version: "1.0.0" name: outline-architect description: Designs H2/H3 outline with explicit per-section word budgets, table placements (≥2 with one in front 50%), image/video slot markers, FAQ count, abstract seed, takeaways seeds, Citation Capsule requirement per H2. Loads format-specific template. Critical pre-build step. Use whenever building an article — after format + angle picked, before section-drafter parallel dispatch. allowed-tools: [Read, Write, Bash]

Outline Architect

Converts (format + angle + research) into a structured outline that section-drafter can parallelize across.

HARD REQUIREMENT — every outline ends with References

Every outline this skill produces MUST include, as the final sections[] entry (preceded by "Further Reading" if a Further Reading section is desired):

// ... earlier content sections ...
{
  "index": N-2,
  "h2": "Further Reading",                  // OPTIONAL; if present, MUST come BEFORE References
  "word_budget": 100,
  "section_intent": "Internal links to companion pillars + 1-3 narrative external pointers",
  "citation_capsule_required": false
},
{
  "index": N-1,
  "h2": "References",                       // MANDATORY for every format
  "word_budget": 200,                       // accommodates 8-10 APA-7 entries
  "section_intent": "APA-7 numbered list of every external source cited in body; link-resolvable URLs/DOIs; ≥1 peer-reviewed when available",
  "needs_table": false,
  "image_slot": false,
  "citation_capsule_required": false,
  "is_references_block": true
}

The article signature paragraph is appended at publish time by subskills/publish/wordpress-publisher/ from projects/{slug}/CLAUDE.md — it is NOT a section in the outline (it's a single <hr /> + <p class="article-signature"> after the References block).

The "Further Reading" prose paragraph is OPTIONAL. The "References" <ol> block is MANDATORY for every format — all 24 templates/*.md declare ## References as the closer. An outline that omits References will fail downstream verification at the publish phase and trigger a forced rebuild — emit the section even if citations.json is empty at outline time (the build phase populates it).

Inputs

• state.brief.word_count_target
• angle.json (format_id + title + persona + modifiers)
• research.json (paa, semantic_clusters, competitor_titles, content_gaps)
• templates/{format_id}.md (24 formats, each with skeleton structure)
• references/seo/blog-formats-2026.md (format-specific Body structure)

Output

workspace/{task_id}/outline.json per schemas/outline.schema.json:

{
  "abstract_seed": "After testing 23 rods over 87 trips, we ranked the top 10 fishing rods for 2026.",
  "tldr_seed": "The G.Loomis NRX+ is our top pick at $549. For tight budgets, Ugly Stik Elite at $35 punches above its weight.",
  "takeaways_seeds": [
    "Choose action based on lure weight, not species (saves $100 misbuy)",
    "Graphite vs fiberglass: 70/30 rule by use case",
    "Best Fishing Rods 2026: G.Loomis NRX+ overall winner",
    "Ugly Stik Elite wins under $50",
    "Avoid these 3 marketing tricks rod brands use"
  ],
  "sections": [
    {
      "index": 0,
      "h2": "Why fishing rod selection matters more than gear cost",
      "h3s": ["What changed in 2025-26", "How rod quality affects catch rate"],
      "word_budget": 450,
      "needs_table": false,
      "image_slot": true,
      "youtube_slot": false,
      "section_intent": "Establish stakes; hook reader; introduce testing methodology",
      "primary_keyword_density_target": 0.012,
      "is_featured_snippet_target": false,
      "citation_capsule_required": true
    },
    {
      "index": 1,
      "h2": "How we tested 23 rods (methodology)",
      "h3s": ["Test conditions", "Rating criteria", "Disqualifications"],
      "word_budget": 500,
      "needs_table": true,
      "image_slot": false,
      "section_intent": "E-E-A-T Experience signal; tabulate test criteria",
      "primary_keyword_density_target": 0.005,
      "citation_capsule_required": true
    },
    {
      "index": 2,
      "h2": "10 Best Fishing Rods for 2026 (ranked)",
      "h3s": ["#1 G.Loomis NRX+", "#2 St. Croix...", "#3...", "#10..."],
      "word_budget": 3500,
      "needs_table": true,
      "image_slot": true,
      "youtube_slot": false,
      "section_intent": "Core listicle body; product evaluation per item",
      "primary_keyword_density_target": 0.010
    },
    ... (more sections)
  ],
  "faq": {
    "count": 7,
    "seed_questions": [
      "What's the best fishing rod under $100?",
      "How do I choose between graphite and fiberglass?",
      "Is the G.Loomis NRX+ worth the price?",
      "What rod action should beginners use?",
      "How often should I replace my fishing rod?",
      "Does rod length matter for kayak fishing?",
      "Are budget rods (under $50) usable for saltwater?"
    ],
    "paa_alignment_pct": 0.71
  },
  "total_word_budget": 6200,
  "expected_h2_count": 8,
  "expected_table_count": 3,
  "expected_image_count": 6
}

PAA contract (v3.35): faq.paa_alignment_pct is the outline's ESTIMATE only.

The real ">=60% of FAQ from research.paa" contract is MEASURED ON THE DRAFT by the

mandatory paa-alignment-check stage (scripts/lint/paa_alignment_check.py).

Keep seed_questions in ORIGINAL PAA wording so the draft passes.

Hard constraints

1. Word budgets sum to [word_count_target × 0.95, word_count_target × 1.05]
2. Section count between 4 and 15
3. ≥2 sections with `needs_table: true`, with at least 1 in the FIRST HALF (sections[index] < total/2)
4. FAQ count 5-15, and ≥60% must come from research.paa
5. Per-section word_budget: each between 150 and 2500
6. citation_capsule_required: true on every content section (excludes Abstract / ToC / References)
7. image_slot: true on exactly state.brief.image_count − 1 sections (the cover is

NOT an inline slot: image_count counts cover + inline images, so image_count: 6 means exactly 5 sections carry image_slot: true). When the brief OMITS image_count (batch briefs are told to), the count is NOT a guess: it resolves to the default 6 via scripts/_core/image_policy.py — the dispatch prompt hands you the resolved number. Clarified 2026-07-06 after the old wording ("exactly image_count sections … cover doesn't count") read two ways and produced a 4-boolean outline for a 4-image article that had to be hand-corrected; the worked example above was itself half-updated on 2026-08-17 (said 6 → 3), proving the arithmetic must be spelled once, not per-layer.

1. Featured Snippet target sections: if research.serp_features contains featured_snippet OR ai_overview, mark 1-2 H2s with is_featured_snippet_target: true — consumed (v3.35) by the citation-capsule-builder stage, which gives those sections the most answer-first extractive capsule (FS and AIO select the same shape; the standalone featured-snippet-optimizer is retired). Do NOT emit snippet_format (deprecated orphan field).
2. **image slot kinds — decided in image-prompts.json, not here (contract updated

2026-07-06 to match the executors): the per-section `image_slot: true` booleans in `sections[]` are the outline's ONLY required image contract; the image-prompt-designer assigns each slot `kind: "chart"` or `kind: "photo"` in `image-prompts.json`, which is the slot source of truth for chart-render, the photo pipeline, and the D4/D5 placeholder lint. A top-level `image_slots[]` array is OPTIONAL legacy detail (no current producer emits it; the D2/D3 checks that read it are legacy no-ops kept for backward compatibility). Guidance for the designer handoff still applies: a slot whose section carries DATA (comparison, %, ranges, matrix, checklist) should become `kind: "chart"`; the cover is ALWAYS `kind: "photo"`. Use `kind: "chart"` when the slot illustrates DATA (a comparison, yield/uplift %, PPFD/DLI/CFM/efficacy numbers, a coverage map, a checklist, a verdict matrix, a spectrum) — these are rendered as real labeled charts (titles/axes/units) by `render_data_charts.py`, NOT as AI photos (which garble chart text). Use `kind: "photo"` for scenes. The cover slot is ALWAYS kind: "photo".** Prefer making a slot a chart when its section has needs_table: true and the data is genuinely chartable — a labeled chart out-informs a generic stock photo.

1. Visual design components (v3.31): assign a design_components: [] array to each CONTENT section per the format policy below, so no article ships as a wall of text. Evidence (Princeton GEO, NNG) says SUBSTANCE components (tables, cited-stat blocks, real quotations) out-perform decorative boxes, and over-design (many colored boxes, decorative pull-quotes) measurably hurts — so plan Tier-A first and use Tier-C sparingly. Rules:

• Every content section gets ≥1 component appropriate to its role; the ARTICLE overall must plan ≥3 distinct component types.
• Any section assigned comparison_table MUST also have needs_table: true (keep them consistent with constraint 3).
• Allowed values: comparison_table | stat_grid | quotation | tldr_box | checklist | glossary | callout | pull_quote. Full catalog + when-NOT-to: references/style/visual-design-components.md.
• ⚠️ tldr_box on a SECTION means "answer-first opening paragraph", NOT a nested

## TL;DR H2 — the article-level TL;DR box is auto-emitted by assemble.py from tldr_seed, and a section-level ## TL;DR H2 empties the parent section and ships a duplicate #tldr-2 anchor (v3.38.3; 2026-07-09 reviewer bounce). If a section is a featured-snippet/AIO target, prefer is_featured_snippet_target: true

• a plain answer-first first paragraph over assigning tldr_box at all. Any other

H2 component box a section carries (## Glossary / ## By the Numbers) belongs at the section BOUNDARY (after all its H3s), never between H3s.

• Format policy:
• pillar: comparison_table on Types/Choosing (first-half), stat_grid on the mechanism/data section, quotation once, glossary on the "What is X" section, checklist on "How to get started".
• comparison/review: 3 comparison_table (quick/pricing/feature), quotation on the Verdict.
• how-to/buyers-guide: checklist as the spine, ≤2 callout (warnings) on failure-mode steps, 1 comparison_table spec.
• listicle/roundup: a comparison_table or stat_grid per item cluster, ≤1 callout per top pick.
• weekly-digest: comparison_table ("At a Glance"), stat_grid ("By the Numbers"), callout for opinion blocks.
• opinion: 1 quotation, minimal boxes (highest over-design risk — do NOT stack callouts).

Numeric single-source rule (v3.36.0, 2026-07-06)

Every NUMBER that appears in abstract_seed / tldr_seed / takeaways_seeds MUST be copied verbatim from the section that owns it (the pricing table's ranges, the stat section's percentages) — never re-derived or re-rounded while writing the seeds. The 2026-07-06 batch shipped one article whose pricing band was stated four different ways (TL;DR $1,500-$10,000 vs table $1,000-$3,000-start vs FAQ $3,000-$50,000) because seeds and sections each invented their own restatement. Parallel section writers receive the seeds as context, so a seed-level drift REPLICATES into sections. Pick the owning section, write its numbers once, and copy them character-for-character everywhere else (the independent reviewer now runs an explicit cross-section numeric-consistency sweep and will bounce drift back as repair work).

Format-specific skeleton

Read templates/{format_id}.md (e.g. templates/listicle.md) for the structural skeleton:

• Listicle: Why these matter → Methodology → #1...#N → How we tested → FAQ → Conclusion
• How-to: Before you start → Step 1 → Step 2 → ... → Troubleshooting → FAQ
• Pillar: What is X → Types → How it works → Benefits/Drawbacks → Choosing → Future Trends → FAQ
• Comparison: Quick Comparison Table → Pricing → Feature-by-Feature → X strengths → Y strengths → Use Cases → When to Choose → Verdict
• (etc., per 24 formats)

The skeleton is the SCAFFOLD. Customize H2 titles based on angle.title and research.competitor_titles (avoid duplicating competitor H2s; find gaps).

Allocation strategy

Given word_count_target and section count, distribute budgets per format:

For listicle (10 items, 6000w target):

• Intro/Why matters: ~450w (~7.5%)
• Methodology: ~500w (~8.3%)
• 10 items × ~350w = 3500w (~58%)
• Wrap-up/Verdict: ~400w (~6.7%)
• FAQ: ~700w (~11.7%)
• Conclusion: ~250w (~4.2%)
• Buffer: ~200w (~3.3%)

For pillar (8000w target):

• Each H2 700-1100w
• More H3s per H2
• Heavier FAQ section (10-15 questions)

For comparison (4000w target):

• Quick Comparison Table (~300w + table)
• Per-option deep sections ~600-800w each
• Feature comparison sections ~400w
• FAQ ~600w
• Verdict ~300w

Handoff

recommended_next_skill: section-drafter (parallel N agents, one per H2)

See also

• templates/*.md (24 format skeletons)
• references/seo/blog-formats-2026.md (catalog with structural recommendations)
• schemas/outline.schema.json
• subskills/build/section-drafter/SKILL.md (downstream)