work-skill AGENTS.md

Work project rules

Read skills/QUALITY.md before changing this project or any nested Work skill. Treat it as the project's quality bar.

• Keep every focused skill under skills/work-* self-contained. Do not reference ../QUALITY.md, sibling skills, or files outside its own directory; focused skills may be installed independently.

• Quality software must not break at real boundaries, must not demand attention without a reason, must know its limits, fix failures quickly, and respect human attention and authorship.

• Practice empathy, focus, and impute: understand the user's real context, eliminate unimportant opportunities, and present every surface and handoff with care that truthfully reflects the quality underneath.

• Treat taste as trained judgment: study references, use enduring fundamentals, invite critique, balance usability with character, and explain qualitative decisions beyond personal preference or current trends.

• Restore deliberate friction where cheap implementation would bypass judgment: require a real problem, user, validation question, and kill or selection criterion before turning generated options into product surface.

• Never preserve backwards compatibility. Prefer a clean break, update all in-scope consumers together, and remove obsolete aliases, adapters, wrappers, flags, and migration shims. If an external contract cannot be removed, stop and escalate; do not invent an internal compatibility layer.

• Choose the simplest implementation that fully meets the current requirements. Avoid speculative abstractions, configuration, and indirection.

• Grow the system in layers. Start from the smallest version that works end to end, and add each new capability on top of a product that already works. Never trade a working product for unfinished complexity.

• Keep components modular and concerns clearly separated.

• Prefer established, well-maintained libraries when they reduce overall complexity or improve reliability. Do not reimplement common functionality without a clear reason.

• Lean on the dependencies already in the project before writing your own implementation or adding packages. Do not assume a library lacks a capability without checking its documentation and types.

• Make architectural decisions for the long term. Do not accept a stopgap that only works for now and is meant to be replaced later.

• Keep user data, recoverability, security, and truthful failure behavior safe; those are not compatibility shims.

• Keep agents bounded executors. The parent owns taste, decisions, integration, user-facing copy, and proof.

• Before work begins, state the exact goal, source of truth, acceptance checks, boundaries, and what does not count. Treat concrete artifacts and validation evidence as the only completion proof.

• Make plans self-contained for a lower-capability executor with no hidden context: resolve judgment-heavy choices and name exact files, symbols, ordered changes, invariants, failure cases, validation commands, and expected evidence.

• For test work, inspect implementation, existing tests, public APIs, and actual user flows before writing anything. Add only deterministic tests for observed behavior, provide the four-line test plan first, and report tests deliberately omitted and any real bug discovered.

• For parallel investigation, diversify by approach family, preserve independent hypotheses early, track evidence and blockers, and challenge exact contracts and failure boundaries. Reopen blocked routes only for a materially new mechanism, within the user's scope and token budget.

• Once an agent starts, let it run until it reports complete, blocked, or needs user input. Never kill it just because a poll, tool call, or arbitrary wall-clock interval elapsed; interrupt only for explicit cancellation/goal change, safety, duplicate ownership, an approved resource limit, or clear repeated no-progress evidence, and record the reason.

• Treat discovery validation as structural only. After changing workflow behavior, forward-test the affected skill with a realistic request in a disposable task and inspect its raw launch, goal, handoff, and read-only boundaries; routing or goal-lifecycle changes require representative $work, $work-plan, and $work-checklist smoke tests before release.

• Use Work / 2. Ideate for an interactive, question-by-question sparring session; it does not spawn agents or edit files, and it must end in a sharper decision or test rather than an idea dump.

• Use Work / 4. Test for direct meaningful-test implementation: inspect real behavior first, publish the four-line test plan, add only justified deterministic tests, and optimize for confidence rather than coverage; it does not spawn agents.

• Use Work / 5. Concern for a compact, read-only implementation sanity check; it reports only evidence-backed concerns and unknowns, never edits or a full fix plan.

• Use Work / 6. Checklist to capture an ideated direction, plan, requirements, or concerns as accountable items. Capture mode does not create a goal or edit; Execute mode registers one parent-owned goal and works every item without agents.

• Use Work / 9. Quality when the standard should be passed into a conversation; it never spawns agents or edits files.