Large Doc Mastery

Large Document / Codebase Mastery

Summaries lose information. "Analyze this 80k-token codebase" with no structure produces a vague paragraph about architecture that misses the three critical bugs and the two deprecated modules that half the new code still imports.

The skill is not consuming large context — Claude does that automatically. The skill is structuring the task so the synthesis is precise: specific questions, specific output formats, explicit cross-reference instructions, and verification passes that confirm the model used the actual document rather than its priors.

---

SLASH COMMANDS

| Command | Action | | --- | --- | | /map | Build a structural map of the document (sections, components, entities) | | /index | Find every location in the document that touches a specific topic | | /extract | Extract all instances of a type: decisions, risks, assumptions, TODOs, etc. | | /cross-ref | Find all connections between two topics, sections, or entities | | /contradictions | Surface any internal contradictions or inconsistencies in the document | | /gaps | Identify what's absent — what the document should address but doesn't | | /trace | Follow an entity (function, concept, person, requirement) through the full document | | /timeline | Extract a chronological sequence of events, changes, or decisions | | /compare | Directly compare two sections or files | | /synthesize | Answer a specific question using the full document as context | | /verify | Check whether a specific claim is supported by the document | | /hotspots | Identify the highest-complexity or highest-risk areas | | /summary-lossy | Produce a short summary with explicit list of what was dropped |

---

HIGH-LEVEL WORKFLOW

User provides large document / codebase
    │
    ├─ Phase 1: Document Orientation
    │     Build structural map; identify key entities and sections
    │
    ├─ Phase 2: Scope Definition
    │     Clarify the synthesis goal — don't analyze everything equally
    │
    ├─ Phase 3: Targeted Extraction
    │     Pull specific patterns, entities, or facts with location references
    │
    ├─ Phase 4: Cross-Reference Pass
    │     Connect extracted elements across sections or files
    │
    ├─ Phase 5: Synthesis
    │     Answer the synthesis question with evidence citations
    │
    └─ Phase 6: Verification
          Confirm claims are in the document, not in prior knowledge

---

PHASE 1 — DOCUMENT ORIENTATION

Before any analysis, build an orientation map. This prevents anchor bias — where the first section dominates the analysis and the rest is skimmed.

Orientation map format

DOCUMENT MAP: [title / description]
Total scope: [estimated sections / files / pages]

STRUCTURE
  [Section/File 1]: [1-line description] — [key entities defined]
  [Section/File 2]: [1-line description] — [key entities defined]
  ...

KEY ENTITIES
  [Entity type]: [list] — (e.g., Functions, Classes, Requirements, Actors)

DENSITY DISTRIBUTION
  Heavy:    [sections with most content / complexity]
  Light:    [sections with least content]
  Missing:  [sections referenced but absent]

---

PHASE 2 — SCOPE DEFINITION

Large documents contain far more than any single analysis needs. Before extracting, define the synthesis aperture — what lens is this analysis using?

Synthesis aperture types

| Aperture | Question being answered | Best for | | --- | --- | --- | | Decision audit | What decisions were made, by whom, and why? | Meeting transcripts, PRDs, ADRs | | Risk surface | What can go wrong, and where? | Architecture docs, contracts, plans | | Dependency map | What depends on what? | Codebases, system designs | | Consistency check | Does the document contradict itself? | Legal docs, specs, policies | | Entity trace | How does X evolve through the document? | Code, narratives, regulations | | Gap analysis | What's missing that should be here? | Requirements, audits | | Change surface | What changed between versions / sections? | Diffs, revisions, amendments |

Scope question rules

• Ask ONE scope-defining question if the synthesis goal is ambiguous
• If the user says "analyze everything," default to: Risk surface + Dependency map
• Never attempt all apertures simultaneously — output becomes noise

---

PHASE 3 — TARGETED EXTRACTION

Extract with location anchors. Every extracted element must cite where it came from.

Extraction format

EXTRACTION: [type] — [aperture]

[ID]. [Extracted element]
     Location: [section / file / line / page]
     Context:  [1 sentence: surrounding context that makes this significant]
     Flags:    [RISK | DECISION | ASSUMPTION | TODO | CONTRADICTION | DEPENDENCY]

Extraction rules

• Cite location for every item. Uncited extractions cannot be verified.
• Do not paraphrase. Quote directly, then interpret in the Context field.
• Flag before filtering. Extract everything matching the type, then rank.
• Note absences. If a required element (e.g., error handling, auth check) is absent, flag as MISSING.
• Distinguish document claims from your inferences. Use "The document states..." vs "This implies..."

High-value extraction targets by document type

| Document type | High-value extraction targets | | --- | --- | | Codebase | TODO/FIXME, dead code, circular deps, missing error handling, hardcoded values | | Legal / contract | Obligations, conditions, termination clauses, liability caps, ambiguous definitions | | PRD / spec | Unverified assumptions, missing acceptance criteria, conflicting requirements | | Research / report | Methodology limitations, sample sizes, confidence intervals, conflicting findings | | Architecture doc | Single points of failure, unstated assumptions, missing components |

---

PHASE 4 — CROSS-REFERENCE PASS

The insight that justifies large-context analysis only appears when you connect elements that are far apart in the document.

Cross-reference format

CROSS-REFERENCE: [entity A] ↔ [entity B]

Connection 1:
  [Entity A] location: [cite]
  [Entity B] location: [cite]
  Relationship: [DEPENDS_ON | CONTRADICTS | MODIFIES | REFERENCES | UNDEFINED]
  Significance: [1 sentence: why this connection matters]

Connection 2:
  ...

SUMMARY
  Total connections found: [N]
  Critical connections (high-risk or decision-affecting): [list]
  Unresolved: [connections implied but not made explicit in the document]

Cross-reference rules

• Always cite both locations — without both citations, the connection cannot be verified
• CONTRADICTS always surfaces first — contradictions are the highest-value finding
• UNDEFINED means the relationship exists but the document doesn't specify it — flag these

---

PHASE 5 — SYNTHESIS

Answer the synthesis question using the extracted and cross-referenced evidence.

Synthesis output format

SYNTHESIS: [question / aperture]

FINDING
[Direct answer in 1–3 sentences]

EVIDENCE
1. [claim] — Source: [location]
2. [claim] — Source: [location]
3. [claim] — Source: [location]

PATTERNS
[Cross-document pattern that only appears with full context]

ANOMALIES
[Elements that don't fit the pattern; outliers; exceptions]

CONFIDENCE
[HIGH / MEDIUM / LOW] — [1 sentence: what limits confidence]

WHAT'S NOT IN THE DOCUMENT
[What the synthesis question required but the document doesn't contain]

---

PHASE 6 — VERIFICATION

After synthesis, run a verification pass to confirm the output is document-grounded.

Verification checklist

For each major claim in the synthesis:

• [ ] Can I cite the exact location in the document where this appears?
• [ ] Am I paraphrasing accurately, or am I inferring beyond what's stated?
• [ ] Is this in the document, or is it from my training data?
• [ ] If I'm connecting two things, are both connections cited?

Verification format

VERIFICATION PASS

Claim: "[claim from synthesis]"
  Status: DOCUMENT-GROUNDED | INFERENCE | TRAINING KNOWLEDGE
  Citation: [location, or NONE]
  Note: [if inference or training knowledge: flag for user review]

Common verification failures in large-context tasks

| Failure | Signal | Fix | | --- | --- | --- | | Prior contamination | Claim is true but not in the document | Flag as TRAINING KNOWLEDGE | | Location drift | Cited location doesn't contain the claim | Re-read and re-cite | | Paraphrase inflation | Summary adds meaning the original doesn't have | Quote directly | | Recency bias | Last section dominates synthesis | Re-run extraction on early sections | | Gap hallucination | Document "says" something it actually omits | Add to WHAT'S NOT IN THE DOCUMENT |

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Sandeeprdy1729
• Source: Sandeeprdy1729/claude-design-skill
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.