--- name: Generate Spec V2 slug: generate-spec-v2 category: AI Engineering description: Generate Spec V2 creates a code-based implementation spec in specs/ for a feature, fix, or refactor. Use it when planning work that needs Mermaid flows, key types, phased diffs, and checks before implementation. github: "https://github.com/saffron-health/libretto/tree/main/.agents/skills/generate-spec-v2" language: TypeScript stars: 877 forks: 67 install: "npx degit https://github.com/saffron-health/libretto/tree/main/.agents/skills/generate-spec-v2 ~/.claude/skills/generate-spec-v2" installs_to: ~/.claude/skills/generate-spec-v2 source_path: .agents/skills/generate-spec-v2/SKILL.md collection_size: 23 category_size: 2451 collection_url: "https://dirskills.com/collections/saffron-health/libretto" added: 2026-08-22T05:21:18.787Z last_synced: 2026-08-22T05:21:18.787Z canonical_url: "https://dirskills.com/skills/generate-spec-v2" --- # Generate Spec V2 Generate Spec V2 creates a code-based implementation spec in specs/ for a feature, fix, or refactor. Use it when planning work that needs Mermaid flows, key types, phased diffs, and checks before implementation. **Install:** ```bash npx degit https://github.com/saffron-health/libretto/tree/main/.agents/skills/generate-spec-v2 ~/.claude/skills/generate-spec-v2 ``` ## README # Generate an implementation spec Write a spec in `specs/` that an engineer can implement without repeating the research. Keep it lean, grounded in the current code, and focused on the requested result. ## Research the change 1. Read the request and repository instructions. 2. Find the entry points, key types, state boundaries, callers, tests, and commands tied to the change. Trace each changed runtime path far enough to show an accurate diff. 3. Read external docs only when a dependency or API affects the design. Prefer official docs and record the links used. 4. Ask a question only when the answer would change the result and the code cannot answer it. Otherwise, make the narrowest sound assumption and record it. Use the amount of research the task needs. Do not impose a fixed research process. ## Choose the scope - Plan the smallest complete change that gives the requested result. - Do not add abstractions, config, infrastructure, migrations, or cleanup unless the result needs them. - State goals and non-goals from the request and code. Do not pause for confirmation when the scope is clear. - Make each phase fit one commit and leave the repository working. - Use as many phases as the work needs. Do not set a phase or line-count limit. ## Write the file Choose a short kebab-case name and create `specs/.md`. Use this structure: ````markdown # ## System flow Put one or more Mermaid diagrams immediately after the title, before all prose. Show the main runtime flow and changed parts. Label current and proposed paths or use separate diagrams when that is clearer. Keep node text short and use valid Mermaid syntax. ```mermaid flowchart TD A[Entry point] --> B[Service] B --> C[Observable result] ``` ## Problem overview Explain the current problem and why it matters in a few plain sentences. ## Solution overview Explain the proposed change and its key design choice in a few plain sentences. ## Goals - State the user-visible or system-level results that must hold. ## Non-goals - State what this spec leaves out. ## Important files, docs, and websites - [`path/to/file.ts`](../path/to/file.ts) — State what the implementer will change or learn here. List only sources that help implement the change. ## Implementation ### Phase 1: Explain the intent and what becomes true after this phase lands in one or two sentences. #### Important types ```ts // path/to/types.ts type ImportantInput = { id: string }; type ImportantResult = | { status: "ok"; value: Value } | { status: "error"; reason: FailureReason }; ``` #### Call stack diff Show how this phase changes the current call path. Keep the entry point and enough parents to make ownership clear. ```diff requestHandler -└── existingService - └── dataStore +└── validateInput + └── existingService + └── dataStore ``` #### Code diff preview Show a short unified diff of the main edit. Use real file and symbol names, enough surrounding code to place the change, and `...` for parts the implementer will fill in. Do not try to write the full patch. ```diff // path/to/handler.ts async function requestHandler(input: ImportantInput) { - return existingService(input); + const valid = validateInput(input); + return existingService(valid); } ``` - [ ] Make one concrete implementation change, with file and symbol names. - [ ] Wire the change into its nearest caller or consumer. - [ ] Smoke the main failure case or boundary by hand. Do not commit this check until the feature is package-level end-to-end testable. - [ ] Run the exact command that proves the phase works. ```` ## Phase rules - Give every phase four or five checklist steps, including checks. - Keep every phase small enough for one clear commit and leave the codebase working. - Make each phase produce visible or testable progress. - Name exact files, symbols, behavior, and commands when the codebase provides them. - Include `Important types`, `Call stack diff`, and `Code diff preview` in every code phase. Keep them inside that phase; do not collect call stacks or code diffs in a global section. - Show inputs, outputs, state, events, errors, or unions that set the phase contract under `Important types`. Use the project language. - Make the call-stack diff start from the current path and mark the proposed path with unified diff signs. For UI work, a component render or event-handler path counts as the call stack. - Make the code diff a short preview of the main edit, not a full patch. Include a file path comment and preserve useful surrounding control flow. - Use `Not applicable — no code path changes` only for a true docs, data, or config phase. Do not invent types or call paths. - Follow [Testing](#testing) for what to check and when to commit tests. - Avoid setup-only or refactor-only phases unless later work cannot land safely without them. Fixture setup for package-level tests is allowed once the feature is end-to-end testable. ## Testing Specs commit only package-level end-to-end tests. Act and observe the way a user of that package would: - Electron / web UI: start the app separately when needed, drive the live renderer with Playwright through `pnpm halo-web`, and assert visible elements, roles, labels, and text. - Services and other APIs: call the public methods, then read results through the same public API or another real collaborator a user of that package would use. Do not use mocks such as `vi.fn`, `vi.mock`, or hand-rolled fake collaborators. Do not spec internal unit tests, or assert implementation details such as internal file layouts or exact formatting of private outputs. If a package-level end-to-end test is hard to build and none already exist for the area, do not add a lower-level test instead. Committed tests must read like end-user code or interactions: short, easy to follow, and free of setup noise. Put shared setup and teardown in Vitest fixtures (`test.extend`), not ad-hoc helpers or manual cleanup. See the [Vitest fixtures documentation](https://vitest.dev/guide/test-context.html#test-extend). Until the feature is end-to-end testable at the package, each phase still includes a check. Write that check as a smoke step the implementer runs by hand and does not commit: - [ ] Smoke the main failure case or boundary by hand. Do not commit this check. Once the feature is end-to-end testable, add the fixtures needed for those high-level tests, then commit the tests. Make fixture setup its own phase when it is more than a small add-on; fold it into the phase that first makes the feature testable when it is small: - [ ] Add Vitest fixtures that set up the package the way a user would. - [ ] Commit a short high-level test that acts and observes through the public API or live UI. ## Final check Confirm that Mermaid diagrams appear only at the top and match the plan; each code phase has key types, an accurate call-stack diff, a code-diff preview, and four or five steps; links and commands are real; committed tests are package-level end-to-end and earlier phases use uncommitted smoke checks; and the full plan covers every goal without pulling in a non-goal.