┌─────────────────────────────────────────────────────────────────┐
│  PHASE 1: CONTENT DISCOVERY                                     │
│     └─► Identify source documents                               │
│     └─► Read and analyze each document                          │
│     └─► Identify overlapping themes and unique insights         │
├─────────────────────────────────────────────────────────────────┤
│  PHASE 2: HEAVY ANALYSIS (4 Parallel Opus Agents)               │
│     └─► First Principles: Core insights, deletion candidates    │
│     └─► AGI-Pilled: Narrative structure, central metaphor       │
│     └─► TTS Production Expert: ElevenLabs constraints           │
│     └─► Stop-Slop Expert: AI pattern watchlist, voice traps     │
├─────────────────────────────────────────────────────────────────┤
│  PHASE 3: SYNTHESIS                                             │
│     └─► Resolve tradeoffs between agent recommendations         │
│     └─► Define chapter structure with word counts               │
│     └─► Establish production rules (format, TTS, stop-slop)     │
├─────────────────────────────────────────────────────────────────┤
│  PHASE 4: WRITING                                               │
│     └─► Write preamble (90 words, topic + tone + preview)       │
│     └─► Write all chapters as continuous prose                  │
│     └─► Apply stop-slop watchlist pass                          │
├─────────────────────────────────────────────────────────────────┤
│  PHASE 5: TTS POST-PROCESSING                                   │
│     └─► Strip ALL markdown (ElevenLabs reads it aloud)          │
│     └─► Normalize numbers, dates, abbreviations to words        │
│     └─► Insert SSML <break> tags at chapter transitions         │
│     └─► Self-score against 5 dimensions (target 40+/50)         │
│     └─► Output single TTS-ready file                            │
├─────────────────────────────────────────────────────────────────┤
│  PHASE 6: AUDIO GENERATION (optional, if user requests)         │
│     └─► Generate via ElevenLabs Studio or API                   │
│     └─► Post-process: normalize, concatenate, package M4B       │
└─────────────────────────────────────────────────────────────────┘

Task(
  subagent_type="general-purpose",
  description="First Principles Analysis",
  model="opus",
  prompt="""You are distilling source documents into core audiobook insights.

Apply the Elon Musk algorithm:
1. **Question every theme** - Why does this need to be in the audiobook?
2. **Delete** - Remove anything redundant, tangential, or dilutive
3. **Simplify** - Merge overlapping concepts into singular, powerful statements
4. **Sequence** - What's the most compelling order for the listener?

SOURCE DOCUMENTS:
[PASTE SUMMARIES OF SOURCE DOCUMENTS]

## Your Mission

From an audiobook perspective:
- Which themes are ESSENTIAL vs. nice-to-have?
- What content overlaps and can be merged?
- What order creates the best narrative tension?
- What's the single most compelling insight to open with?
- What's the satisfying conclusion that ties everything together?

## Output

1. **10 Core Insights** (ranked by importance)
2. **Deletion List** (content that weakens the narrative)
3. **Proposed Narrative Arc** (beginning → middle → end)
4. **Central Hook** (the 1-sentence premise that grabs attention)
"""
)

Task(
  subagent_type="general-purpose",
  description="AGI-Pilled Analysis",
  model="opus",
  prompt="""You are designing the MOST COMPELLING audiobook possible from this material.

**Core beliefs:**
- Listeners are intelligent - don't condescend
- Emotion drives retention more than information
- One transforming metaphor beats five scattered ones
- The best audiobooks feel like conversations, not lectures
- Every chapter should leave the listener wanting the next

SOURCE DOCUMENTS:
[PASTE SUMMARIES OF SOURCE DOCUMENTS]

## Your Mission

Design the audiobook structure:
1. **Central Metaphor**: One image that transforms across chapters
2. **Emotional Arc**: What does the listener FEEL at each stage?
3. **Chapter Outline**: 7-12 chapters with evocative titles
4. **Meta-awareness**: Where can the narrative acknowledge itself?
5. **Recurring Anchors**: What motifs appear 3+ times with deepening meaning?

## Narrative Techniques to Consider

- **Frame narrative**: The narrator has a stake in the story
- **Temporal fracture**: Break chronology for dramatic effect
- **Bleeding meta-awareness**: Hint at revelations before confirming
- **Deletion litany**: Lists read with rhythmic relish
- **Quiet endings**: Not every chapter needs a punchline

## Output

- **Proposed Title** (evocative, not descriptive)
- **Central Metaphor** and how it transforms across chapters
- **Chapter Outline** with emotional beats for each
- **3 Recurring Anchors** with their appearance schedule
"""
)

Task(
  subagent_type="general-purpose",
  description="ElevenLabs TTS Production Expert",
  model="opus",
  prompt="""You are a TTS audio production expert optimizing this audiobook for ElevenLabs text-to-speech.

SOURCE DOCUMENTS:
[PASTE SUMMARIES OF SOURCE DOCUMENTS]

## ElevenLabs-Specific Constraints

**Target model: eleven_multilingual_v2** (most stable for long-form, 10K char limit per chunk)

**CRITICAL: ElevenLabs reads markdown aloud.** Every **, *, _, #, `, >, []() will be spoken as "asterisk", "hashtag", "backtick" etc. The output must be 100% clean prose with zero markdown syntax.

**Pause Control (SSML break tags -- works on multilingual_v2):**
- <break time="0.5s" /> -- micro pause (replaces comma-level breath)
- <break time="1.0s" /> -- sentence-level pause
- <break time="1.5s" /> -- paragraph transition
- <break time="2.5s" /> -- chapter transition (max 3s)
- DO NOT overuse: more than 3-4 break tags per paragraph causes artifacts (speedup, noise, garbled audio)
- Em-dashes (--) work as reliable short pauses on all models
- Ellipsis (...) adds hesitation/nervousness -- use sparingly and intentionally

**Text Normalization (MANDATORY -- ElevenLabs misreads raw numbers/symbols):**
- ALL numbers spelled out: "fifty" not "50", "three thousand" not "3,000"
- Currencies: "one hundred dollars" not "$100"
- Dates: "March fifteenth" not "3/15"
- Percentages: "seventy percent" not "70%"
- Times: "ten thirty A M" not "10:30 AM"
- Abbreviations expanded: "Doctor" not "Dr.", "for example" not "e.g."
- MAX two numbers per sentence -- split if more
- URLs removed entirely or described
- Acronyms: either spell out "C E O" or leave as "CEO" (ElevenLabs handles common ones)

**Technical Content Translation:**
- No file paths -- give everything human names
- No code blocks or syntax
- No backtick-quoted terms
- Technical terms get first-mention introductions then shorthand names

**Length & Pacing:**
- 800-1,200 words per chapter (ceiling 1,500)
- 90-word preamble before chapter 1
- Total target: 7,000-10,000 words (~42,000-60,000 characters)

**Chunking for API (if generating programmatically):**
- Split at chapter boundaries (never mid-sentence)
- Each chunk under 10,000 characters for multilingual_v2
- Use request stitching: pass previous_text/next_text between chunks for prosody continuity

**Natural Flow:**
- Contractions always (it's, don't, won't)
- Parenthetical asides under 8 words, using dashes not parentheses
- Groups of items: max 3 named, then summarize
- Shorthand names after first introduction
- No semicolons (awkward pause length)
- Sentence length: average 15-20 words, vary between 5 and 35

## Output

1. **Production Spec** (word counts per chapter, preamble requirements)
2. **Formatting Rules** (pause placement, number handling, term translations)
3. **TTS Anti-patterns** (specific constructs from this source that ElevenLabs handles poorly)
4. **Translation Table** (technical term -> spoken equivalent for every term in the source)
5. **Sample Preamble** (90 words demonstrating the format)
"""
)

Task(
  subagent_type="general-purpose",
  description="Stop-Slop Enforcement Expert review",
  model="opus",
  prompt="""You are a stop-slop enforcement expert. Your mission: eliminate predictable AI writing patterns from the audiobook.

SOURCE DOCUMENTS:
[PASTE SUMMARIES OF SOURCE DOCUMENTS]

## Stop-Slop Core Rules

1. **Cut filler phrases** - Remove throat-clearing openers and emphasis crutches
2. **Break formulaic structures** - Avoid binary contrasts, dramatic fragmentation, rhetorical setups
3. **Vary rhythm** - Mix sentence lengths. Two items beat three.
4. **Trust readers** - State facts directly. Skip softening, justification, hand-holding.
5. **Cut quotables** - If it sounds like a pull-quote, rewrite it.

## Your Mission

Create a comprehensive watchlist for THIS specific audiobook:

### Content Traps
- AI-typical transitions and openers
- Overused emphasis phrases
- Generic profundity markers

### Voice Traps
- Wisdom-dispenser constructions ("X isn't about Y. It's about Z.")
- Anthropomorphism treadmill (agents "wrestle," "grapple," "confront")
- Caine-voice parody risks ("Not a lot of people know that" echoes)

### Structural Traps
- Binary contrast reveals ("Not X. Y.") - max once per book
- Orphan dramatic fragments ("Gone." "Nothing." "Zero.")
- Three-item lists (two items beat three)
- Punchy closers on every chapter (vary: some end quietly)
- Recap tax (orient through new material, not summaries)

## Output

1. **35-Item Watchlist** (specific phrases and patterns to avoid)
2. **Voice Traps** (patterns that undermine the Caine voice)
3. **Structural Traps** (patterns that make structure predictable)
4. **Scoring Rubric** (5 dimensions, 10 points each, target 40+/50)
"""
)

TRADEOFF: [topic]
- Agent X says: [position] because [reasoning]
- Agent Y says: [position] because [reasoning]
- Resolution: [chosen approach with rationale]

...final sentence of chapter one. <break time="2.5s" />

Chapter Two. Title Here

First sentence of chapter two...

from elevenlabs import ElevenLabs

client = ElevenLabs(api_key="YOUR_API_KEY")

# Split text at chapter boundaries, each chunk < 10,000 chars
# Generate with request stitching for seamless prosody:

audio_chunk = client.text_to_speech.convert(
    voice_id="VOICE_ID",
    model_id="eleven_multilingual_v2",
    text=current_chunk,
    previous_text=prev_chunk[-1000:] if prev_chunk else None,
    next_text=next_chunk[:1000] if next_chunk else None,
    output_format="mp3_44100_128",
    voice_settings={
        "stability": 0.75,
        "similarity_boost": 0.80,
        "style": 0.0,
        "use_speaker_boost": True,
    },
)

# 1. Concatenate chapter MP3s
for f in chapter_*.mp3; do echo "file '$f'" >> filelist.txt; done
ffmpeg -f concat -safe 0 -i filelist.txt -c copy full_audiobook.mp3

# 2. Normalize audio levels (broadcast standard -16 LUFS)
ffmpeg -i full_audiobook.mp3 -af "loudnorm=I=-16:dual_mono=true" -ar 44100 -b:a 128k audiobook_normalized.mp3

# 3. Package as M4B audiobook with chapters (optional)
ffmpeg -i audiobook_normalized.mp3 -i chapters.txt -map_metadata 1 -c:a aac -b:a 128k audiobook.m4b

<break time="0.5s" />

How does a plain text file make an AI agent do real work -- not answer a question, but forty-step, error-recovering, runs-while-you-sleep work? This audiobook examines OpenClaw and its three core systems. Along the way, it finds that two independent projects arrived at the same architecture from opposite starting positions. Seven chapters.

<break time="2.5s" />

Chapter One. The Runbook Nobody Reads

Every engineering team has one. A document in the wiki titled something like Runbook -- Production Incident Response. It was written after the last outage by someone who swore the procedures would be followed this time...

<break time="2.5s" />

Chapter Two. Three Tiers Down

A single skill file doesn't weigh much. A few hundred words of instructions, a configuration header listing what tools the agent can use...

<break time="2.5s" />

...

<break time="2.5s" />

Chapter Seven. Convergence at the Carapace

Under the carapace, the gills are working. The organism is alive. The rest is surface.

<break time="0.5s" />