| name | spec-start |
| description | Runs one headless pass that turns an underspecified prompt into a topic directory under docs/plans/ with spec.md and qna.md only, using the same design discipline as an interactive brainstorm but without blocking questions. qna.md uses a flat ## Q-NN log (Kind bullets + optional ### answers; one example ### under Q-01). Use for one-shot spec generation, GitHub-bound planning agents, or when the user invokes spec-start without back-and-forth clarification. For issue URL as canonical prompt plus branch/PR flow, pair with spec-start-github. |
| disable-model-invocation | true |
spec-start (headless spec kickoff)
For GitHub issue as canonical prompt (fetch issue, agent/<n>-spec-* branch, local or harness PR), use spec-start-github together with this skill.
Help turn ideas into fully formed designs and specs without natural collaborative dialogue: run once, infer missing intent from the prompt and repo, and write the results to disk.
Start by understanding the current project context (files, docs, recent commits). Do not ask questions one at a time or wait for approval between sections—when something is unclear, capture it in qna.md as ## Q-NN entries (see qna.md format below).
Hard-gate (spec-start)
Do not invoke any implementation skill, write production code, scaffold a project, or take any implementation action. Output is documentation only inside the topic directory (plus an optional git commit only if the caller explicitly asked to commit).
This applies to every run regardless of perceived simplicity.
Anti-pattern: "This Is Too Simple To Need A Design"
Every topic deserves the same discipline. A todo list, a single-function utility, a config change — all of them. "Simple" topics are where unexamined assumptions cause the most wasted work. The design can be short (a few sentences for truly straightforward topics), but you must still produce spec.md and qna.md, scaled to complexity.
Checklist (headless mapping)
You must complete these in order. Treat them as a single uninterrupted pass—no blocking AskQuestion, no "approve this section," no mid-run wait for the user.
- Explore project context — check files, docs, recent commits relevant to the prompt; follow existing repo structure and naming.
- Assess scope — if the request describes multiple independent subsystems (for example chat, file storage, billing, and analytics in one prompt), flag it immediately in
qna.md as one or more ## Q-NN entries with Kind: decomposition or scope (independent pieces, how they relate, suggested build order). Do not silently narrow scope without recording that in qna.md.
- Resolve "clarifying questions" without the user — for anything you would normally ask one at a time (purpose, constraints, success criteria), infer the best answer from context and record it in
qna.md as ## Q-NN entries (use Kind: assumption with confidence and blast radius, or Kind: open when inference is too weak). Prefer multiple-choice style reasoning internally; do not paste quiz questions to the user.
- Propose 2–3 approaches — with trade-offs and your recommendation; lead with the recommended option and explain why.
- Write the design — in sections scaled to complexity (see Presenting the design below), directly into
spec.md (no separate "chat" design pass).
- Write Q&A —
qna.md as a flat list of ## Q-NN entries (see qna.md format). Include example ### answer headings only under Q-01; other questions stay bullet-only until real discussion adds answers.
- Spec self-review — quick inline check for placeholders, contradictions, ambiguity, scope (see Spec self-review below). Fix issues inline before finishing.
- Stop — do not auto-invoke planning or implementation skills. Optionally note that a follow-up session may produce an implementation plan once reviewers have read
qna.md (and any review artifacts other skills add later, such as comments.md). When reviewers add comments.md, use spec-refine to fold that feedback into spec.md and qna.md.
Process flow (headless)
digraph spec_start {
"Explore project context" [shape=box];
"Scope too large?" [shape=diamond];
"Record decomposition as Q entries in qna.md" [shape=box];
"Infer Q entries for qna.md" [shape=box];
"Propose 2-3 approaches" [shape=box];
"Write spec.md sections" [shape=box];
"Write qna.md" [shape=box];
"Spec self-review (fix inline)" [shape=box];
"Done" [shape=doublecircle];
"Explore project context" -> "Scope too large?";
"Scope too large?" -> "Record decomposition as Q entries in qna.md" [label="yes"];
"Scope too large?" -> "Infer Q entries for qna.md" [label="no"];
"Record decomposition as Q entries in qna.md" -> "Infer Q entries for qna.md";
"Infer Q entries for qna.md" -> "Propose 2-3 approaches";
"Propose 2-3 approaches" -> "Write spec.md sections";
"Write spec.md sections" -> "Write qna.md";
"Write qna.md" -> "Spec self-review (fix inline)";
"Spec self-review (fix inline)" -> "Done";
}
The process
Understanding the idea
- Check out the current project state first (files, docs, recent commits).
- Before simulating "detailed questions," assess scope: if the request bundles several independent subsystems, say so up front in
qna.md using ## Q-NN entries (Kind: decomposition or scope) and outline how to split work. If the project is too large for one coherent spec, still produce one spec.md for the slice you can cover, and make the boundary explicit in qna.md.
- For appropriately-scoped topics, internally focus on purpose, constraints, and success criteria; express conclusions in
spec.md and uncertainties as ## Q-NN rows in qna.md.
Exploring approaches
- Propose 2–3 different approaches with trade-offs.
- Present options with your recommendation and reasoning.
- Lead with your recommended option and explain why.
Presenting the design (written, not interactive)
- Once you have a working picture of what is being built, write the design into
spec.md.
- Scale each section to its complexity: a few sentences if straightforward, more if nuanced (on the order of a couple hundred words per heavy section when needed).
- Do not ask after each section whether it looks right; instead, use self-review and
qna.md to catch gaps.
- Cover, as applicable: architecture, components, data flow, error handling, testing.
Design for isolation and clarity
- Break the system into smaller units that each have one clear purpose, communicate through well-defined interfaces, and can be understood and tested independently.
- For each unit, you should be able to answer: what does it do, how do you use it, and what does it depend on?
- Can someone understand what a unit does without reading its internals? Can you change the internals without breaking consumers? If not, the boundaries need work.
- Smaller, well-bounded units are easier to reason about and edit reliably. When a conceptual unit feels too large, that is a signal to split the design description or call out follow-up work in
qna.md.
Working in existing codebases
- Explore the current structure before proposing changes. Follow existing patterns.
- Where existing code has problems that affect the work (for example a file that has grown too large, unclear boundaries, tangled responsibilities), include targeted improvements as part of the design—the way a good developer improves code they are touching.
- Do not propose unrelated refactoring. Stay focused on what serves the current goal.
Topic directory layout
Pick a short topic slug from the user prompt (kebab-case, lowercase). Use today’s date in YYYY-MM-DD form (UTC if no timezone is given).
Create a topic directory (create parent directories as needed):
docs/plans/YYYY-MM-DD-<topic>/
The repo may already contain flat Markdown files under docs/plans/; this layout is the convention for new topics produced by this skill so each topic can grow a folder (spec.md, qna.md, optional assets) without colliding with those files.
Within it, for this skill’s initial pass, you must create these files (each ending with a newline):
| File | Role |
|---|
spec.md | Canonical design / spec for the topic. |
qna.md | Flat ## Q-NN Q&A log: assumptions, open points, scope, and decomposition—each as its own question row; answers are optional ### children (see qna.md format). |
You may add optional files in the same directory when visualization helps (for example architecture.svg, state-machine.png, or other assets). Reference them from spec.md or qna.md with paths that work in a GitHub-style file view (repo-relative paths from the repo root, or paths relative to the topic directory—pick one convention per topic and use it consistently).
Do not create, edit, or seed comments.md. That avoids accidental commits of review scratch space before another workflow defines it. spec-refine processes an optional comments.md reviewers add in the same topic directory (see that skill for format and workflow).
Other future artifacts for the same topic (appendices, plan.md, exports, comments.md once owned elsewhere) should live in the same directory so one topic stays in one place.
spec.md — required sections
Use clear prose; adapt depth to scope.
- Title and one-line status (for example
Status: Speculative (spec-start; headless)).
- Context — why this exists; link to issues or PRs if identifiers appear in the prompt.
- Goals and Non-goals — bullet lists.
- Architectural approach — 2–3 options with trade-offs, one recommended path, and what is explicitly deferred.
- Further sections as needed: components, data flow, error handling, security or threat notes if relevant, testing strategy, rollout.
- References — paths, prior docs, external links.
- Q&A index — short bullets, each pointing to a
Q-NN heading in qna.md (same anchor as ## Q-NN in that file). Prefer entries whose Kind is open or that otherwise need reviewer attention; avoid long duplication of qna.md body text.
qna.md format
qna.md is a single flat stream of entries. Do not use top-level sections such as “Assumptions” / “Open questions” / “Scope”—those concerns are expressed only as separate ## Q-NN — Short title blocks (prefer real question titles). Stable IDs run Q-01, Q-02, … in document order.
Heading outline
- Optional
# title line at the top (for example # Q&A — <topic>). If you skip it, the file starts at ##.
- Each tracked item is
## Q-NN — Short title only—no extra ## band headings for grouping.
- Each answer from discussion (GitHub review threads, follow-ups, etc.) is a
### … heading under that question, with unique slug text in the heading so anchors stay distinct. Use sibling ### blocks for multiple answers; do not use a bold pseudo-section title or horizontal rules as the primary way to separate answers.
- Rich answer bodies may use lists, fences, Mermaid, and—when needed—
#### inside a given answer for internal structure.
Bullets under each ## Q-NN (not headings—keeps the outline shallow)
- Kind: one of
assumption | open | scope | decomposition (extend this list only when this skill is updated).
- Detail: the substantive statement (one tight paragraph or a few bullets).
- Optional Context (or fold into Detail): why it matters, blast radius, suggested resolution—whatever applies.
Kind-specific bullets (include when relevant)
assumption: Confidence: high | medium | low — Blast radius if wrong: …open: Suggested resolution: … (spike, owner, policy choice).scope / decomposition: Proposed handling: … (split order, deferral, boundary).
Answers (### children)
- Do not add answer
### headings under every ## Q-NN on the initial pass—that makes the file noisy before any discussion.
- Example only on
Q-01: under ## Q-01, add exactly one illustrative ### answer block. Label the example body so reviewers know it is a template (for example a short note that the block demonstrates shape, optional attribution, and optional permalink-on-next-line—not project truth).
- Additional real answers: on any
## Q-NN, append new ### Answer — … headings with unique slugs (for example ### Answer — @handle — YYYY-MM-DD or ### Answer — PR #123 review). If the permalink is long, put it on the line immediately under the heading.
- Chronology: when mirroring GitHub threads, prefer oldest
### first, newest last under the same ## Q-NN.
Ordering
- Prefer a logical order (for example scope/decomposition first, then assumptions, then open)—not mandatory.
- Do not insert extra
## section headers to group kinds.
Minimal example (structure only; replace with real topic content)
# Q&A — example-topic
## Q-01 — Are we assuming feature X ships behind a feature flag?
- **Kind:** assumption
- **Detail:** …
- **Confidence:** medium
- **Blast radius if wrong:** …
### Answer — example only (template)
_(This `###` block demonstrates the answer shape. Delete or replace when recording real discussion.)_
Optional permalink line: `https://github.com/org/repo/pull/1#discussion_r…`
## Q-02 — What is explicitly out of scope for this spec?
- **Kind:** scope
- **Detail:** …
## Q-02 has no ### children until real answers exist.
Diagrams, Markdown visuals, and optional asset files
Assume reviewers will read the topic in a GitHub-like environment: Markdown is rendered with syntax highlighting, tables, task lists, and common extensions such as Mermaid (standard fenced code blocks with the mermaid info string), which is enough for many architecture, sequence, state, and dependency pictures. Prefer embedding those diagrams directly in spec.md or qna.md when they clarify options, trade-offs, decomposition, or Q&A items—especially where a diagram replaces a long ambiguous paragraph. For qna.md, prefer large Mermaid or code blocks inside a specific answer’s ### body when the graphic belongs to review discussion; keep spec.md as the canonical design surface when both apply.
When inline Markdown is not enough, add extra files in the topic directory (for example SVG, PNG, or other diagram or image formats your hosts display well) and link to them from spec.md or qna.md. Keep filenames stable and descriptive; mention in prose what each asset is for so the spec stays understandable in plain diff views.
Browser or live UI tools remain optional: when mockups, running apps, or side-by-side comparisons would materially improve accuracy—and such tools are available—you may gather evidence without blocking on the user. Summarize what you observed in spec.md or qna.md, or add a ## Q-NN row (Kind: open) describing what still needs visual verification if you cannot capture it.
Do not add diagrams for their own sake: use them where they reduce ambiguity. For purely textual tradeoffs or scope lists, prose is enough.
Spec self-review
After writing spec.md and qna.md, look at them with fresh eyes:
- Placeholder scan: Any "TBD", "TODO", incomplete sections, or vague requirements? Fix them or capture them as
## Q-NN entries in qna.md with explicit uncertainty.
- Internal consistency: Do sections contradict each other? Does the architecture match the feature descriptions? Do
spec.md and qna.md agree? If you added Mermaid or links to sibling files, do blocks parse and paths resolve from the repo root as intended?
- Scope check: Is this focused enough for a single implementation plan, or does it need decomposition (tracked as
## Q-NN rows with Kind: decomposition or scope)?
- Ambiguity check: Could any requirement be read two ways? If so, pick one for
spec.md and record the discarded interpretation in qna.md as a ## Q-NN row, or leave the requirement out of spec.md until resolved and track it in qna.md.
qna.md shape: Are Q-01, Q-02, … IDs unique and in order? Does Q-01 include exactly one example ### answer block per this skill? Do real answer headings use unique slug text? When permalinks appear, do they look well-formed?
Fix any issues inline. No need to loop self-review more than once unless edits reintroduce problems.
Commit
Commit only if the user or automation explicitly asked to commit. Otherwise leave changes in the working tree and report paths.
Key principles (carried over, adapted)
- YAGNI ruthlessly — remove unnecessary features from all designs.
- Explore alternatives — always propose 2–3 approaches before settling.
- Be explicit about uncertainty —
qna.md replaces incremental validation with humans during the run; leave room for other skills to attach review-oriented files in the topic directory later.
Aftercare (message to user)
In the final reply, list the topic directory path, summarize the recommended approach in one paragraph, and surface the top three ## Q-NN items from qna.md that most need reviewer attention (for example Kind: open, or high blast-radius assumptions). If you added optional diagram or image files, list them too so reviewers know what to open. After comments.md exists, spec-refine is the skill that merges review feedback back into spec.md / qna.md.
