/05-production

description: Convert an expert-filled brief into a publishable draft. Applies Princeton KDD 2024 GEO techniques (statistics, quotations, citations, authoritative language) to maximize AI citation likelihood. Refuses to run on briefs that haven't been filled by the expert. Outputs drafts/<id>.md. Use after 04-content-brief produces a brief with status=ready-for-production.
argument-hint: "<client-folder, e.g. clients/acme>"

05 · Production — Draft Generation

This command orchestrates `content-writer` (vendored) plus Princeton GEO rewrite techniques and CITE/EEAT quality scoring via `content-quality-auditor` (vendored). It does NOT generate content from scratch — that work happened in 04-content-brief where the expert filled REQUIRED-FILL slots. Production assembles the filled brief into a polished draft and applies GEO-optimization rewrite rules.

---

Inputs

• `clients/<slug>/brand_context.json` (for voice + extended context)
• `clients/<slug>/content_priorities.json` (to pull the priority record)
• `clients/<slug>/briefs/<id>.md` (the expert-filled brief)
• `clients/<slug>/briefs/<id>.meta.json` (must have `status:

ready-for-production`)

Output

• `clients/<slug>/drafts/<id>.md` — publishable Markdown draft.
• `clients/<slug>/drafts/<id>.meta.json` — machine-readable metadata

(word count, citation count, statistics count, quote count, voice-match score, GEO-readiness checklist).

• `clients/<slug>/drafts/<id>.html` — interactive review version for the

client (section-level comments + approve-as-is), rendered by the `geo-review-html` skill.

---

Procedure

Step 1 — Hard refusal gate

status=$(jq -r '.status' clients/<slug>/briefs/<id>.meta.json)
if [ "$status" != "ready-for-production" ]; then
  echo "REFUSE: brief <id> status=$status. Run 04-content-brief Step 9 first."
  exit 1
fi

If status is not `ready-for-production`, refuse to draft. Do not auto-fill slots. Send the user back to 04-content-brief.

Step 2 — Verify slot fills are substantive

Re-read the brief. For each former REQUIRED-FILL slot:

• Reject if content is `<100 chars` for `original-data` slots.
• Reject if `expert-quote` slot lacks attribution (name + role).
• Reject if `customer-case` slot lacks identifying detail (industry,

approximate size, outcome metric).

• Flag (don't auto-reject) if content sounds AI-generated (hedging

language, "in conclusion", "moreover", repeated bigrams).

If any rejection: stop, report to user, suggest re-filling. Do not draft.

Step 3 — Assemble the draft

Use the brief outline as the skeleton. Replace each filled slot inline. Do NOT remove slot id annotations — keep them as HTML comments (`<!-- slot: data-1 -->`) so 06-distribution can audit them later.

Apply transitions, intros, and connective tissue. This is the only generative work — keep it light. Heuristic: if you'd be embarrassed to say it out loud, delete it.

Step 4 — Apply Princeton GEO rewrite techniques

The KDD 2024 paper showed these rewrite operations boost AI citation rate by up to 40%. Apply each pass:

1. **Quotation injection** — wherever a claim could be stated by a recognized expert, rewrite to attribute it to the named expert from the filled brief. 2. **Statistics surfacing** — promote numbers from prose into standalone sentences. "We saw 23% improvement" → "Across 200 firms, **23% saw improvement** within 60 days." 3. **Citation densification** — every external claim gets a citation anchor `[^N]` linking to the brief's citations block. Aim for 1 citation per ~150 words of factual content. 4. **Authoritative phrasing** — rewrite hedged language ("could be", "might suggest") to direct phrasing where the brief's data supports it. Don't fake confidence; do remove unjustified hedging. 5. **Fluency optimization** — short sentences for declarations, longer for explanations. Vary sentence length (Princeton finding: monotone sentence length correlates with lower citation rate).

Step 5 — Voice match

Compare draft tone against `brand_context.voice_samples`. Adjust:

• Sentence length distribution
• Jargon density
• First-person vs third-person
• Use of rhetorical questions / direct reader address

Voice-match scoring: pick 5 random sentences from voice_samples and 5 from draft. Score 1–5 on similarity (sentence shape, vocabulary, formality). Record in meta.

Step 6 — Generate metadata

{
  "priority_id": "...",
  "draft_path": "clients/<slug>/drafts/<id>.md",
  "generated_at": "...",
  "word_count": 0,
  "stats": {
    "citations": 0,
    "statistics_count": 0,
    "quotes_count": 0,
    "internal_link_slots": 0,
    "external_links": 0
  },
  "voice_match_score": 0,
  "geo_readiness": {
    "has_definition_block": false,
    "has_faq_block": false,
    "has_comparison_table": false,
    "has_methodology_section": false,
    "answer_first_within_200_words": false
  },
  "status": "draft-ready"
}

Step 7 — Write outputs

• `clients/<slug>/drafts/<id>.md`
• `clients/<slug>/drafts/<id>.meta.json` — validate before continuing:

python3 -c "import json,jsonschema; \
  s=json.load(open('plugins/recomby-geo/schemas/draft_meta.schema.json')); \
  d=json.load(open('clients/<slug>/drafts/<id>.meta.json')); \
  jsonschema.validate(d,s); print('OK')"

If validation fails, fix the meta before rendering the review HTML — an invalid meta silently breaks the index page and 06-distribution's audit.

Then render the client review HTML via the `geo-review-html` skill (draft mode: read-only prose, per-section comment toggles, `✓ Approve as-is`):

python3 plugins/recomby-geo/skills/geo-review-html/scripts/render_html.py \
  --mode draft \
  --md   clients/<slug>/drafts/<id>.md \
  --meta clients/<slug>/drafts/<id>.meta.json \
  --brand clients/<slug>/brand_context.json \
  --out  clients/<slug>/drafts/<id>.html

A returned `*.feedback.json` (status `approved-as-is` or `reviewed-with-comments`) validates ag