Record proof that a task works by capturing commands, pass/fail results, changed files, artifacts, risks, and the next safest task.
Implement exactly one planned slice after scope is frozen, then run relevant verification, summarize changed files, and stop.
Limit an agent's blast radius before implementation by defining allowed files, read-only areas, forbidden operations, commands, and stop conditions.
Convert a mini-spec into 3-7 vertical implementation slices when a small project needs observable, independently verified progress.
Run a lightweight GO or NO-GO gate before using small dashboards, ML workflows, agent workflows, scheduled jobs, or decision-impacting outputs.
Clarify ambiguous AI-engineering project ideas before specs or code by challenging vague terms, identifying assumptions, and updating durable context.
Debug failures systematically by freezing feature work, reproducing the issue, testing hypotheses, applying the smallest fix, and rerunning verification.
Preserve discovered bugs as durable BUGS.md entries with reproduction steps, expected behavior, hypothesis, status, verification, and resolution.
Add focused deterministic tests, fixtures, golden examples, or smoke checks for workflows that can run while still being wrong.
Create a compact SPEC.md for fast solo AI-engineering projects when the goal is clear enough to define behavior, constraints, and verification without full PRD ceremony.