datasheet-intelligence（数据表情报）

Datasheet Intelligence

Objective

• Produce evidence-grounded hardware answers and code from PDF/DOCX/XLSX datasheets.
• Prefer fast mode by default; use --structured only when table/header fidelity is required.

Context Policy

• Keep SKILL.md minimal and procedural.
• Run scripts/toc.py, scripts/search.py, scripts/read.py directly before loading extra references.
• Load references/usage.md only for detailed flags or format-specific examples.

Prerequisites

Use uv with this skill's pyproject.toml and uv.lock. Do not rely on PEP 723 inline script metadata.

# Install uv if needed
curl -LsSf https://astral.sh/uv/install.sh | sh

Use one of these execution contexts:

• Recommended (works from any directory): uv run --project skills/datasheet-intelligence...
• Alternative: cd skills/datasheet-intelligence && uv run...

All command examples below use the recommended --project style.

Mandatory Execution Loop

1. MUST identify candidate pages first with scripts/toc.py or scripts/search.py before large reads.
2. MUST read targeted ranges with scripts/read.py --pages and expand iteratively.
3. MUST verify every critical claim (address, bit position, reset value, formula) with source file + page/section.
4. MUST rerun extraction on mismatch or ambiguity (search -> read -> search).
5. MUST follow Tip: / Try: guidance from script errors, then rerun.
6. MUST not finalize the answer until critical code settings are mapped to citations.

Workflow

PDF Datasheets

Choose the strategy by document size.

Small (< 50 pages)

1. Run scripts/toc.py (use --structured if bookmarks are missing).
2. Run scripts/read.py --pages for relevant sections.
3. Add --structured if tables are broken.

Medium (50-150 pages)

1. Run scripts/toc.py --filter to narrow sections.
2. Run scripts/search.py to locate exact pages.
3. Run scripts/read.py --pages for focused ranges (use --structured for table-heavy ranges).

Large (>= 150 pages)

Never read the whole document at once.

1. Run scripts/toc.py to map sections and page ranges.
2. If scripts/toc.py reports no bookmarks, switch immediately to scripts/search.py (search-first) instead of full --structured TOC.
3. Skip low-value sections (Legal, Revision History, Ordering Info, Package Drawing).
4. Run scripts/search.py for exact keyword locations (--unique-pages recommended for long documents).
5. Run scripts/read.py --pages in 10-30 page chunks.
6. Iterate: read -> discover new keywords -> search again -> read again.

High-priority large-PDF sections:

DOCX / XLSX

1. Run scripts/search.py first to find candidate paragraphs/rows.
2. Run scripts/read.py for targeted reading.
3. Use scripts/read.py --structured when layout/table structure is critical.
4. If no hits, expand keywords and retry search before full reading.

Quick Commands

Use --structured only when table/header fidelity is required.

SKILL_DIR="skills/datasheet-intelligence"

# 1) Find candidate pages first
uv run --project "$SKILL_DIR" "$SKILL_DIR/scripts/search.py" docs/rp2040.pdf "IC_CON" "I2C0_BASE" --unique-pages

# 2) Read only selected ranges
uv run --project "$SKILL_DIR" "$SKILL_DIR/scripts/read.py" docs/rp2040.pdf --pages 464-470

# 3) Switch to structured mode only if layout fidelity is critical
uv run --project "$SKILL_DIR" --with docling "$SKILL_DIR/scripts/read.py" docs/rp2040.pdf --pages 464-470 --structured

For full flags and format-specific examples, read references/usage.md.

Operational Rules

1. Start with TOC for PDF workflows.
2. If bookmarks are missing, switch to search-first flow and avoid full structured TOC for very large PDFs.
3. Keep explicit project context in every command (uv run --project...).
4. Read enough neighboring context to avoid missing table headers/footnotes.
5. Cross-check register values against Address Map / Register List sections.

Output Contract

1. MUST provide evidence for each critical claim (address, bit position, reset value, formula): source file + page/section.
2. MUST map important code settings to evidence locations.
3. MUST mark unverifiable values as unverified.
4. MUST report table/prose conflicts and separate uncertain items.

Resources

See usage.md for detailed examples.