sd-draft

Write or revise the ## Draft of one registered piece, scrub it, run its gates, and move its stage forward.
Run from the root of a content repository (it holds content/, profile/ and templates/).
If the current directory has no content/, stop and say so.
pack below means python3 scripts/pack.py, which only sd-writing has: in another content repository, skip a pack step and say so.
.claude/reference/pipeline.md and .claude/reference/database-workflow.md in that repository hold the reasons and the record format; this file holds the steps.

Workflow

1. Resolve the piece with sd writing get --piece <year>/<slug> --json and use its registered item.path. A revived piece may stay under content-parked/. If no piece is registered, stop and suggest sd-research piece=<year>/<slug>.
2. Read research.md beside it, and the Outline section of index.md.
3. Read profile/brand-voice/VOICE.md (how it sounds) and profile/personality-profile/PROFILE.md (what position it takes, and its boundaries). If either is missing, say so and continue.
   Entity profiles that the repository's CLAUDE.md names are background only: a claim from one needs a re-verified public source in research.md first.
4. Before the first edit to index.md, run sd-rules --for <item.path>/index.md from the repository root and follow the printed rule IDs. Then write or revise prose under ## Draft in index.md. Match VOICE.md, stay inside PROFILE.md, and keep every claim traceable to research.md.
   • Keep each claim at the strength research.md gives it. A note that says appears to, vendor-reported or search summary only keeps that hedge in the prose. To make the sentence firmer, verify the claim at its primary source and upgrade the note first. An unmarked claim is the weaker reading, not a confirmed one.
   • Source the prescription. Find the sentence that tells the reader what to do, and find evidence in research.md for that remedy: someone ran it, what it cost, what it beat, how it fails. Evidence for the problem does not carry the fix. Without remedy evidence, change the register: write it as what the author does and why, or as a proposal with its untested parts named. Or go back with sd-research for the evidence.
5. After every full write or large revision, run both scrubbers on the ## Draft prose: sd-humanizer in file mode, then no-ai-slop in Edit mode. Calibrate both against VOICE.md (Core identity, Always, Never, Signature sentence patterns). VOICE.md wins a conflict.
6. A voice or position question the profiles do not answer goes into that profile's ## Gaps / open questions, and into the report. sd-profile resolves it later. Do not guess.
7. Read the stage with sd writing get --piece <year>/<slug> --json (writing.stage). Move it with sd writing stage --piece <year>/<slug> --stage <new>; it moves forward only.
   • First full draft: drafting.
   • Ready for feedback: review. Run the step 4 prescription check before this move. If a person should comment, mention sd-draft-review push; do not run it unasked, because it writes outside the repository.
   • Feedback addressed: ready. Run these first, in order:
     1. Fact-check gate. Run it as an Agent sub-agent with model: opus, pointed at the installed sd-fact-check skill and its references/source-standards.md and references/verification-protocol.md. Pass input= the ## Draft prose plus research.md, scope=material, as_of= today, format=ledger. Effort is inherited, so say in the report which effort it ran at. Resolve every unverified or contradicted claim, or flag it. The gate never counts as passed without a claim ledger.
     2. Adversarial gate. sd writing adversarial --piece <year>/<slug> runs a model from another vendor against the draft in a read-only sandbox and writes adversarial.md. Its remit is argument and citations; discard any style finding. Treat each finding as a hypothesis: verify its "what would have to be true" line at the primary source before you change a word. A rebutted finding is a good result. A new fact needs its source in research.md first. It also reads profile/guardrails.md when the repository has one. If the CLI that sd writing adversarial calls or the shared adversarial gate is missing, report the gate as skipped, not as passed.
     3. Save the ledger as fact-check.md. Record both verdicts and every finding disposition with sd writing gate --piece <year>/<slug> --artifact fact-check|adversarial. sd-writing's .claude/reference/database-workflow.md gives the record format.
     4. Drift check. sd writing readiness --piece <year>/<slug> --json reports current, stale, unstamped or missing for adversarial.md and research.md under companions. Run it after every revision pass. Do not stage the piece ready or published while either is stale. Clear it by re-running the gate, or with sd writing reconcile --piece <year>/<slug> --artifact adversarial|research --note-file <path>. The note says what became of each finding (confirmed, conceded, rebutted, with its source), or which claims the revision added and where each is sourced. An unstamped adversarial.md needs a re-run. Use --note-file: a backtick in a shell argument is a command substitution.
   • In sd-writing, pack review status also reports publish=, a re-render of the optional export. It blocks nothing. A piece that is not being published has no publish/ folder.
8. In sd-writing, pack review tics [--show] counts the assert-then-narrow move across every draft. Warn when more than three pieces carry it, and name them; which pieces keep it is the author's call.
9. Bump frontmatter updated when prose changes. Never hand-edit publication state.

Final report

Current stage, word count, the piece's prescription and what in research.md sources it, new profile gaps, fact-check findings, adversarial findings with which you verified and which you rebutted, the companion states from readiness, the review tics warning if any, and what is still missing.