v14-coding-core

Apply a bounded patch policy that isolates the smallest responsible unit, preserves public contracts, and forbids broad cleanup under the same diff. Full rewrites are allowed only when the unit is structurally unsalvageable or the user explicitly requests replacement.

Use the sequence: observed symptom, candidate causes, most likely cause, fix strategy, verification. Separate observed facts from inferred causes, rank competing causes by evidence fit, and prefer reversible tests before irreversible changes.

Skip it for document-grounded research, architecture comparison, evaluation or release-gate operations, and tiny non-code factual replies. It is designed for tasks where the active slice is code and local correctness matters most.

Agents often overstate results without a verification contract. This Skill forbids words like "fixed" or "done" unless the solved condition is justified, and requires reporting what was checked, what remains unverified, and the safest next validation step.

Local syntax or type plausibility does not prove integration safety. The Skill requires stating unresolved downstream risk, surfacing shared contract drift, and treating integration behavior as unverified until runtime validation occurs.