Files
Phelan_Varrent/docs/superpowers/specs/2026-03-10-manuscript-analysis-skill-design.md
2026-03-10 14:16:45 -05:00

8.5 KiB

Manuscript Analysis Skill — Design Spec

Date: 2026-03-10 Purpose: Reusable skill to analyze manuscript chapters for repeated phrases, continuity errors, and story craft quality.


Context

Book 1 of The Shade Series is complete at ~104K words across 21 chapters. Before publication and during future book development, the author needs repeatable quality checks that catch:

  1. Repeated/overused phrases across the manuscript
  2. Continuity errors, timeline inconsistencies, and world-building contradictions
  3. Story craft issues (pacing, character arcs, foreshadowing, POV consistency)

The challenge: the full manuscript exceeds what fits in a single context window. This skill solves that with a multi-agent pipeline where each agent gets a fresh context window and a focused task.


Invocation

/manuscript-analysis [book-number] [chapter-range]

Examples:

  • /manuscript-analysis 1 1-5 — analyze chapters 1-5 of book 1
  • /manuscript-analysis 1 16-21 — analyze chapters 16-21
  • /manuscript-analysis 2 1-7 — future book 2

The skill parses the book number and chapter range, validates that the chapter files exist (e.g., chapters/book1/ch01-final.md through ch05-final.md), and creates the output directory if needed.


Architecture: Multi-Agent Pipeline

The orchestrator (main session) dispatches subagents via the Agent tool. Each subagent gets a clean context window. The orchestrator does NOT perform analysis itself — it coordinates and aggregates.

/manuscript-analysis 1 1-5
         |
    ORCHESTRATOR
         |
    Phase 1: PHRASE ANALYSIS
         |-- Agent 1A: Grep-based phrase indexing (target chapters + Grep across full book)
         |-- Agent 1B: Literary phrase analysis (target chapters + index from 1A)
         |
    Phase 2: CONTINUITY CHECK
         |-- Agent 2A: Cross-reference target chapters against all reference docs
         |-- (Agent 2B: if range > 7 chapters, split the work)
         |
    Phase 3: STORY CRAFT
         |-- Agent 3A: Narrative analysis against story summary + outline
         |
    Phase 4: AGGREGATION
         |-- Orchestrator reads intermediate reports, writes consolidated report

Why Subagents?

Each agent gets a fresh ~200K token context window. Without subagents, analysis of even 5 chapters plus reference docs would consume the entire session context, degrading output quality and preventing further work.

Why Grep for Phrase Indexing?

The Grep tool searches the filesystem directly — it does not load files into the context window. This lets Agent 1A query the full 104K-word manuscript for phrase frequencies without exceeding its context budget. The tradeoff: Grep only catches exact string matches, not semantic repetition. Agent 1B handles the semantic/structural analysis.


Phase Details

Phase 1: Repeated Phrase Analysis

Agent 1A — Phrase Indexer

  • Loads: target chapter files only (~33K tokens for 5 chapters)
  • Process:
    1. Read target chapters fully
    2. Extract candidate phrases: multi-word expressions (2-4 words) appearing 3+ times within target range
    3. Also flag distinctive single words used excessively (adverbs, specific adjectives)
    4. For each candidate, use Grep to count occurrences across ALL chapter files in the book
    5. Record: phrase, count in target range, count in full book, which chapters contain it
  • Output: notes/analysis/book{N}-phrase-index-ch{XX}-{YY}.md
  • Context budget: ~40K tokens

Agent 1B — Phrase Analyst

  • Loads: target chapter files + phrase index from 1A + CLAUDE.md voice guidelines
  • Process:
    1. Read chapters with a literary editor's eye
    2. Identify repetitive sentence structures (not just repeated words)
    3. Flag overused transitions, beats, and narration patterns
    4. Distinguish intentional voice patterns (Phelan's established mannerisms per CLAUDE.md) from authorial habits
    5. Note phrases that work as intentional callbacks vs. unintentional repetition
  • Output: notes/analysis/book{N}-phrase-analysis-ch{XX}-{YY}.md
  • Context budget: ~45K tokens

Phase 2: Continuity & Consistency Check

Agent 2A — Continuity Checker

  • Loads: target chapters + story summary + timeline + exploits log + relevant character files + economy + location files
  • Checks:
    1. Timeline: Bell references, day-of-week, elapsed time vs. timeline-book{N}.md
    2. Characters: Name spelling, knowledge consistency, physical descriptions vs. character files
    3. World-building: Currency amounts vs. economy.md, location details vs. location files, magic usage vs. runic-flow rules
    4. Plot threads: Threads from earlier chapters (per story summary) referenced appropriately? Dangling threads?
    5. Internal consistency: Facts within the target range don't contradict each other
  • Output: notes/analysis/book{N}-continuity-ch{XX}-{YY}.md
  • Context budget: ~85K tokens (for 5 chapters)

Splitting rule: If chapter range > 7 chapters, split into two agents (2A + 2B), each handling half the range. Both load the full reference docs.

Phase 3: Story Craft Feedback

Agent 3A — Story Craft Analyst

  • Loads: target chapters + story summary + book outline + CLAUDE.md + book-specific CLAUDE.md
  • Analyzes:
    1. Premise & conflict: Central conflict clear and advancing? Stakes escalating?
    2. Character arcs: Phelan's growth visible? Supporting cast serving their functions?
    3. Foreshadowing: Seeds planted and paid off? Setups subtle enough?
    4. POV consistency: Every paragraph sounds like Phelan? No slips into omniscient?
    5. Pacing: Action/investigation/quiet beats distributed well?
    6. World-building integration: Detail woven into action or dumped as exposition?
    7. The Noise convention: Parenthetical tangents at right frequency and doing narrative work?
    8. What's working well: Specific callouts of strong passages and techniques
  • Output: notes/analysis/book{N}-storycraft-ch{XX}-{YY}.md
  • Context budget: ~75K tokens

Phase 4: Aggregation

The orchestrator reads all intermediate reports and writes a consolidated final report:

  • Organized by severity: Critical (plot holes, continuity breaks) > Important (repeated phrases, voice slips) > Suggestions (craft improvements)
  • Deduplicates where multiple agents flagged the same issue
  • Summary at top with counts per category
  • Output: notes/analysis/book{N}-analysis-ch{XX}-{YY}.md

Context Budget Summary (5-chapter range)

Agent Chapter Text Reference Docs Total
1A (Phrase Indexer) ~33K tokens minimal ~40K
1B (Phrase Analyst) ~33K tokens ~10K ~45K
2A (Continuity) ~33K tokens ~50K ~85K
3A (Story Craft) ~33K tokens ~40K ~75K

All well within the ~200K token window per agent.


Output Structure

notes/analysis/
  book1-phrase-index-ch01-05.md      # Raw phrase frequency data
  book1-phrase-analysis-ch01-05.md   # Literary phrase analysis
  book1-continuity-ch01-05.md        # Continuity findings
  book1-storycraft-ch01-05.md        # Story craft findings
  book1-analysis-ch01-05.md          # Final consolidated report

Previous analysis files are overwritten on re-run (same book/range).


Reference Files Used

File Used By Purpose
chapters/book{N}/chXX-final.md All agents Source manuscript text
world/story-summary-book{N}.md Phases 2, 3 Narrative context proxy for chapters not being directly analyzed
world/timeline-book{N}.md Phase 2 Timeline consistency reference
world/magic/exploits-log.md Phase 2 Magic system consistency
world/magic/runic-flow-rules.md Phase 2 Magic system rules
world/economy.md Phase 2 Currency/pricing reference
world/locations/*.md Phase 2 Location detail reference
characters/*.md Phase 2 Character detail reference
outline/book{N}-outline.md Phase 3 Intended vs. actual pacing
CLAUDE.md Phases 1B, 3 Voice guidelines, style rules
chapters/book{N}/CLAUDE.md Phase 3 Book-specific premise/themes

Skill File Location

~/.claude/skills/manuscript-analysis/SKILL.md


Verification

After implementation, test with:

  1. /manuscript-analysis 1 1-3 — small range, should complete with all 4 intermediate + 1 consolidated report
  2. Check that phrase index contains Grep counts across all 21 chapters (not just 1-3)
  3. Check that continuity report references timeline and character files
  4. Check that story craft report addresses all 8 analysis categories
  5. Verify consolidated report has severity tiers and deduplication