Engineering Artifact Reading

Read and extract from engineering documents accurately. This is the reading half of the research loop, made explicit: most research failures are reading failures — answering from one page, from a title, or from memory.

Input Gate

Before searching, name the question, the artifact/scope (or say it is a literature review), the jurisdiction/edition, and the effort budget (user-set or thorough). Ask ONE clarifying question if the ask is too vague, then proceed. See references/input-gate.md.

Workflow

1. Map the document first. Before answering, parse enough to know the structure: title, revision/edition, table of contents, sections, tables, figures, appendices, and where the normative (required) content lives vs the informative (guidance) content.
2. Anchor every extract to a location. Page/section/table/figure numbers for PDFs and standards; sheet + zone for drawings; clause + line for code. Keep table values, figure labels, part numbers, and quoted snippets tied to their location so they can be re-checked.
3. Read the pages the answer depends on. Do not answer from a single visible page when the question spans methods, tables, notes, or appendices. For a datasheet, read the notes under the tables — that is where the limits live.
4. Cross-check — when a conclusion depends on multiple parts of a document (a value in a table, a note, a referenced section), verify they agree. Check normative references when the source points at another document.
5. Record provenance — save extracted notes with their source locations as artifacts for the research run. Never strip the location when handing findings to synthesis.

Reading by artifact type

• Standard/code: section + edition + table/figure number; distinguish normative (shall) from informative (should/may) and commentary.
• Datasheet: the part number and revision; table notes; absolute maximum ratings vs recommended operating conditions; test conditions behind specs.
• Drawing: title block (revision, scale), sheet layout, zone grid, dimension units and tolerances, general notes.
• Paper/report: abstract, methods, results, limitations — do not infer results from the abstract.
• Specification: scope, referenced documents, requirement clauses, and the verification method each requirement is measured against.

Subagent Dispatch Mode

When dispatched as an isolated subagent for document parsing (e.g., by /proposal):

1. Receive file paths — one or more document paths to parse
2. Extract content — for each file, run the reader that matches the type:

   # PDFs: page-anchored text with [[page N]] markers
   node scripts/extract-pdf.mjs <pdf-path-or-url> --json
   # text / Markdown / JSON documents
   node scripts/extract-document.mjs <file-path> --json

   The PDF reader (scripts/extract-pdf.mjs, wrapping the optional pdf-parse dependency) stamps [[page N]] boundaries and reports the source sha256, so extracts stay anchorable after the binary is deleted.
3. Parse the JSON result — extract:
   • markdown — full document content in markdown format
   • pages — page count
   • method — extraction method used (text, pdf-parse, vision)
   • warnings — any quality issues (scanned PDF, missing optional dependency, etc.)
4. Return structured content with:
   • Full text (markdown format)
   • Page count
   • Extraction method used
   • Source locations for key sections (page numbers)
   • Warnings and quality notes
5. Record provenance — document what was extracted, how, and any limitations

Output format for subagent return:

{
  "source": "path/to/file.pdf",
  "method": "pdf-parse",
  "pages": 5,
  "content": "full markdown content...",
  "sections": [
    {"title": "Research Areas", "page": 1, "text": "..."},
    {"title": "Requirements", "page": 2, "text": "..."}
  ],
  "warnings": []
}

If extraction method is vision: general artifact-reading may use the host's explicit read_file or vision capability, but /proposal deterministic intake must stop and require a separately recorded text transcription; it must not silently feed vision output into the binder.

Boundaries

• Never fill a gap you could not read. Mark it blocked and say what was unreadable and why (paywall, missing page, unreadable scan, binary format).
• Do not answer from a title, filename, or snippet when a direct read is possible — that is the integrity commandment, applied to reading.
• Research-only, not for final engineering sign-off. Outputs support engineering research but must be reviewed by a licensed engineer for any design application.