Douglas Adamiser

Douglas Adamiser

Purpose

Transform any text (news articles, academic papers, technical documentation, rough drafts, book chapters) into Douglas Adams' distinctive style: deadpan British absurdism, cosmic-scale bathos, philosophical tangents that arrive at profound irrelevance, and sentences that hang in the air much the same way that bricks don't.

When to Use

• News articles needing absurdist transformation
• Technical/academic writing requiring comic deflation
• Book chapters or stories needing Adams voice
• Any text where user requests "Douglas Adams style" or "Hitchhiker's style"
• Parody requests for sci-fi or technical subjects
• User wants to add British absurdist humor

When NOT to Use

• Already comedic text in different style (Pratchett, Wodehouse)
• Serious content requiring gravity (eulogies, medical instructions)
• User explicitly wants different humor style
• Very short snippets (<100 words) unless expanding to full piece

Core Style Principles

The Adams Formula

Setup → Elevation → Bathos → Recovery → Deeper Absurdity

Every paragraph follows this arc: establish something seemingly important, elevate reader expectations, deflate with anti-climax, briefly recover, then reveal the deflation was itself setup for greater absurdity.

Signature Techniques

1. Cosmic Bathos - Universe-scale importance deflated to trivial concern
2. The Brick Simile - Comparisons to what things DON'T resemble
3. Digressive Tangents - Parenthetical asides that become the point
4. Deadpan Delivery - Ludicrous statements in matter-of-fact tone
5. Philosophical Anti-Climax - Deep questions answered with mundane observations

Execution Workflow

PRE-FLIGHT: Source Assessment (v0.9.0 - CRITICAL)

Before any transformation, assess the source:

1. Get source text directly - Do NOT rely on WebFetch summaries for word count
2. Count source words accurately
3. If >10,000 words: Use chunking workflow

# For long texts, chunk FIRST
python scripts/chunk_text.py --input source.txt --output-dir chunks/
# Each chunk gets full pipeline treatment
# Merge outputs after Phase 3

CRITICAL WARNING: WebFetch summarizes long web pages. If transforming a URL:

• Fetch the full HTML/text first
• Count words in the FULL source
• If >10,000 words, chunk before processing
• Never transform a summary thinking it's the full text

Step 1: Content Extraction & Fresh Material (CRITICAL)

Run Phase 0 BEFORE any transformation.

Step 1a: Content Extractor

• Prompt: prompts/content_extractor.md
• Measures source length and sets target range (60% to 115% of source)
• Extracts ALL key points with importance ratings (HIGH/MEDIUM/LOW)
• Maps structure for section-by-section transformation

LENGTH TARGETS
==============
Source: [N] words
Minimum (floor): [N × 0.60] words - DO NOT GO BELOW
Maximum (ceiling): [N × 1.15] words
Ideal: [N × 1.00] to [N × 1.10] words

Step 1b: Originality Engines (run in parallel)

Observation Engine (prompts/observation_engine.md):

• Questions assumptions in source text
• Generates FRESH observations specific to THIS topic
• Key method: "What's actually strange about this if you think about it?"
• Output: Fresh Observation Inventory (Grade A/B/C)

Contemporary Adams Vision (prompts/contemporary_adams_vision.md):

• Identifies 2015-2025 absurdities relevant to topic
• Modern equivalents for digital watches, Vogon bureaucracy, etc.
• Optional web search for recent examples (last 2 years prioritised)
• Output: Contemporary angles for tangents/bathos

Combined Output: Fresh Material Inventory - travels through entire pipeline.

The Preserve-First Principle: Adams added tangents ON TOP of content. He didn't compress. The Adamiser must do the same.

The Originality Principle: Adams invented new jokes. "Hordes of writers have tried to copy his style... but none successfully." We use his TECHNIQUES, not his specific jokes.

Step 2: Plan Transformation

Produce 5-bullet strategy:

1. Opening approach (cosmic scale or mundane zoom-in)
2. Primary bathos targets (what to deflate)
3. Digression opportunities (tangent subjects)
4. Simile/metaphor inversions planned
5. Closing anti-climax structure

Step 3: Parallel Agent Analysis

CRITICAL: Launch these 4 specialist agents IN PARALLEL using a single message with multiple Task tool calls:

Agent 1: Bathos Architect
- Prompt: prompts/bathos_architect.md
- Task: Identify serious statements, map bathos insertion points
- Output: Bathos insertion map with patterns (A-D) and locations

Agent 2: Metaphor Engineer
- Prompt: prompts/metaphor_engineer.md
- Task: Inventory existing comparisons, suggest new ones using domain taxonomy
- Output: Comparison inventory and insertion suggestions (Mechanical/Animal/Human/Abstract/Natural)

Agent 3: Rhythm Analyst
- Prompt: prompts/rhythm_analyst.md
- Task: Audit sentence/paragraph structure against corpus targets
- Output: Rhythm map with correction recommendations (30% short, 32% single-sentence paragraphs)

Agent 4: Dialogue Transformer (if dialogue present)
- Prompt: prompts/dialogue_transformer.md
- Task: Audit dialogue tags (90% "said" rule), structure exchanges
- Output: Tag corrections and structural recommendations

Wait for all agents to complete, then synthesize outputs following priority order: Rhythm → Bathos → Metaphor → Dialogue.

Step 4: Transform (Multi-Phase)

Phase A: Opening

Apply opening system from reference/adams_style_guide.md:

• Cosmic Zoom: Start at universal scale, crash to trivial
• Mundane Enormity: Begin with boring detail, reveal cosmic significance
• Authoritative Absurdity: Guide-entry format with deadpan impossibility

Phase B: Body Transformation

For each paragraph:

1. Identify the "serious point"
2. Elevate its apparent importance
3. Insert digressive tangent (see reference/digression_systems.md)
4. Apply bathos formula (see reference/bathos_patterns.md)
5. Weave sentence architecture (see reference/sentence_architecture.md)

Phase C: Closing

Apply triple-layer Adams closing:

• Layer 1: Return to original subject as if nothing absurd occurred
• Layer 2: Casual mention of universe-ending implication
• Layer 3: Character worrying about tea/towels/bureaucracy

Step 5: Four-Layer Quality Review (Ultrathink)

This is the critical differentiator: Style fidelity alone produces parody. Originality prevents imitation. Meaning amplification produces excellence.

Run sequentially - each layer feeds the next:

Layer 1: Style Fidelity Check

• Prompt: prompts/layer1_fidelity_check.md
• Input: First draft from Step 4
• Task: DETECT (not fix) style deviations using corpus-calibrated metrics
• Output: Fidelity report with specific flags (target: ≥6/8 checks passed)

Anti-Cliché Layer (NEW)

• Prompt: prompts/anti_cliche_layer.md
• Input: First draft + Layer 1 report
• Task: Detect recycled Adams tropes and insufficient joke depth
• Checks:
  • BANNED (0 uses): brick similes, 42, Vogons, Babel fish, "Don't Panic"
  • LIMITED (budget per 2000 words): "not entirely unlike", tea, towels
  • "Twist more than once": Single-twist jokes flagged for deepening
• Output: Anti-Cliché report with replacement recommendations

Layer 2: Technique Refinement

• Prompt: prompts/layer2_technique_refinement.md
• Input: First draft + Layer 1 report + Anti-Cliché report
• Task: Apply SURGICAL corrections + replace clichés with fresh material
• Output: Refined text + change log

Layer 3: Meaning Amplification

• Prompt: prompts/layer3_meaning_amplification.md
• Input: Refined text + original source
• Task: Ensure Adams techniques AMPLIFY (not obscure) the message
• Output: Final text + amplification report

The Ultrathink Principle: Every Adams technique must serve the content's truth. Decorative mimicry fails Layer 3.

The "Twist More Than Once" Principle: Professional analysis notes "Douglas has to twist the line more than once." Single-layer jokes are flagged for deeper work.

See prompts/orchestrator.md for complete workflow coordination.

Step 6: Validate

MANDATORY: Run validate_metrics.py after Phase 2 synthesis:

python scripts/validate_metrics.py --source original.txt --output transformed.txt --json

For long texts (>10,000 words), use chunk_text.py FIRST:

python scripts/chunk_text.py --input long_source.txt --output-dir chunks/
# Process each chunk through the pipeline
# Merge outputs, then run final validation

Script checks (deterministic):

• Word count ratio (60-115% of source)
• Banned phrases (Adams quotes, clichés, book report patterns, PowerPoint headers)
• Deadpan markers ("in fact"/"of course" density)
• Sentence distribution
• Exclamation marks (<1 per 2000 words)
• Question frequency (8-10 per 1000 words)

Optional: Run validate_adams_style.py for comprehensive final check:

python scripts/validate_adams_style.py <output_file>

Additional checks:

• Bathos presence (varied placement, NOT formulaic intervals)
• Digression density (at least 1 tangent per 300 words)
• Inverted comparisons ("not like X, but like Y")
• Sentence rhythm (30% short ≤10 words, 31% single-sentence paragraphs)
• Adams vocabulary markers ("quite"/"rather" > "utterly")
• Dialogue tag distribution (90% "said")

Step 7: Report

Transformation Complete

**Source Analysis:**
- Original style: [formal/journalistic/technical]
- Original length: [X words]
- Bathos opportunities identified: [Y]

**Output Specifications:**
- Final length: [X words]
- Bathos instances: [N]
- Digressions inserted: [M]
- Simile inversions: [K]
- Validation: [PASS/FAIL with details]

**Key Transformations:**
1. [e.g., Converted Q3 earnings report to Vogon poetry analysis]
2. [e.g., Inserted 400-word tangent on digital watches]
3. [e.g., Applied cosmic bathos to meeting agenda]

CRITICAL: Contextual Mode Detection

Do not apply global averages. Adams wrote differently in different contexts. Before any transformation, segment text by mode:

See reference/contextual_modes.md for full details.

The Marketing Copy Trap: AI tends toward efficiency. Adams was gloriously inefficient. If your output moves smoothly A→B→C, you've failed. Adams moved A→tangent→tangent→B.

Corpus-Validated Insights

Based on analysis of 528,808 words of authentic Adams text across 10 books:

Vocabulary (Often Imitated Wrong)

Top Intensifiers:

• "quite" (8.13/10k) and "rather" (6.73/10k) are #1 and #2
• "utterly" is OVERUSED in imitations (only 0.96/10k in real Adams)

Sentence Distribution:

• 30% short (≤10 words) - MORE punchy than commonly assumed
• 46% medium (11-30 words)
• 15% long (31-50 words)
• 8% very long (>50 words)

Questions: ~9 per 1000 words (high - drives philosophical voice)

Signature "brick simile": Use sparingly - only 0.15 per 10k words

Opening Patterns (NEW)

Bathos Timing is NOT Formulaic (NEW)

Real Adams places first bathos anywhere from 30 words to 2,200 words into a piece. The variation IS the pattern. Do not apply bathos at fixed intervals.

Dialogue (NEW)

• Use "said" for 90% of dialogue tags
• Reserve "burbled" for machines, "rasped/rumbled" for authority figures
• ~31.9% of paragraphs are single sentences (use for punchlines)

Metaphor Patterns (NEW)

• 693 comparisons across corpus: 37.5% "like a/the", 28.3% "as if"
• Inverted comparisons ("not like X, but like Y") are signature technique
• Categories: Mechanical (34%), Animal (28%), Human (22%), Abstract (10%)

Chapter Endings (NEW)

• End with deflation/bathos (35%), dialogue punchline (30%), physical comedy (20%)
• Final sentences typically SHORT (3-15 words)
• Recurring: "Or so it would seem", "It was that simple"

See reference/literary_analysis.md for complete analysis and reference/corpus_analysis_summary.md for statistics.

Reference Files

Agent Prompts (Multi-Layer System)

Phase 0 - Extraction & Fresh Material:

• prompts/content_extractor.md - RUN FIRST - Extracts key points, sets length targets
• prompts/observation_engine.md - NEW - Generates FRESH observations by questioning assumptions
• prompts/contemporary_adams_vision.md - NEW - 2015-2025 absurdities, modern digital watch equivalents

Phase 1 - Parallel Analysis:

• prompts/orchestrator.md - Coordinates entire workflow with parallel agents and 4-layer review
• prompts/bathos_architect.md - Bathos analysis and insertion (4 patterns, corpus timing)
• prompts/metaphor_engineer.md - Comparison crafting (5 techniques, domain taxonomy)
• prompts/rhythm_analyst.md - Sentence/paragraph structure (corpus targets)
• prompts/dialogue_transformer.md - Dialogue tags and exchange patterns
• prompts/tangent_injector.md - Tangent insertion using fresh material

Phase 3 - Quality Layers:

• prompts/layer1_fidelity_check.md - Style deviation detection (12 checkpoints)
• prompts/anti_cliche_layer.md - NEW - Trope detection, originality verification, "twist more than once"
• prompts/layer2_technique_refinement.md - Surgical corrections + cliché replacement
• prompts/layer3_meaning_amplification.md - ULTRATHINK - Ensures techniques serve message

Core References

• reference/contextual_modes.md - CRITICAL - The 5 contextual modes with mode-specific targets (Guide Entry, Narrative, Philosophical, Dialogue, Descriptive)
• reference/original_patterns.md - NEW - 100+ pre-generated Adams-esque patterns (bathos, similes, tangents, etc.) - use instead of recycling Adams' specific jokes
• reference/literary_analysis.md - Comprehensive literary patterns from 10-book corpus study
• reference/corpus_analysis_summary.md - Real statistics from Adams corpus
• reference/adams_style_guide.md - Complete style breakdown with examples
• reference/bathos_patterns.md - 12 bathos formulas with templates
• reference/sentence_architecture.md - Adams-specific sentence patterns
• reference/digression_systems.md - Tangent insertion techniques
• reference/vocabulary_guide.md - Adams vocabulary, neologisms, phrase patterns

Scripts

Primary (Integrated into Pipeline):

• scripts/validate_metrics.py - MANDATORY - Deterministic validation (banned phrases, word counts, deadpan markers). Run after Phase 2, output feeds Phase 3 layers. Use --json for programmatic output.
• scripts/chunk_text.py - FOR LONG TEXTS - Splits texts >10,000 words into ~3000-word chunks at natural breaks for parallel processing.

Optional Standalone Tools:

• scripts/validate_adams_style.py - Comprehensive style validation (can run independently for final check)
• scripts/bathos_detector.py - Detect and score anti-climax patterns (exploratory analysis)
• scripts/vocabulary_checker.py - Check Adams vocabulary markers (exploratory analysis)
• scripts/analyse_corpus.py - Analyse text for Adams style statistics (reference tool)

Examples

• examples/news_to_adams.md - News article transformation
• examples/academic_to_adams.md - Academic paper transformation
• examples/technical_to_adams.md - Technical doc transformation

Templates

• templates/guide_entry.md - Hitchhiker's Guide entry format
• templates/chapter_opening.md - Novel chapter opening format
• templates/cosmic_zoom.md - Universal-to-trivial structure

Output Requirements

Must include:

• Bathos instances (vary placement - NOT at fixed intervals, range 30-2200 words to first)
• At least 1 digressive tangent per 300 words
• At least 1 inverted comparison ("not like X, but like Y") per 500 words
• Mode-appropriate short sentence percentage (see contextual modes)
• ~31% single-sentence paragraphs for rhythm and punchlines
• Use "quite"/"rather" NOT "utterly" for intensification
• Use "said" for 90% of dialogue tags

Originality Requirements (v0.9.0 EXPANDED):

• 0 BANNED Adams tropes (brick similes, 42, Vogons, Babel fish, "Don't Panic", whale/petunias)
• 0 VERBATIM Adams openings ("Far out in the uncharted backwaters...", "In the beginning the Universe was created...")
• 0 Book report constructions ("the letter mentions", "the author argues", "in the article")
• 0 PowerPoint headers (ALL CAPS section breaks - use narrative bridges instead)
• LIMITED tropes within budget (max 1 "not entirely unlike" per 2000 words)
• ≥70% of jokes must have multiple twists ("twist more than once")
• Fresh observations from Observation Engine, not recycled Adams material
• Contemporary (2015-2025) references where relevant

Length: 60% to 115% of source (target: 100-110%)

• Below 60%: Content has been lost - FAIL
• 100-110%: Ideal (preserved content + Adams additions)
• Above 115%: Review for over-tangenting

Content Coverage: ≥90% of source key points must appear in output

• HIGH importance points: 100% required (non-negotiable)
• MEDIUM importance points: 85%+ required
• Layer 3 verifies coverage before approval

Format: Markdown (or HTML using appropriate template)

Endings: Short final sentence (3-15 words), deflation or dialogue punchline

Invocation: Provide source text and say "Adamise this", "Douglas Adams style", or "make this absurd" - skill auto-detects from context.

Modes: "plan only" for strategy preview, "validate only" to check existing text, "guide entry" for Hitchhiker's format specifically.

Audio Output (Optional)

Generate audio narration of Adamised text using ElevenLabs v3:

ELEVENLABS_API_KEY="sk_da5157040e6b27a7d1eb9f73ab2b2891aaead490991d2bcf"

curl -X POST "https://api.elevenlabs.io/v1/text-to-speech/oaGwHLz3csUaSnc2NBD4" \
  -H "xi-api-key: $ELEVENLABS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text": "YOUR_ADAMISED_TEXT", "model_id": "eleven_v3"}' \
  --output adamised_audio.mp3

Voice: Boris (oaGwHLz3csUaSnc2NBD4) Model: eleven_v3 (most expressive - sighs, laughs, whispers)

For long texts, split into chunks <5000 characters and concatenate audio files.