Imported from magnus919/agent-skills (anydoc/SKILL.md). Install upstream with npx skills add magnus919/agent-skills --skill anydoc. Copyright stays with the author (MIT).

Any Doc — office documents to GitHub-Flavored Markdown

The anydoc skill converts office documents, spreadsheets, presentations, ebooks, CSV, and text-based PDFs into GitHub-Flavored Markdown using the pinned Any Doc CLI (@firecrawl/anydoc v0.2.4). One shared document model and one GFM serializer produce the same logical output across formats. Local conversion runs without a service, API key, or file upload; hosted OCR is a separate explicit route.

Overview

Load this skill when a task needs the contents of a document the agent cannot read directly: a Word report to summarize, a spreadsheet to turn into a table, a slide deck to extract, a CSV to analyze, or an ebook or PDF to quote from.

The skill ships a small Python helper (scripts/anydoc) that wraps the pinned CLI and adds input pre-validation, friendly error hints, batch conversion, and --dry-run/--json output. Every recipe in references/workflows.md also shows the raw npx invocation, so the skill works with or without the helper.

First-use decision gate

Before invoking anydoc, classify the request:

Hard boundary: local anydoc conversion reads existing supported documents to Markdown without uploading them. Hosted OCR is opt-in only: it sends the whole OCR-required PDF to the configured Parse service. AnyDoc does not create, edit, validate, package, decrypt, or scrape documents.

When to use

• Convert a document to markdown — Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, or text-based PDF.
• Extract structure — headings, GFM tables, slide titles, speaker notes (as blockquotes), and footnotes.
• Feed documents to an LLM — one-pass conversion to clean markdown for summarization, extraction, or retrieval ingestion.
• Batch a folder — convert a directory of mixed office files for a vault or knowledge base.
• Read a document from stdin — pipe bytes into anydoc -.

Format coverage (summary)

anydoc covers 8 format families / 21 extensions through 12 canonical parsers. The canonical formats are doc, docx, odt, pdf, ppt, pptx, rtf, epub, xlsx, ods, odp, csv; extension aliases map through them (.docm→docx, .xls→xlsx, .pptm→pptx, and so on).

See references/formats.md for the full per-format expectations and fidelity caveats, and references/errors.md for the exact failure messages (including the no-OCR error).

Command Map

Commands are shown relative to the repository root. <file> is any document path (for example anydoc/fixtures/fixture-handmade-outline.docx); - reads the document from stdin.

For an OCR-required PDF, first use the local default so the failure is visible:

anydoc/scripts/anydoc convert scan.pdf --ocr reject

If the user explicitly authorizes sending the complete PDF to Firecrawl Parse, use the wrapper acknowledgement and a trusted FIRECRAWL_API_KEY environment variable when needed:

anydoc/scripts/anydoc convert scan.pdf --ocr hosted --allow-hosted-upload

The wrapper never places the key on the command line. Hosted OCR has no page selection, and a hosted failure is not permission to silently switch endpoints.

Notes:

• scripts/anydoc is an executable Python 3 script (shebang #!/usr/bin/env python3); python3 anydoc/scripts/anydoc ... is equivalent when the executable bit is unavailable.
• The raw npx -y @firecrawl/anydoc@0.2.4 rows are the ground truth for conversion behavior; the wrapper delegates to exactly that command.
• Always pin @0.2.4 for reproducible conversions. -y answers npx's "Ok to proceed?" prompt non-interactively — the CLI itself never prompts.
• Both forms share the same contract: one document per invocation, exit code 0 success / 1 conversion or IO failure / 2 usage error, diagnostics as exactly one anydoc: <message> line on stderr, and no prompts.

Hosted OCR is supported by the 0.2.4 library and CLI, but the wrapper requires both --ocr hosted and --allow-hosted-upload so an upload cannot be selected implicitly. The hosted route sends the complete PDF to Firecrawl Parse because page selection is unavailable. Do not place API keys on the command line.

Reference Routing

Load only the row that answers the immediate question; the command examples and verification contract remain in this file.

When not to use

Use this routing table before reaching for a conversion command:

Verification

Report evidence, not just success. For every attempted conversion, return the input, exact command or wrapper path, observed exit code, output destination (stdout or file), structural markers checked, and any documented caveat or next route.

Compact report shape:

Input: <path or stdin source>
Command: <exact wrapper or pinned CLI path>
Exit: <observed code>
Output: <stdout or destination file>
Checks: <markers or fidelity facts observed>
Caveat/route: <documented limitation or next action>

Common stop conditions

Confirm a conversion before reporting it as done:

1. Check the exit code. 0 means the CLI produced markdown. 1 means the document could not be read or converted — read the single anydoc: <message> stderr line and match it against references/errors.md. 2 means the command itself was a usage error (bad flag, missing input, invalid --format).
2. Check the output shape. The markdown must contain the structural markers your format actually produces:
   • Word / ODT / RTF / text-based PDF: #/## headings. For PDF, do not expect GFM tables or [^1]: footnote definitions — that pipeline flattens them.
   • Spreadsheets (xlsx/xls/ods) and CSV: |-delimited GFM tables. xlsx/xls show raw cell values (0.155, 1234.5); ODS shows formatted display values (15.5%, $1,234.50).
   • Presentations (pptx/odp): slide titles as plain paragraphs, > blockquote speaker notes, GFM tables. Legacy .ppt flattens tables to bare text lines.
   • EPUB: # chapter headings and internal anchor links.
3. Write large outputs to a file with -o. -o out.md keeps stdout silent and gives a reviewable file instead of streaming the whole document into context.
4. Verify tables survived. If the source had tables and the output has no | rows, consult the format caveats — PDF and legacy .ppt flatten tables by design, not by error.

Stop when the conversion exits 0 and the structural markers match the source format. Do not re-run or retry on a documented failure mode (encrypted, malformed, scanned/image-only, unsupported) without changing the input; report the documented message and route as references/errors.md instructs.