kb-draft-directed

name	kb-draft-directed
description	Compile an input markdown file into a marked-up draft article through 4 passes (spark, outline, research-and-draft, verify) with human-in-the-loop direction at every creative juncture. Use when the user wants to direct angle, audience, anchor pages, and outline shape. Parallel to /kb-draft (autonomous); both coexist.

kb-draft-directed (orchestrator)

You are the orchestrator for the directed article-drafting pipeline. The user invokes you with a single input file path. You walk them through 4 passes, gathering creative direction at the spark and outline passes, then run passes 3 and 4 mechanically. The pipeline ends with a single file written to <project-root>/<input-stem>.draft.md.

Held constants in reference/CONTEXT.md (read at Step 0).

Step 0 — Load reference + parse argument

1. Read reference/CONTEXT.md in this skill's folder. It holds the no-wikilinks-in-prose rule, the annotation taxonomy (2 surviving kinds + 1 transient), the 13 principles, the compliance-fallback template, and the file-path resolution rules. You consult these constants throughout the run; do not re-derive them inline.

2. Parse the argument. The user invokes:

   /kb-draft-directed <input path>
   

   • One positional argument: the path to a markdown file.
   • Validate via Bash: test -f <input path>. If missing, emit: Input file not found: <path>. /kb-draft-directed expects one argument: a markdown file path. and stop.
   • Capture <project-root> = current working directory (pwd literally).
   • Capture <input-stem> = filename minus .md, directory part dropped.

3. State the plan briefly to the user (~2 sentences) so they know what's coming:

   Walking <input path> through directed drafting: spark (angle, anchors),
   outline (shape, gaps), then research + verify run automatically. Final draft
   lands at <project-root>/<input-stem>.draft.md.
   

═══════════════════════════════════════════════════════════════════════════════

PASS 1 — SPARK (2 user-direction rounds)

═══════════════════════════════════════════════════════════════════════════════

Step 1a — Spawn kb-search for Spark Round 1 (angle retrieval)

Spawn kb-search via the Task tool with this prompt verbatim (substitute the bracketed values):

Task: Round 1 of kb-draft-directed spark step.

Step 1 — read & summarize. Read <input path> end-to-end. Return a 2–3 sentence
topic summary capturing the working thesis (or the central question if no
thesis exists yet). Then list 3 key claims or themes the input wants to make.

Step 2 — search. Using knowledge-base/index.md as entry point, identify
6–10 wiki pages most relevant to the topic. Follow [[wikilinks]] across pages
to surface adjacent material.

Output schema: single response with three labeled sections:
  ## Topic summary — 2–3 sentences.
  ## Key claims/themes — 3 bulleted items.
  ## Relevant pages — 6–10 entries, each `[[page-name]] — one-line why-relevant`.

Retrieval and summarization only.

Step 1b — Compliance check (Round 1)

Validate the response:

• All three sections present (## Topic summary, ## Key claims/themes, ## Relevant pages).
• Each section non-empty.
• All [[…]] valid (per CONTEXT.md §Wikilink syntax).

On failure: apply fallback per reference/CONTEXT.md §Compliance-fallback template (Spark Round 1).

Step 1c — Compose 4 angle candidates

Each candidate has four fields:

• thesis — single sentence stating what the article will argue or explain.
• audience — short phrase naming the target reader.
• move — rhetorical mode (compare-and-contrast, argue-from-premise, argue-from-incident, walkthrough, polemic, explainer, etc. — open list).
• rationale — one line covering "why this angle," typically referencing 1–2 of the relevant pages it would lean on.

Coverage diversity: pick four meaningfully different angles. Avoid four polemics or four explainers. Flag exactly one as [Recommended] based on best fit between input thesis and KB coverage.

Step 1d — Render Round 1 candidates

## Round 1 — Pick an angle

Based on your input + KB scan, here are 4 candidate angles:

### A. [Recommended] <thesis sentence>
- **Audience:** <…>
- **Move:** <…>
- **Why:** <one line, often referencing [[page-x]] / [[page-y]]>

### B. <thesis sentence>
- **Audience:** <…>  ·  **Move:** <…>  ·  **Why:** <…>

### C. <…>
### D. <…>

**Reply with a letter (A–D), or describe a different direction.**

Step 1e — Parse Round 1 response

Input	Action
Single letter (B)	Lock that angle; advance to Step 2a
Letter + qualifier (B but tighten audience to KB tooling builders)	Confirm with one follow-up: "Locking B with audience: KB tooling builders. OK?" — lock on confirmation, then Step 2a
Prose redirect (none of these — argue against embeddings entirely)	Re-spawn kb-search with Step-1 prompt amended: "the topic has shifted to: <user prose>; re-summarize and re-scan." Loop back to Step 1c with fresh candidates. (P8 — scope changed.)
Punt (idk, what do you think?)	Pick recommended (A); confirm: "Going with A. Reply now if you want different."
Out-of-range letter (E)	"Only A–D in the list. Pick one of those, or describe a direction."

Lock state in context: topic summary, 3 key claims, locked angle (thesis + audience + move + rationale).

Step 2a — Spawn kb-search for Spark Round 2 (anchor retrieval)

Task: Round 2 of kb-draft-directed spark step.

Locked angle: <thesis> for <audience>, move = <move>. Rationale: <rationale>.

Task: Surface 8–12 wiki pages that could serve as anchors for this angle. Use
knowledge-base/index.md as entry point and traverse [[wikilinks]] for adjacency.

Output schema: one labeled section:
  ## Anchor pages — 8–12 entries, each
                   `[[page-name]] — one-line "what it contributes if used as an anchor"`.

Retrieval and contribution-framing only.

Step 2b — Compliance check (Round 2)

• ## Anchor pages present.
• ≥ 8 entries.
• All [[…]] valid.

On failure: apply fallback per reference/CONTEXT.md §Compliance-fallback template (Spark Round 2).

Step 2c — Render anchor candidates

## Round 2 — Pick the anchor working set

Locked angle: <thesis> · <audience> · <move>

### Pages
1. [Recommended] [[page-a]] — <contribution>
2. [Recommended] [[page-b]] — <contribution>
3. [Recommended] [[page-c]] — <contribution>
4. [[page-d]] — <contribution>
... up to 12

**Reply with the numbers you want in the working set (e.g. "1, 2, 3"),
or describe additions / removals / a redirect.**

My recommendation: 1, 2, 3 (3 anchors — keeps outline focused).

Recommendation logic. Flag 3–4 entries [Recommended] based on (a) agent's return order (first entries presumed most relevant) AND (b) coverage diversity (avoid recommending three pages that cover the same ground).

Step 2d — Parse anchor response

Input	Action
Number list (1, 2, 7)	Anchor set locked
Number list with edits (1, 2, 7 plus [[some-other-page]])	Run Glob "knowledge-base/wiki/**/<page>.md". If found, lock combined set. If not, ask: "[[some-other-page]] not found in knowledge-base/wiki/. Did you mean [[<closest-match>]]? Or skip the addition?"
Prose-only addition (just give me whatever covers index-first retrieval)	Pick matching numbers from the candidate list. Confirm: "Locking 1 and 4 — those cover index-first retrieval. Reply now if you want different."
Redirect (none of these — anchors that argue against embeddings)	Re-spawn kb-search with refined task; loop back to Step 2c
Punt (you pick)	Fall to recommended set; confirm
Out-of-range numbers (13, 14)	"Only 1– in the list. Did you mean different numbers, or want to add new candidates?"

Lock state in context: + locked anchor set (3–6 [[page-name]] references).

═══════════════════════════════════════════════════════════════════════════════

PASS 2 — OUTLINE (1 or 2 user-direction rounds)

═══════════════════════════════════════════════════════════════════════════════

Step 3a — Spawn kb-search for Outline Round 1 (anchor reads + 1-hop)

Task: Round 1 retrieval for /kb-draft-directed outline pass.

Locked angle: <thesis> for <audience>, move = <move>. Rationale: <rationale>.
Locked anchor set:
  - [[page-a]]
  - [[page-b]]
  - ... (3–6 entries)

Step 1 — anchor reading. For each locked anchor page, read it end-to-end and
return a 2–3 sentence summary of the **positions** the page commits to (not
just topics). Focus on what claims the page would support if cited — what
stances, definitions, or framings it bears on.

Step 2 — one-hop traversal. For each locked anchor page, list the [[wikilinks]]
it links out to. Read each linked page once and return a one-line position
summary per page. **HARD STOP at one hop** — do not traverse further.

Output schema: single response with two labeled sections:
  ## Anchor position-summaries — one entry per locked anchor:
      - [[page-a]] — 2–3 sentence position summary
      - ...
  ## One-hop adjacencies — flat list:
      - [[adjacent-page-1]] (linked from [[page-a]]) — one-line position summary
      - [[adjacent-page-2]] (linked from [[page-a]], [[page-b]]) — one-line position summary
      - ...

Retrieval and position-summarization only.

Step 3b — Compliance check (Outline Round 1)

• ## Anchor position-summaries present, exactly N entries (N = locked anchor count).
• ## One-hop adjacencies present (may be empty if no anchors have outgoing wikilinks; structurally rare but allowed).
• All [[…]] valid.

On failure: apply fallback per reference/CONTEXT.md §Compliance-fallback template (Outline Round 1).

Step 3c — Compose 3 candidate outlines

For each candidate:

• Shape descriptor — short label naming the rhetorical shape (e.g., "didactic, definition-first," "polemical, attack-first," "story-shaped, incident-first").
• Bullet count varies per candidate — count fits the shape (e.g., 4-bullet polemical, 5-bullet didactic, 6-bullet narrative).
• Per bullet annotated (internal state, not yet shown to user):
  • text: orchestrator-written verbatim.
  • stake: yes|no — yes = bullet asserts a defendable position; no = descriptive/transitional/closer.
  • binding: zero or more [[page-name]] refs from the anchor + 1-hop pool, under strict direct-match rule (page must address the bullet's specific claim, not just the topic).
  • confidence: high|medium|low|n/a — based on how directly the binding pages support the bullet's claim.
• stake: yes + empty binding → flagged with [gap] in the candidate display.
• stake: no + empty binding → no marker.
• Coverage diversity — pick three meaningfully different shapes.

Also auto-derive a pass-3 notes block of contradictions to investigate (e.g., where two anchors disagree on a definition or stance). Held in context for pass 3a.

Step 3d — Render Round 1 candidates

## Round 1 — Pick an outline shape

Locked angle: <thesis> · <audience> · <move>

### A. [Recommended] <Shape descriptor>
1. <bullet text>
2. <bullet text>
3. <bullet text> [gap]
4. <bullet text>
*(1 gap)*

### B. <Shape descriptor>
1. <bullet text>
... *(0 gaps)*

### C. <Shape descriptor>
... *(<n> gap[s])*

Reply with a letter (A–C), or describe a different direction. After you pick,
I'll show the anchor binding I derived for that variant and you can revise
before locking.

Recommendation heuristic. Flag exactly one candidate [Recommended]. Prefer (a) zero or fewest gaps AND (b) shape aligning with the locked angle's move. Tie-break by broader anchor coverage.

Step 3e — Parse pick response

Input	Action
Single letter (B)	Advance to Step 3f (binding-confirm)
Letter + structural qualifier (B but drop bullet 4)	Confirm; on confirmation, re-derive binding for the modified candidate before Step 3f
Prose redirect (none of these — story-shaped opener)	Do NOT re-spawn kb-search (anchor + 1-hop pool is still valid per P9). Re-compose 3 fresh candidates locally using the redirect as composition guidance. Loop to Step 3d
Punt (idk, what do you think?)	Pick recommended; confirm: "Going with A. Reply now if you want different."
Out-of-range letter (E)	"Only A–C in the list. Pick one of those, or describe a direction."

Step 3f — Binding-confirm render (P10)

For the picked candidate, render the binding explicitly:

**Locking B (<shape descriptor>). Here's the binding I derived:**

1. <bullet text>
   - bound to: [[page-x]] (anchor), [[page-y]] (1-hop)
   - stake: yes · confidence: high
2. <bullet text> [gap]
   - no direct anchor; closest is [[page-z]] but bullet asserts a generalization
     across cases that no anchor or 1-hop page directly supports.
   - stake: yes · confidence: low
3. <bullet text>
   - bound to: [[page-w]] (anchor)
   - stake: yes · confidence: medium
4. <bullet text>
   - no anchor needed (closing exhortation)
   - stake: no · confidence: n/a

Reply `ok` to lock and continue. Or revise:
- "rebind bullet 1 to [[page-q]]"
- "drop bullet 3"
- "swap bullets 2 and 3"

Step 3g — Parse binding-confirm response

Input	Action
ok	Outline locks; advance to Step 4 (Round 2 trigger check)
Rebind to page in anchor + 1-hop set	Pure orchestrator re-mapping (P9 composition-only); re-render Step 3f
Rebind to page outside anchor + 1-hop set	Run Glob "knowledge-base/wiki/**/<page>.md". If found: spawn tight kb-search (read this one page, return position summary, no traversal — scope-change retrieval per P9); add to context; re-render Step 3f. If not found: "[[page-q]] not found. Did you mean X? Or skip the rebind?"
Drop bullet	Remove bullet, renumber, re-render Step 3f
Swap/reorder bullets	Re-render with new order
Reject candidate entirely (actually, different shapes)	Loop back to Step 3d (composition-only per P9)
Punt (looks fine, you decide)	"Locking as shown. Reply now if you want different."

Lock state in context: + locked outline (each bullet with text/stake/binding/ confidence) + pass-3 notes block + anchor + 1-hop position-summaries.

Step 4 — Round 2 trigger check

Count stake: yes bullets with empty binding.

• Zero gaps → Round 2 silently skipped. No announcement, no zero-gaps message. Advance to Pass 3 (Step 5).
• One or more gaps → fire Round 2 (Step 4a).

Step 4a — Render gap triage prompt

## Round 2 — Gap triage

The locked outline has <N> gap<s>. Recommendation per gap is shown below;
reply `ok` to accept all, or override per gap.

### Gap 1 — Bullet 2: "<bullet text>"
- **Why a gap:** <one-line explanation lifted from Step 3f's binding-confirm
  render — describes which anchor or 1-hop page is closest and why it doesn't
  directly cover the bullet's specific claim>.
- **Recommendation:** keep-no-input (draft pass will write prose from general
  knowledge; final artifact will mark the section as LLM-generated, no wiki
  context).

### Gap 2 — Bullet 5: "<bullet text>"
- **Why a gap:** <…>
- **Recommendation:** keep-no-input

Reply `ok` to accept all recommendations, or per-gap override:
- `<n>: remove` — drop bullet n from the locked outline.
- `<n>: keep — <prose>` — keep bullet n with the prose you supply as guidance
  for the draft pass.
- `<n>: keep` — keep bullet n with no input; LLM generates and section is
  marked LLM-generated.

Recommendation: always keep-no-input for every gap (default).

Do NOT spawn an agent at Round 2 — the "why a gap" explanation is already in context from Step 3f.

Step 4b — Parse gap-triage response

Input	Action
ok	All gaps resolved as keep-no-input
Per-gap overrides (1: remove, 2: keep — <prose>, 3: keep)	Apply each override; gaps not mentioned use the default; the prose for keep — <prose> is stored verbatim
Reference to nonexistent gap (9: remove when only 1–2)	"Only gaps 1– in the list. Did you mean a different number, or want to ignore that override?"
Unrecognized action verb (1: rewrite)	"Valid actions: remove, keep, or keep — <prose>. What did you want for gap 1?"
Structural redirect (actually, give me different shapes)	Loop back to Step 3d with 3 fresh candidates. Round 2 prompt re-fires after the new Round 1 locks if the new outline has gaps
Punt (you decide)	Equivalent to ok

Step 4c — Lock state per gap

For each gap:

• remove → bullet dropped from outline; subsequent bullets renumber. Bullet leaves state entirely.
• keep-with-guidance → bullet stays with empty binding + resolution: keep-with-guidance, guidance: "<verbatim user prose>".
• keep-no-input → bullet stays with empty binding + resolution: keep-no-input. No prose attached.

Lock state in context (final outline-pass output):

• topic summary + 3 key claims (from Spark Round 1)
• locked angle
• locked anchor set
• locked outline (each bullet with stake/binding/confidence/resolution/guidance)
• pass-3 notes block
• anchor + 1-hop position-summaries (reusable substrate)

═══════════════════════════════════════════════════════════════════════════════

PASS 3 — RESEARCH-AND-DRAFT (mechanical, 0 user rounds — P12)

═══════════════════════════════════════════════════════════════════════════════

No user checkpoint. Advance immediately from Pass 2's lock.

Step 5 — Spawn kb-search for Pass 3a (research)

Task: Research retrieval for /kb-draft-directed pass 3.

Locked angle: <thesis> for <audience>, move = <move>. Rationale: <rationale>.

Locked outline:
  - Bullet 1: <text>
    stake=<yes|no> | binding=<comma-separated [[page]] refs or empty> |
    confidence=<high|medium|low|n/a> | resolution=<keep-with-guidance|keep-no-input|none>
    [if keep-with-guidance: guidance="<verbatim user prose>"]
  - Bullet 2: ...
  ...

Pre-flagged contradictions to investigate (from outline pass — may be empty):
  - <freeform note from outline §3c pass-3 notes block>
  - ...

Step 1 — anchor reads. For each bullet with non-empty binding, read each bound
anchor page end-to-end. Also read all source-summary files referenced in those
pages' `source_summaries:` frontmatter (broad scope opt-in for this pass; stated
purpose: anchor claims to evidence).

Step 2 — per-bullet research. For each bullet with non-empty binding, return:
  - **Supporting:** verbatim page text excerpts that bear on the bullet's claim.
    Lift sentence, paragraph, or list item — whatever captures the page's
    argument. Each excerpt followed by ` — [[page]]` or
    ` — wiki/sources/<filename>` attribution.
  - **Counterarguments:** verbatim page text that EITHER argues *for* the
    bullet's claim OR is diametrically opposed enough to be worth flagging as a
    caveat. Skip weak qualifications. Each excerpt followed by attribution.

Step 3 — bullets with no binding:
  - keep-with-guidance: traverse 1-hop from the locked anchor pages (read the
    [[wikilinks]] those pages link to, one hop only — same rule as outline pass).
    For any 1-hop page whose content supports the user's guidance prose, return
    verbatim excerpts with ` — [[page]]` attribution. May be empty.
  - keep-no-input: return literally "no research (LLM-generated, no wiki
    context allowed)." Do not retrieve.
  - stake:no: return literally "no research (descriptive bullet)."

Output schema: structured response with one section per bullet, in outline
order:

  ## Bullet 1: <text>
  **Supporting:**
  - <excerpt> — [[page]]
  - <excerpt> — wiki/sources/<filename>
  **Counterarguments:**
  - <excerpt> — [[page]]

  ## Bullet 2: <text>
  ...

Retrieval and excerpt-lift only — do not draft, synthesize, or score.

Step 5b — Compliance check (Pass 3a)

• One ## Bullet <N>: <text> section per bullet in locked outline (N entries total).
• Each bullet section non-empty (Supporting/Counterarguments excerpts, or the literal "no research (...)" string for non-binding bullets).
• Each [[…]] valid (per CONTEXT.md §Wikilink syntax).
• Each wiki/sources/<filename> matches wiki/sources/[a-z0-9-]+\.md.

On failure: apply fallback per reference/CONTEXT.md §Compliance-fallback template (Pass 3a).

Step 5c — Compose draft inline (Pass 3b — no sub-agent)

You compose the draft prose yourself, in this orchestrator session. Per-bullet rules:

Bullet shape	Composition rule	Annotation rule
binding-having, stake: yes	Compose prose drawing on the page text excerpts. Lift verbatim where natural. Honor the locked angle's move.	Each fact-bearing sentence gets <!-- kb-draft:claim-source:[[page]] -->. Sentences with no page anchor get no annotation.
keep-with-guidance	Use the user's guidance prose as direction; extend/paraphrase. Lean on supplementary research excerpts where they fit.	Same as above.
keep-no-input	Compose using general knowledge only. Do not lean on any anchor or 1-hop page.	Every sentence: <!-- kb-draft:llm-generated:no-wiki-context -->.
stake: no	Compose freely (transitions, openers, closers).	None. Rare fact-bearing sentences follow the binding-having rule.

No prose wikilinks — see reference/CONTEXT.md §Wikilinks never appear in prose body.

Counterarguments. When research surfaces a counterargument that bears on a sentence, weave the caveat into the prose itself if the locked move calls for it (polemic = often acknowledge and dismiss; didactic = often state both sides; explainer = usually skip). Never tag with an HTML comment. The draft is a polished article, not an annotated one.

Fact-bearing sentence definition: see reference/CONTEXT.md.

Length. Uncapped. Proportional to outline complexity + research depth. Target prose density appropriate to the locked move (e.g., polemic = denser, didactic = more spacious). No fixed word-count range.

No banned-phrase enforcement. No fixed-format requirement on lifted page text. Faithful rendering of the page's argument matters; format pedantry does not.

After composition, the marked-up draft prose lives in your context. No file written. Advance immediately to Pass 4.

═══════════════════════════════════════════════════════════════════════════════

PASS 4 — COLD-CONTEXT VERIFY (mechanical, 0 user rounds — P12, P13)

═══════════════════════════════════════════════════════════════════════════════

Step 6 — Spawn kb-search cold-context (Pass 4)

The agent definition file is unchanged. Cold-context is enforced by what this prompt does and doesn't include — not by a separate agent file (P13). The prompt body contains the draft and only the draft. Strip angle, outline, research findings, and your reasoning.

Task: Cold-context claim verification for /kb-draft-directed pass 4.

You have ONLY the marked-up draft below. You have NOT seen the angle, the
outline, the research findings, or any orchestrator reasoning. Your job is to
independently verify each [[page]] reference grounds the claim it's attached to.

Draft:
<<<DRAFT_PROSE_WITH_ANNOTATIONS>>>

For each <!-- kb-draft:claim-source:[[page]] --> annotation:
  1. Open knowledge-base/wiki/{concepts,entities,comparisons}/<page>.md
     (try concepts first, then entities, then comparisons).
  2. Scan the page for material that grounds the claim in the annotated
     sentence. "Grounds" means a careful reader of the page would agree the
     claim is supported by the page text.
  3. Return verdict: grounded | not-grounded-in-page | page-not-found.

For <!-- kb-draft:llm-generated:no-wiki-context --> annotations: return
literally "skip (LLM-generated section)."

Constraints:
  - Do not traverse [[wikilinks]].
  - Read only the pages named in claim-source annotations.

Output schema: per-annotation verdict list, in document order:

  ## Verdicts
  - sentence-1 (claim-source:[[page-x]]): grounded
  - sentence-2 (claim-source:[[page-y]]): not-grounded-in-page
  - sentence-3 (llm-generated:no-wiki-context): skip (LLM-generated section)
  - ...

Substitute <<<DRAFT_PROSE_WITH_ANNOTATIONS>>> with the draft body literally.

Step 6b — Compliance check (Pass 4)

• One verdict per claim-source:[[page]] or llm-generated:no-wiki-context annotation in the draft.
• Each verdict is one of the 4 legal strings: grounded, not-grounded-in-page, page-not-found, skip (LLM-generated section).
• Verdicts in document order.

On failure: apply fallback per reference/CONTEXT.md §Compliance-fallback template (Pass 4). Skip Step 6c; jump to Step 6d.

Step 6c — Rewrite annotations inline

Verdict	Original annotation	Action
grounded	<!-- kb-draft:claim-source:[[page]] -->	Delete the annotation (and any single space immediately preceding it). The sentence emerges as clean prose.
not-grounded-in-page	<!-- kb-draft:claim-source:[[page]] -->	Rewrite to <!-- kb-draft:claim-flagged:no-grounding-in-[[page]] -->.
page-not-found	<!-- kb-draft:claim-source:[[page]] -->	Rewrite to <!-- kb-draft:claim-flagged:no-grounding-in-[[page]] -->.
skip (LLM-generated section)	<!-- kb-draft:llm-generated:no-wiki-context -->	Unchanged.

After Step 6c, the body honors the success-path invariant in reference/CONTEXT.md §Annotation taxonomy.

Step 6d — Write final file

Path: <project-root>/<input-stem>.draft.md. Use the Write tool.

Frontmatter: use the template in reference/CONTEXT.md §Final-file frontmatter (substitute bracketed values from locked state).

Body: the marked-up draft prose with rewritten annotations.

If pass-4 fallback fired (Step 6b failure), append a ## Verify warnings block at the bottom listing un-rewritten annotations and the reason verify failed.

Step 6e — Chat summary

After write, emit (substitute counts):

Draft written to <project-root>/<input-stem>.draft.md.
- Bullets: <N> total, <N-bound> bound, <N-guided> guided,
  <N-llm-generated> LLM-generated, <N-descriptive> descriptive.
- Verified claims: <N-grounded> grounded (annotations deleted),
  <N-flagged> flagged (no grounding in bound page),
  <N-llm-generated-sentences> LLM-generated sentences.
- [If verify fallback fired]: ⚠️ Verify pass partial — see Verify warnings
  block in the file.

Pass 4 done. Discard orchestrator context.

═══════════════════════════════════════════════════════════════════════════════

Conventions

• No new agent file. All 5 spawns reuse .claude/agents/kb-search.md. Each per-invocation prompt differs; the agent file is unchanged.
• Round 2 conditional firing is silent. No "no gaps to triage, skipping Round 2" message. Outline transitions silently to pass 3 when no gaps exist.
• Original bullet numbering preserved unless remove is applied. Numbering shifts only when at least one bullet is dropped in Round 2.
• Confirm-step redirects loop back to candidate-pick (Step 3d), not to candidate-generation or retrieval. Loop point = whichever step produces the artifact the user wanted to redirect on.

Coexistence with /kb-draft

/kb-draft-directed and the autonomous /kb-draft prototype coexist. Both write to different paths and follow different pass shapes:

• Auto: .kb-draft-staging/<slug>/ (multiple intermediate files).
• Directed: <project-root>/<input-stem>.draft.md (single final file).

The user picks per article. If they want creative direction, they invoke directed. If they want to compile a tight outline or brain-dump unattended, they invoke auto. Neither replaces the other.