answer-contract

Part of the Evidence-first Agents suite — tooling that makes AI agents accountable instead of just capable: answer-contract (output discipline) · skill-spec (spec discipline) · memory-wiki (memory discipline) · skill-os (the assembly line). Same author, same zero-dependency philosophy.

The output-discipline layer for AI coding agents. Four rules that stop your agent from padding, rambling, and shipping one mediocre answer.

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

i-have-adhd shapes your agent's sentences; answer-contract shapes its documents and decisions. Install both and the whole output chain is covered.

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.

This is not a style preference. Anthropic's own Claude Code system prompt enforces the same discipline at harness level — "Lead with the outcome"; "Being readable and being concise are different things, and readable matters more"; outcomes reported faithfully. Your agent already wants these rules. This skill makes them explicit and auditable.

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; assumptions labeled (disk-actual sizes, symlink errors skipped — occurred 0 times); fallback pre-committed (numbers mismatch → re-run the reproduction command). The table below is the real executed output:

repo	files	size
repo-a	4,135	83.89 MB
repo-b	966	59.60 MB
repo-c	465	46.59 MB
repo-d	576	11.78 MB
repo-e	195	1.63 MB
repo-f	62	0.42 MB
total	6,399	203.91 MB

Worked case 2 — a copywriting task

Prompt: "Write a Xiaohongshu (RED) promo post for a regional specialty noodle brand."

Default single-pass behavior: one emoji-dense paragraph of trend-speak aimed at everyone — which persuades no one, with no alternatives.

With the four rules: three hook-type options compared (homesickness / craft-contrast / late-night-pain) across expected strengths and risks; one recommended; one premium cut (hook hybrid with shot list); assumptions labeled (brand tone set to "modern heritage" — vetoable); a pre-committed kill signal (week-1 CTR < account average −30% → switch to the cheapest-to-produce option).

What's in this repo vs. what's not

• In this repo: the four rules, when to fire each, the pre-send check, two worked cases, install instructions.
• Not in this repo (by design): the task-class implied-need checklists, the full option-matrix templates, calibration data from live usage, and vertical adaptation packs. These are delivered through collaboration — see COLLABORATION.md.

Install (60 seconds)

Single file. Zero dependencies. No telemetry, no network calls, no background processes — your prompts and outputs never leave your machine.

Copy SKILL.md into your agent's skills directory:

git clone https://github.com/chenhz01/answer-contract.git
cp answer-contract/SKILL.md ~/.claude/skills/answer-contract/SKILL.md   # Claude Code
# or the equivalent skills path of your agent

Then invoke with /four-rules, or add a one-line reference to your CLAUDE.md / AGENTS.md. Turn it off anytime with "stop four rules".

Authorship

Produced through human-AI collaboration: the rules and worked cases were drafted by an AI coding agent, then reviewed, edited, and approved by a human maintainer before release. The evidence is executed, not claimed — the code-case table is real script stdout from the same session.

Status

Actively maintained — small, regular releases. See CHANGELOG. On the roadmap: vertical adaptation packs for copywriting and decision-memo workflows (delivered via collaboration, see COLLABORATION.md).

Honest boundary

This skill constrains output shape and decision hygiene. It does not make a weak model smart. Evidence status: the code case above was produced and executed by the same model that ships this skill (single-session controlled A/B, before = unmodified single-pass output) — not an independent blind study; treat the effect sizes accordingly.

License

Documentation and skill text: CC BY-NC-ND 4.0 — attribution required, no commercial use, no derivatives. Explicitly prohibited for use in LLM training, fine-tuning, or distillation without prior written consent.

canary: CANARY-ZS-4R-20260912-8E2D41B7