feature-prompt

feature-prompt

Produces the smallest useful handoff for the next step — not an implementation spec. grill-with-field will challenge the plan, sharpen domain terms, update GLOSSARY.md, and offer ADRs only for hard decisions.

Output contract

Final prompts use this shape:

Project:
[Full Project Matrix code, repo, path, or context. Use one line.]

What is needed:
[The change to make. One short paragraph or 2-4 bullets.]

Why it is needed:
[The problem, user pain, business reason, or workflow gap — so grill-with-field
can challenge tradeoffs, not just wording — and what happens if nothing is
built.]

Expected end result:
[Observable done state. Prefer user-visible behavior, passing checks, or a
demo flow; seeds the later acceptance criteria. Name the empty, loading,
error, and recovery behavior, each role that sees it, and one thing that must
not happen.]

Known limits:
[Hard constraints, non-goals, compatibility needs, or exclusions
already known. When intake was split, list the deferred slices as explicit
non-goals. Always include the zero-attribution rule.]

Open questions:
[Conditional. Real unresolved questions for grill-with-field to attack.]

The first five sections are always present; omit Open questions when none remain.

Rules

Zero attribution: never add or leave co-author, AI, or tool attribution in any output. Include this rule in the prompt's Known limits.

Intake and inference

• Start from free-form intake. Infer first; ask only when What is needed is unclear or project/context cannot be inferred safely.
• Use repo evidence when cheap: Project Matrix, cwd, GLOSSARY.md, GLOSSARY-MAP.md, and ADR names. Do not run a broad code scan by default.

Scope and slicing

• Before drafting, apply the fog test: can you state the destination in one line and name every open decision as a sharp question, right now? If not, stop and suggest /wayfinder — it charts those decisions and resolves them one per session. Do not write a PR-sized prompt for foggy work. Fog, not size, is the test: a large mechanical change with nothing left to decide is not fog.
• Default to one thin vertical slice, small enough to review and merge independently. Broad intake gets split: draft this prompt for slice 1 only; Known limits records the deferred slices. A prompt that would force a very long grilling session → ask to split scope before finalizing; user away → split it yourself and note the split at the top of the response.
• When the user intends parallel agent threads, split slices to minimize shared-file coupling.

Context and evidence

• Use the full Project Matrix code verbatim whenever one exists. Never abbreviate project codes or invent shorthand.
• Reference existing files/modules/seams. Put "reuse existing seam vs create new seam" under Open questions only while unresolved; otherwise record the settled choice and its evidence in the contract section that names the seam, never as a new section. If search finds no existing seam, name the terms searched.
• Include only non-obvious context: constraints, architecture quirks, and domain rules the model cannot infer cheaply from repo scans. Omit stack facts the code already shows.
• Dependency internals central and unclear → suggest fetching targeted source (e.g. opensrc) as a follow-up context step before deep grilling.
• Keep domain words intact. Ground candidate definitions in evidence and use the approval flow below; unresolved definitions belong to grill-with-field.

Questions and output discipline

• Do not create Domain terms, Decisions, Dependents, Risks, Doc anchors, Integration, Constraints, or Acceptance sections. Fold useful facts into the six contract sections; user-stated hard decisions go under Known limits.
• All unclear decisions land under Open questions; ungrillable ones ("needs to feel/see it") also route to /handoff + /prototype before deep grilling continues.
• Keep Open questions to the highest-leverage unknowns (usually 1-5). Drop trivia that can be decided during implementation.
• A vague answer to a sharp question ("as fast as possible", "all users") is not an answer — ask once for a number or a named segment; still vague → park it under Open questions as written.
• Do not implement the feature, create a spec, or edit ADRs. GLOSSARY.md changes happen only through the approved candidate-terms flow below.
• Keep the final prompt spartan, direct, plain English — it is a generated artifact, never compressed shorthand.

Candidate context terms

Only when cheap repo evidence or user-requested exploration reveals domain terms missing from or stale in GLOSSARY.md — never run extra exploration just to fill this. The shared flow — what qualifies, presentation, the away-fallback, applying approvals — lives in references/context-terms.md. Show candidates before finalizing the prompt, apply only approved updates, and report the edited path before saving. Terms still needing review stay under Open questions.

Agent use

Sub-agents: dispatch local lanes automatically for independent work — never cloud agents; announce the lane count at dispatch and report each lane as it completes. Lanes are for fast, independent context checks only — no worker agents, no code edits.

Final output

1. Draft the final prompt.
2. For non-trivial or inferred prompts, show it once for correction, listing inferred facts (project, slice boundary, end result) separately from what the user stated. User away → save as drafted and note at the top that it is unconfirmed.
3. If candidate context terms were found, run the shared approval flow.
4. Verify the pre-save checklist and save to the path below, then re-open the saved file and confirm sections and path match the approved draft, or the explicitly unconfirmed draft when the user is away, before reporting.
5. Add only:

Context updated: <relative GLOSSARY.md path> [only if edited]
Saved to: <relative path written>
Next: pass this final prompt to the `grill-with-field` skill.
Suggested next skills (optional):

- /grill-with-field: challenge assumptions, sharpen domain terms, and confirm decisions.
- /handoff + /prototype: when open questions are ungrillable and need a higher-fidelity spike.
- /to-spec: if this needs a formal spec (PRD) after grilling.

File output

Path

<artifacts-root>/specs/prompts/NNNN-<feature-slug>-prompt.md

Resolve <artifacts-root>: the *.code-workspace directory if one exists, else the per-context root (GLOSSARY-MAP.md at repo root; legacy CONTEXT-MAP.md), else the repo root.

• NNNN — scan <artifacts-root>/specs/adr/ and specs/prompts/ for the highest existing four-digit number and increment. This avoids numbers occupied when the prompt is created. ADRs have an independent sequence and may later reuse a prompt number; the recorded prompt path is the link.
• <feature-slug> — kebab-case from What is needed, max 4 words, ASCII.
• -prompt — fixed suffix marking the artifact type.

Conflict handling

• Create specs/prompts/ lazily. Never overwrite a number already used by another artifact. Never delete unrelated files.
• A prior same-slug prompt counts as hand-edited when git shows commits or working-tree changes to it that this session didn't make; if git can't tell (untracked file), assume hand-edited.
• A same-slug prompt created in this session and unchanged since → update in place. Hand-edited → show the diff and ask: overwrite, new numbered revision, or abort. If the user is away, write a new numbered revision — never overwrite hand edits unconfirmed.

File body

Write exactly the final prompt body — no preface, no generated-by header — drop-in usable as input to grill-with-field.

Pre-save checklist

• Only the six allowed section headers appear, in contract order
• Project: names the full Project Matrix code verbatim (when one exists)
• For a new file, NNNN unique across specs/adr/ and specs/prompts/; an in-place update retains its own number; slug ≤ 4 words, kebab-case ASCII, -prompt suffix
• Split intake: deferred slices recorded under Known limits; the zero-attribution rule is present
• No placeholder text (TBD, TODO, [...]) remains, and no two sections contradict each other
• File body is the prompt only — drop-in, no preface