced and labeled (assumption:), or confirmed before full execution
Structured Output	every response	Conclusion first; tables over prose; no opener, no recap, no closer
3-Option-Plus-Premium	non-code tasks with >1 defensible approach	≥3 options compared across named dimensions + exactly one recommendation + one premium path
Assumption-Circuit-Breaker	before sending	Self-attack the conclusion, list assumptions, state coverage (N/N verified), pre-commit a fallback and kill signal

Worked case 1 — a code task (executed, not claimed)

Prompt: *"Count files and total size per repo in a vendor directory, compare them."*

**Default single-pass behavior:** 20 lines of plausible-looking code that was never run, wrapped in "this should work" — no numbers, no verification.

**With the four rules:** conclusion first — `6 repos, 6,399 files, 203.91 MB`, top repo 83.89 MB; the exact 10-line script; coverage stated as `6/6 directories, full scan`; assumpti