← ClaudeAtlas

spec-probelisted

Run before implementing anything non-trivial: surface omitted edge cases and must-NOT constraints into a verification-tiered TASK_CONTRACT.md, so the verifier has something to check against. Use when starting a feature, a bugfix with unclear intended behavior, or any task where a reviewer would have to guess what correct means. Also use when asked to spec-probe, write a task contract, or check what the spec left out.
ucsandman/claude-skills · ★ 5 · AI & Automation · score 75
Install: claude install-skill ucsandman/claude-skills
# Spec probe A verifier catches only what the spec named. When an obligation was never stated and no convention determines it, the comparison is vacuous — any behavior is "consistent with the spec," and the verifier passes the defect at ~93% confidence. Measured: 100% confident false pass with the edge omitted, 98% catch once the same edge is written into the spec (n=210, Wilson 95% CI). Re-running a bigger model never closes that gap; a ~30x cost increase buys no additional recovery. So do not strengthen the reader. Widen the artifact, before code exists. Output is one file: `TASK_CONTRACT.md` at the repo root (or beside the feature's plan). Schema and worked example: `references/contract-schema.md`. Read it before writing one. ## Steps 1. **Enumerate requirements.** One line each, from the request plus whatever spec exists. Terse prose is the failure mode here — a requirement with no discernible shape yields no edges. Expand a one-liner into its actual obligation before classifying it. 2. **Shape-classify each requirement** as one or more of: `numeric-range`, `collection`, `text`, `stateful`, `io`. Do this yourself as a reading task. Do NOT delegate it to a regex or a keyword table: a deterministic classifier under-fires on roughly half the applicable edges. If you delegate it to a subagent, that subagent is sonnet or better. 3. **Run the eight-category edge checklist** against each shape. Ask every one out loud; a category you skip is a category yo