Skip to content
Back to skills

Write Spec

ASecurity

Draft a new spec (SPEC-NNNN) for aireceipts. Use when the user asks to spec a feature, write/create a spec, or turn an idea into a milestone spec. Runs an interview gate before drafting — never drafts on a vague ask.

  • 31 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added September 23, 2026
ai-agentsgo

Security analysis

A100/100

Scanned September 23, 2026

npx -y skills add anandgupta42/aireceipts --skill write-spec --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Write Spec?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Write Spec
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/anandgupta42-write-spec/badge)](https://www.skillsdirectory.com/skills/anandgupta42-write-spec)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: write-spec
description: "Draft a new spec (SPEC-NNNN) for aireceipts. Use when the user asks to spec a feature, write/create a spec, or turn an idea into a milestone spec. Runs an interview gate before drafting — never drafts on a vague ask."
trigger: /write-spec
---

# /write-spec — interview, then draft

## 1. Interview gate (required — do not skip)

Before writing anything, ask the user (or reason explicitly from what they've said) to
pin down:
- **Goal** — what does this deliver, in one sentence?
- **Non-goals** — what is it explicitly not doing?
- **Kill criterion** — what would prove this was a bad idea? If none exists, that's a
  sign the idea is too vague to spec yet — say so and stop.

This is the single highest-leverage step: over-eager drafting on an underspecified ask
is the most common spec-system failure. If the ask is already precise, state your
understanding back in one paragraph and get a confirming nod before drafting.

## 2. Read first

1. `specs/TEMPLATE.md` — required structure and frontmatter.
2. `specs/SPEC-0000-product.md` — the binding invariants I1–I6. Restate the ones this
   spec touches.
3. A recent shipped spec as the voice model — terse, opinionated, concrete, no filler.
4. The real source files the design touches. Ground every claim in `file:line`. Never
   invent code that doesn't exist — if unsure, put it in Open questions (folded into
   Non-goals or a scenario, per TEMPLATE).

## 3. Numbering & frontmatter

- Next id: `ls specs/ | grep SPEC- | sort | tail -1`, then +1.
- Fill YAML frontmatter exactly: `id, title, status: draft, milestone, depends`.
- `depends` lists any SPEC ids this one builds on; `[]` if none.

## 4. Draft against TEMPLATE.md

Purpose / Requirements (`Rn`, each testable) / Scenarios (Given/When/Then) / Non-goals /
Test matrix / Success criteria (checkboxes, always include the unmasked verification
block) / Tombstone (omit — only for rejected specs).

For a milestone (`M`-file), also fill the **Agent team config** section: roles by
directory ownership, dependency waves, critical path, team prompt.

## 5. Discipline

- Status stays `draft`. This skill never sets `approved` — that's the maintainer's button.
  Drafts never self-approve.
- If this is research-derived, link the source.
- Keep it as tight as the reference specs — no restating AGENTS.md, no type definitions
  that belong in code (cite `file:line` / a zod schema instead).

When done: write the file, print the path, and a 2-line summary of what's still open —
then run `/validate-spec` on it. A draft without a `## Validation` record never reaches
the maintainer.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.