Prompt Library

Replace bracketed fields. Good prompts describe outcomes and evidence; they do not micromanage every keystroke.

For multi-step feature work, use the staged prompts in Spec Map Build Prove Learn instead of asking one prompt to build an entire product.

Implement a feature

Implement [outcome] for [user]. Inspect the relevant code first. Preserve [constraints]. Keep the change narrowly scoped and follow existing patterns. Acceptance criteria: [criteria]. Verify with [commands/manual checks]. Report changed files, verification results, assumptions, and remaining risks.

Diagnose without changing code

Diagnose [symptom]. Do not modify files. Reproduce if safe, trace the relevant path, rank hypotheses, and cite concrete evidence from code/logs. Explain the root cause, affected scope, and smallest reasonable fix.

Fix a bug

Fix [bug]. First reproduce it or establish a failing test. Identify the root cause rather than masking the symptom. Add regression coverage, implement the smallest safe fix, run relevant checks, and summarize why the test would fail before the fix.

Review code

Review [diff/branch/files] for correctness, security, regressions, and missing tests. Prioritize actionable findings by severity. For each finding, name the exact location, failure scenario, and minimal correction. Do not summarize unchanged code; say explicitly if no findings remain.

Refactor safely

Refactor [area] to improve [quality] without changing observable behavior. State the behavior that must remain invariant. Add characterization tests if coverage is weak. Make incremental edits and verify after each coherent stage. Avoid unrelated cleanup.

Explain unfamiliar code

Explain how [feature/path] works from entry point to side effects. Build a call/data-flow map, identify state and invariants, then explain one realistic example. Flag uncertainty and cite file locations. End with the three best places to modify for [goal].

Design a system

Design [system] for [scale/users]. Clarify assumptions, functional requirements, and quality attributes. Compare at least two viable approaches. Cover data model, APIs, failure modes, security, observability, deployment, cost drivers, and evolution. Recommend one approach and explain its tradeoffs.

Create tests

Create tests for [behavior]. Derive cases from public contracts and failure modes. Include happy path, boundaries, invalid input, state transitions, and one regression case. Prefer deterministic tests and minimal mocking. Run them and report what remains untested.

Learn a topic

Teach me [topic] for the purpose of [project]. Start with a mental model, then a minimal example, common failure modes, and tradeoffs. Ask me to predict outcomes in three exercises. Finish with a small project and a rubric for checking my understanding.

Turn a session into knowledge

Convert this work session into durable notes. Extract: problem, root cause, mental model, generalizable pattern, failed approaches, verification, and follow-up questions. Remove chat-specific filler and secrets. Suggest links to existing concepts and one spaced-repetition question.