multimodal-doc-extraction

name	multimodal-doc-extraction
description	Extract structured data from mixed-media documents (scanned invoices, forms, ID cards, spreadsheets embedded in PDFs) with per-field confidence and human-review routing.
argument-hint	["document types","field schema","confidence threshold","review queue"]
allowed-tools	Read, Write
model	claude-fable-5
capability	Multimodal

Multimodal Document Extraction

Claude 5 Mythos + Fable skill — Multimodal. Optimised for Claude 5 Mythos (reasoning) and Claude 5 Fable (action / tool use), and backward-compatible with Claude Sonnet 4.5 and Opus 4.

TL;DR

Real documents mix scans, forms, tables, and stamps. This skill uses Claude 5's multimodal reasoning to extract typed fields, score each field's confidence, and route low-confidence items to a human queue.

When to use this skill

• You are working with Claude 5 Mythos or Fable (or a downstream vendor exposing them).
• The task benefits from multimodal — do not use for tasks a smaller model handles cheaply.
• You need a repeatable, reviewable prompt — not a one-off chat.
• Output has to survive a code review, an audit, or a partner sign-off.

When NOT to use this skill

• Simple lookups a smaller model (Haiku 4 / Sonnet 4.5) can serve at 10× lower cost.
• Real-time UX where 1M-token loads would break the latency budget.
• Tasks where a deterministic script would do — do not model what you can compute.

Inputs (fill these in)

Field	Required	Notes
context_source	✅	Repo path, doc set, ticket, screenshot, or MCP resource id
goal	✅	The single sentence definition of done
audience	✅	Reviewer, end-user, downstream agent, or auditor
constraints	✅	Token budget, step budget, permitted tools, forbidden ops
output_shape	✅	Free text / markdown / JSON schema / patch / PR
citations	⛔ optional	Required for research, legal, medical, financial outputs
escalation	⛔ optional	When to stop and ask a human

Operating Workflow

1. Frame the task

• Restate the goal in one sentence in your own words. If you cannot, ask.
• Identify the smallest input that still contains everything Claude needs.
• Decide whether the task is a read (Mythos), an act (Fable), or read → act (both).

2. Load the context

• Preserve original structure — headers, file boundaries, page numbers.
• Order matters: put stable, high-signal context first; volatile context last.
• Use XML-style tags (<file>, <doc>, <ticket>) to make regions addressable.
• Cache-friendly: keep the top of the prompt stable across calls to hit prompt-cache.

3. Specify the output contract

• Give the exact shape (markdown outline, JSON schema, or diff format).
• State the null policy — what does Claude do when a field is unknown?
• State the citation policy — is every claim traceable to an input region?

4. Run with guardrails

• Cap step count for Fable agents (typical: 8–25 steps).
• Log every tool call, its arguments, its response, and its latency.
• Route destructive ops (rm, DROP, git push --force) through an approval gate.

5. Verify and iterate

• Score output against the contract before returning to the user.
• If invalid, re-ask with the validator error appended — do not silently accept.
• Cache the successful prompt template — it is the reusable artefact.

Output Format

The skill produces:

1. A prompt template (markdown or XML) ready to send to Claude 5.
2. A contract describing the expected response shape.
3. A validator (regex, JSON-schema, or unit-test snippet) that gates the response.
4. A runbook describing how to re-run, refresh cache, and escalate.

Example template head:

<system>
You are Claude 5 Multimodal operating under the "multimodal-doc-extraction" skill.
Follow the output contract exactly. If a required field is unknown, return
null with a `reason` explaining what is missing. Never fabricate citations.
</system>

<context>
  <!-- structured, ordered, addressable -->
</context>

<task>
  <goal>{goal}</goal>
  <constraints>{constraints}</constraints>
  <output_shape>{output_shape}</output_shape>
</task>

Decision Logic

• If the input fits in a single Mythos window → load it whole, no RAG.
• If the input exceeds the window → chunk with overlap, then run a reducer pass.
• If the task requires action → route through Fable with a tool manifest.
• If the task requires verifiable claims → force citations and reject un-cited claims in the validator.
• If confidence is low → escalate rather than guess.

Anti-patterns

• Summarising the input first. Claude 5 does that better than your preprocessor and you lose fidelity.
• Ambiguous tool schemas. Vague names cost you retries — Claude will call the wrong tool.
• Unbounded agent loops. Every Fable run must have a step cap and a wall-clock cap.
• Silent JSON re-parse. If Claude returns invalid JSON, surface the error and re-ask — do not paper over it.
• Cache-hostile prompts. Injecting a timestamp at the top of the prompt kills prompt-cache and doubles your bill.
• Mixing Mythos and Fable in one call. Pick the right variant per step; do not ask Mythos to act or Fable to write a research memo.

Quality Bar (definition of done)

• Output validates against the declared contract on the first pass ≥ 90% of the time.
• Every factual claim in a research-mode output carries a citation.
• Every tool call in an agent-mode output is logged with args, response, and duration.
• The prompt template is checked into the repo and versioned like code.
• A re-run with the same inputs produces a semantically equivalent output.

Worked Micro-Example

Goal. Real documents mix scans, forms, tables, and stamps. This skill uses Claude 5's multimodal reasoning to extract typed fields, score each field's confidence, and route low-confidence items to a human queue.

Inputs.

• context_source: (task-specific — see argument-hint)
• goal: single-sentence definition of done
• output_shape: JSON matching the schema in /schemas/multimodal-doc-extraction.json

Run.

1. Load context in the required order.
2. Send the templated prompt to Claude 5 Multimodal.
3. Validate the response against /schemas/multimodal-doc-extraction.json.
4. On failure, re-ask with the validator error; cap at 2 retries.
5. Persist the successful response and the exact prompt used.

Expected artefact. A machine-readable output plus a human-readable summary that a reviewer can approve in under 5 minutes.

Cadence

Cadence	Action
Per run	Emit trace: prompt hash, model id, tokens in/out, tool calls, validator result
Per week	Review failure traces; tighten the schema, examples, or guardrails
Per month	Rotate examples; retire ones that no longer surface real failure modes
Per quarter	Re-benchmark against the current Claude 5 point release

Companion skills

• prompt-engineering-playbook — reusable prompt patterns across every model.
• mcp-agent-architecture — deeper wiring for MCP-based tool use.
• rag-knowledge-systems — when the corpus outgrows even a 1M window.
• executive-briefs — for turning Claude 5 output into partner-ready one-pagers.

Rules (do not violate)

• Never present Claude 5 output as human-authored for regulated deliverables.
• Never let an agent execute a destructive command without a human gate.
• Never disable the validator "just to unblock a demo" — that is how bad outputs ship.
• Never hardcode secrets into the prompt; use tool-side injection.
• Never remove the AI Content Notice — Claude 5 hallucinates and every output requires human review.

Closing note

Claude 5 Mythos and Fable are powerful because they behave like a careful colleague, not a magic oracle. This skill is the frame that keeps the colleague honest, cheap, and reviewable.