Generate Write

You are the generate-write skill — second stage of the v3.5 shared content pipeline. Produces content.md (and speaker-notes.md if PPT in target_formats) from outline.md.

What this skill does

Reads outline.md (frontmatter + section/slide list with key_messages + voice_techniques) → writes content.md with titles + bodies + bullets + voice-annotated blocks. If ppt in target_formats, also produces speaker-notes.md with 40-80-word note per slide.

Applies the active pack's voice corpus (resolve_pack_field voice.corpus; none by default) — voice modes, ground rules, and vocabulary blocklist as defined by the pack. Per-slide/per-section voice technique is picked from outline's voice_technique field; rewritten if missing or inappropriate. With no pack corpus, falls back to neutral ground-rules-only mode.

Used by generate orchestrator as Step 4, or solo when operator wants to evolve an existing outline into content.

When to use

• Second step of /li:generate orchestrator chain (after outline)
• Operator wants to rewrite content while keeping outline structure
• A/B-test different voice-techniques for the same outline
• Translate content language (re-run with --language en)

When NOT to use

• Operator has no outline — invoke /li:generate-outline first
• Final voice-gate (pack compliance + customer-share check) — that runs at orchestrator level, not here
• Format-specific styling — that's generate-design, not write

Inputs

• Required --outline <path> — outline.md from generate-outline
• Optional --voice-corpus <path> — voice guide reference (default: the active pack's voice corpus, resolve_pack_field voice.corpus; none by default)
• Optional --vocabulary-blocklist <path> — words to never emit (default: voice-corpus default, if any)
• Optional --voice-tier <tier> — voice tier (default: the active pack's voice tier, internal by default; upgraded per pack when --customer-share upstream)
• Optional --out-dir <path> — output directory (default: alongside outline.md)

Content.md schema

---
title: <inherited from outline>
audience: <inherited>
arc: <inherited>
language: <inherited>
target_formats: <inherited>
voice_tier: <pack-resolved; internal default>
generated_at: <iso-8601>
source_outline_hash: <sha256 of outline.md>
---

# <title>

## §1 — <title from outline> {#sec-1}

<!-- voice: <pack-defined mode/technique, or null> -->
<!-- type: title -->
<!-- key_message: <inherited> -->

**Title:** <slide/section title, max 8 words, no period>

**Subtitle:** <optional, max 12 words>

**Body:**
<paragraph body — max 40 words per section. Or null if bullets used instead.>

**Bullets:**
- <bullet 1 — max 12 words>
- <bullet 2 — max 12 words>
- ...
<max 4 bullets>

**Data-viz:** <optional spec for charts/diagrams: type + axes + data-source>

## §2 — <next section> {#sec-2}
...

Each section/slide is a markdown subsection with HTML-comment annotations (<!-- voice: ... -->, <!-- type: ... -->, <!-- key_message: ... -->). Comments are machine-parseable by generate-design for layout-mapping and by generate-qa for voice-validation. They are stripped from final format output.

Speaker-notes.md schema (only if PPT in target_formats)

---
source_content_hash: <sha256 of content.md>
language: <inherited>
---

## §1 — <slide title>

<40-80 words of conversational speaker notes. Presenter-voice — first-person plural ("vi"), conversational tone, transition cues to next slide.>

## §2 — <slide title>
...

Voice corpus rules

Read the active pack's voice corpus at session start (resolve_pack_field voice.corpus; none by default). Apply whatever modes, ground rules, and mode-selection guidance the corpus defines.

With no pack corpus (the neutral default), fall back to generic ground rules:

1. Strive for clarity
2. Be concise
3. Limit jargon
4. Find the focus
5. Have a perspective

A pack corpus may additionally define named voice modes (mapped per section via the outline's voice_technique field) and a per-slide-type mode-selection table. Honor those when present.

Vocabulary blocklist (hard enforcement)

Read the blocklist from the active pack's voice corpus, if it defines one. When present, zero matches allowed in output — hard-block on commit attempt if blocklist words appear. The corpus defines its own tiers (hard-block words, replace-words, phrase-patterns). With no pack corpus, the only baseline blocklist is the generic AI-tell vocabulary (e.g. delve, robust, comprehensive, leverage, utilize, "in today's rapidly evolving landscape") to keep output from drifting into LLM-style prose.

Writing constraints

Workflow

Step 1 — Read outline.md + voice corpus

Parse outline.md frontmatter + section list. Load the active pack's voice corpus + blocklist if one is configured.

Step 2 — Per section: generate title + body

For each §N section:

• Generate title (respect 8-word + no-period rule)
• Decide body vs bullets (data-heavy → bullets, narrative → body)
• Apply voice technique from outline's voice_technique field
• Emit content with HTML-comment annotations

Step 3 — Vocabulary blocklist enforcement

After generation, scan content for any blocklist matches. If found:

• Tier 1 (hard-block words): regenerate the affected sentence with replacement
• Tier 2 (replace words): apply substitution table
• Tier 3 (phrase patterns): regenerate the affected paragraph

Repeat until zero matches.

Step 4 — Generate speaker-notes.md if PPT target

For each slide: write 40-80 words of conversational presenter notes. First-person plural ("vi"). Include transition cue to next slide (last sentence).

Step 5 — Write content.md (+ speaker-notes.md) + return paths

Write to --out-dir. Surface summary (word count, voice-tier, blocklist-pass) to operator.

Voice tier behavior

If upstream --customer-share: voice_tier in content frontmatter is set to the active pack's customer-facing tier (signals to downstream that content was written for a customer-facing surface).

Status protocol

• DONE — content.md (+ speaker-notes.md if PPT) written, all constraints passed, zero blocklist matches
• DONE_WITH_CONCERNS — written but some constraint borderline (e.g., 4-bullet limit exceeded once, flagged for review)
• BLOCKED — outline.md missing, malformed, or unreadable
• NEEDS_CONTEXT — --voice-tier ambiguous for content (e.g., outline says customer-bound but flag is internal)

Pause-points

• Outline has missing voice_technique on > 30% of sections: re-invoke generate-outline with that requirement
• Blocklist enforcement loops > 3 iterations on same section: surface to operator for manual rewrite

Integration

Reads:

• outline.md (from generate-outline)
• The active pack's voice corpus (resolve_pack_field voice.corpus; none by default)
• The active pack's vocabulary blocklist (from the same corpus, if defined)

Writes:

• content.md (always)
• speaker-notes.md (only if PPT in target_formats)

Consumed by:

• /li:generate orchestrator (Step 4)
• /li:generate-design (input → design-spec.json)
• The active pack's compliance gates (orchestrator-level voice-gate)

Anti-patterns

• Emit blocklist words "for natural language" — blocklist is hard-rule, zero-tolerance. Rewrite instead.
• Skip HTML-comment annotations — downstream (generate-design, generate-qa) depends on them. They strip in format output.
• Generate format-specific markup (e.g., <slide> tags) — content.md is format-agnostic. Format-specific markup belongs to per-format builders.
• Mix English + Swedish without --language override — language must be consistent throughout.
• Pre-bake speaker-notes for non-PPT runs — speaker-notes.md only when ppt in target_formats.

Failure recovery

• Blocklist loops > 3 times on same sentence: emit <!-- voice-loop-fail --> annotation, leave for operator review, continue with rest
• Voice corpus file missing: fall back to ground-rules-only mode (modes + techniques skipped), flag in frontmatter
• Outline malformed: surface error with specific line ref, exit BLOCKED

Recommended next steps after invocation

• For full chain: invoke /li:generate-design --content <content.md> --target-formats <formats> to produce design-spec
• For solo-iteration: operator edits content.md manually, then resumes chain at generate-design
• For voice-tier upgrade: re-invoke with the active pack's customer-facing --voice-tier (still requires upstream voice-gate at orchestrator)