AI Agent Workflow Collection

Apply a bounded capability-harness control protocol to substantive, quality-sensitive, current, version-specific, project-dependent, or multi-step work. Route only the smallest useful context-enrichment, evidence, alternative, observable-verification, or evaluation module after the task contract is clear. Do not use for routine direct tasks, cosmetic variants, fixed agent pipelines, or work already owned by another controller.

Systematically diagnose unclear product behavior or root cause by building a feedback signal, testing one hypothesis at a time, and verifying the original symptom. Use for intermittent or flaky symptoms, cross-component failures, performance regressions, regressions after recent changes, or bugs that survived previous fixes. Do not use for obvious direct failures, primary test-design or wait-strategy problems, Claude Code runtime issues, unresolved architecture boundaries, or explicit review/readiness checks.

Create, update, or resume a durable multi-session map of unresolved decisions and their dependencies. Use when the user clearly asks for a decision map, wants to track an open decision frontier across sessions, or wants to resume decision work by ticket. Do not use for ordinary implementation planning, one-session design comparison, progress summaries, handoff or checkpoint state, or direct execution of settled decisions.

Use when the user asks to make or compare a codebase architecture decision involving module boundaries, ownership, interfaces, seams, adapters, domain language, or a throwaway prototype that answers a design question. Also use during implementation only when a non-obvious architecture, ownership, interface, or dependency-boundary decision blocks the next safe edit; a testability concern qualifies only when it demonstrates that kind of boundary pressure. Do not use for ordinary code explanation, straightforward implementation, general test design, flaky tests, mocks, regression coverage, or implementation of an already selected design.

Use when the user clearly asks to record, update, prune, or consult project-governed Markdown reference lessons about repeated mistakes, corrections, or confirmed approaches that must be versioned, shared, reviewable, or otherwise tied to repository history. Do not use for automatically loaded CLAUDE.md or .claude/rules instructions, host auto memory, handoff state, or decision-frontier planning.

Use when the user explicitly challenges the reliability of the agent's current or prior conclusion, evidence, source use, active stage, wrong direction, strategy or workflow drift, stale context, unsupported confidence, hallucination, guessing, source-vs-memory confusion, or example-vs-task confusion, or explicitly asks to reassess one of those concerns. Also use when the user says the agent used the wrong, missing, unread, or outdated source for its current or prior conclusion. Do not use for ordinary status questions, file-inventory questions, routine source-reading requests, general uncertainty, ordinary strategy or approach preference changes, or preventive stage reminders without a stated reliability concern.

Use when the user explicitly asks to review code, assess or address review or PR feedback, verify whether completed software work is done, fixed, passing, or ready, or check a development artifact before sending or shipping. Also use before a done claim for completed changes affecting persisted data or migrations, authentication, authorization, permissions, public compatibility contracts, concurrency, transactions, or destructive behavior. Do not use for ordinary progress summaries, implementation-only requests, high-risk terminology without a completion claim, or an explicitly invoked bundled /code-review.

Use when the user clearly asks to refactor, migrate, rewrite, de-duplicate, or clean up an existing coding-agent instruction artifact, such as a SKILL.md, Claude Code prompt, CLAUDE.md fragment, or equivalent agent configuration, while preserving its intended behavior and trigger boundary. Do not use for ordinary code refactors, code or document review, planning, general-purpose writing, translation, or ordinary prompt and copy editing.