feature-prompt
Turn a feature idea, change request, or rough requirement into a small prompt for grill-with-field. An ADR to a spec is /to-spec; how existing behaviour works is /feature-discovery.
How do I install this agent skill?
npx skills add https://github.com/devarfeen/agent-skills-kit --skill feature-promptIs this agent skill safe to install?
- Gen Agent Trust Hubpass
The skill is safe but includes a surface for indirect prompt injection as it processes untrusted user input and repository documentation to generate structured prompt files and coordinate local sub-agents. It also utilizes Git for local file state verification.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
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 neededis 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 limitsrecords 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 questionsonly 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, orAcceptancesections. Fold useful facts into the six contract sections; user-stated hard decisions go underKnown limits. - All unclear decisions land under
Open questions; ungrillable ones ("needs to feel/see it") also route to/handoff+/prototypebefore deep grilling continues. - Keep
Open questionsto 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 questionsas written. - Do not implement the feature, create a spec, or edit ADRs.
GLOSSARY.mdchanges 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
- Draft the final prompt.
- 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.
- If candidate context terms were found, run the shared approval flow.
- 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.
- 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/andspecs/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 fromWhat 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,
NNNNunique acrossspecs/adr/andspecs/prompts/; an in-place update retains its own number; slug ≤ 4 words, kebab-case ASCII,-promptsuffix - 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
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/devarfeen/agent-skills-kit/feature-prompt">View feature-prompt on skillZs</a>