メニュー
Extract every constraint buried in the task text — limits, exclusions, format rules, 'must/only/never' clauses — into an explicit list, and check the plan or draft against each one.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Check a deliverable against its actual consumer: who reads it, what they do next, what format and register they need, and what they will assert downstream on its basis. Activate when the deliverable has a named audience beyond the requester, or its text will be forwarded, announced, or acted on by someone absent from the conversation.
For any rule-like deliverable — code, regex, query, formula, validator, policy — enumerate the boundary inputs (empty, zero, one, max, duplicate, malformed, hostile) and actually exercise the most breakable three before shipping. Activate when the deliverable is executable or rule-like and will meet inputs the author didn't type.
State every claim at its actual evidence level — verified, inferred, assumed, guessed — and give credences only where an observable could score them; ban certainty language on unverified claims. Activate when a response makes claims about unobserved, future, or unverified state, or when the user asks how confident or likely something is.
Before intervening in a system, write the explicit causal chain from the proposed action to the intended outcome — every arrow marked verified or assumed — and check the weakest assumed arrow first. Activate when an intervention (fix, optimization, policy change, growth action) is proposed on a system whose mechanism has not been written down.
With multiple live hypotheses and a limited test budget, design the cheapest single test whose possible outcomes separate the hypotheses — with a results-to-verdict table committed before running anything. Activate when two or more competing hypotheses are on the table and the next step is choosing what test, experiment, or probe to run.
Enumerate every deliverable the request actually asked for and map each one to the work produced, so nothing asked-for is silently dropped, stubbed, or rounded up to done. Activate when a request contains multiple parts — an enumerated list, several 'and'/'also' clauses, or a spec with more than one required artifact — or before delivering a draft against such a request.
| name | hidden-constraint-scan |
| description | Extract every constraint buried in the task text — limits, exclusions, format rules, 'must/only/never' clauses — into an explicit list, and check the plan or draft against each one. |
| license | MIT |
Trigger (observable): The task text contains restrictive language (must, only, never, without, except, at most, do not) or is long enough (>~150 words) that mid-prompt constraints are likely to fall out of working memory.
When NOT to activate: Short requests (<~50 words) already restated in full; constraints already extracted into a checklist this session; brainstorming where no deliverable is being checked.
A constraint table: verbatim quote | hard/soft | status. Violations named before the deliverable is presented.
Known risk: Inflating preferences into hard constraints and over-constraining the solution. Mitigation: the hard/soft classification step forces the distinction.
Max intended cost: ≤300 added output tokens; one full re-read of the task text.
Evidence status: EXPERIMENTALLY_TESTED — this repair content lifted a weak model on procedural tasks in controlled runs; the effect is task-concentrated, and automatic router selection of micro-skills is a separate, unproven layer.
Lineage: Derived from two documented reasoning-failure modes — ignoring stated boundaries mid-task, and locking onto the first mental representation of a task while later clauses go unread — combined with an evidence-backed completion-integrity principle of partitioning the request into checkable parts.
