Artifact Frontmatter

Emits the skimmable two-layer skim block at the TOP of every phase artifact — TL;DR, phase + complexity tier, what-changed-since-last-version, decisions locked (and which flipped from a Recommended de

SKILL.md

GitHub

Artifact Frontmatter — the 60-second skim layer

Purpose

Every phase artifact opens with a skim layer before any prose (conventions.md §6, two-layer outputs). The reader gets the gist in under a minute; the detailed body follows for anyone who needs to dive in. This skill produces that block. The body is the deep layer; this is the skim layer. Never make the reader parse the whole body to find the decision.

The 60-second test: a reader who reads ONLY this block can answer four questions —

  1. What is this artifact about? (TL;DR)
  2. What phase / tier is it? (header strip)
  3. What changed since the last version? (what-changed)
  4. What's still open? (open [VERIFY WITH PM] items)

If a reader cannot answer all four from the block alone, the block has failed — compress and re-emit.

When to invoke

The phase agent invokes this skill as the first output step, before writing any body section. Order of operations: compose the body internally first so the TL;DR, decisions-locked, and open [VERIFY WITH PM] items are derived from real content — then emit this block first (above the body) in the final artifact. The block is mechanically derived from the artifact/state — it is not hand-prose. Re-emit it whenever the artifact is regenerated (e.g., a v2 PRD after [VERIFY] resolution) so what-changed stays accurate.

The block (fixed order)

Emit in this order. Keep the whole block to one screen (~50 lines markdown); if it overflows, run the compression pass (below) — the budget is the skim layer only, never the body.

1. YAML header strip + title + [AI DRAFT] label

---
spec: <meta.feature_name slug>
phase: <N> — <Phase name>
tier: <Tier label> · <meta.mission_tier>          # e.g. Tier 2 — Standard Spec · IL5
status: <phase.status> · <gate note if any>
artifact: <artifact name + version>
generated: <ISO 8601 timestamp>
---

# <Feature Name> · <Artifact + version>

> **[AI DRAFT]** — Auto-generated skim layer. <one clause on body provenance, e.g. "Body is the human-curated artifact." or "All N [VERIFY] items from vN-1 resolved.">

The [AI DRAFT] label is mandatory on every block (CLAUDE.md output rule — label all AI-generated content until human-reviewed).

2. TL;DR (60-second scan)

## TL;DR (60-second scan) — 2–3 sentences. What this artifact decides/scopes, the load-bearing constraint, and (if relevant) what's deliberately out. Plain language; lead with the BLUF. Bold the one phrase the reader must not miss. This is the single highest-leverage line in the artifact — write it last, after the body, so it reflects what the artifact actually says.

3. What changed since last version

## What changed since <prior version> — a small table or tight bullets. Source: diff against the prior artifact version + resolved context.open_questions. For a first version with no prior, replace with ## Decisions locked this phase (below) and state "First version — no prior to diff." Resolved [VERIFY] items show their resolution; dropped scope shows "Dropped."; flipped decisions show "Flipped.".

4. Decisions locked (flag flips from Recommended)

## Decisions locked this phase — table of the phase's context.key_decisions[]. For any decision the user chose against the agent's Recommended intake default, tag it Flipped (or Scope expansion if it widened scope). This is load-bearing: a reviewer needs to see at a glance where the human overrode the default. Example rows:

# Decision Notes
Q6 Top of list = highest precedence, fixed (no toggle) Flipped from agent recommendation
Q9 Visual builder AND raw CEL are equal first-class surfaces Flipped, doubles editor scope
Q8 Strict additive inheritance — no child override matches existing policy stacking

If the full decision set is long, show the flipped/expansion rows here and point to the body: "Full table in body § Intake Clarifications."

5. Open [VERIFY WITH PM] items — pinned, never buried

## Open going into Phase <N+1> (or ## [VERIFY WITH PM] — pinned, do not bury when there are several). Scan the body for every [VERIFY WITH PM] / [VERIFY WITH ENG LEAD] flag and list each as one line: the question + who must decide + what it blocks. If none: None blocking. plus any P2 polish item. Never let a [VERIFY] flag live only in the body — it must surface here (CLAUDE.md: flag [VERIFY WITH PM] items at the top, not buried).

6. Reading guide (one line, or a small time-budget table)

## Reading guide — at minimum one line: "If you only have 5 minutes, read this block; for the decision rationale, body § X." A time-budget table is the richer form:

If you have… Read
5 min This block. Stop here for gate sign-off conversation.
20 min Add body § Executive Summary, § Scope, § Out of Scope
Full pass Body in full (~N words). The grounding in § X is load-bearing.

7. Source artifacts (linked, not restated)

## Source artifacts — one bullet per upstream artifact this phase consumed, each a link by phase + anchor. This is the skim-layer companion to the dedup skill: the block links sources; the body cites them inline. Never restate an upstream artifact's content here — link it.

Compression pass (when the block overflows one screen)

If the block exceeds ~50 lines: (a) collapse the decisions table to flipped/expansion rows only, pointing to the body for the rest; (b) drop the time-budget table to a single reading-guide line; (c) tighten TL;DR to 2 sentences. Never drop the [AI DRAFT] label, the open-[VERIFY] section, or any flipped-decision row — those are load-bearing.

What this skill does NOT do

  • Does not write the body. It emits only the skim layer; the agent writes the deep layer below the closing ---.
  • Does not invent content. Every line traces to the spec state object or the body. If a field is empty (no decisions, no opens), say so explicitly ("None blocking.") — never fabricate.
  • Does not duplicate upstream content. Source framing is linked (§7), not restated — that is the dedup skill's job, applied in the body.

Self-resolving references (conventions.md §5)

Inside the block, never emit a bare internal code. Every FR-, EC-, threat, or control ID carries an inline gloss at the point of use — FR-10 (admins bulk-assign attributes by policy), not FR-10. A reader skimming this block must never have to open another file to understand a cited code.