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:

If the user needs...	Do this
The contents of an existing supported document	Continue to Command Map.
Generation, editing, validation, EPUB packaging, HTML scraping, or password decryption	Stop and use the route in When not to use.
A format-fidelity or failure decision	Load the matching row in Reference Routing before choosing a command.
A conversion result	Choose stdout, -o, or batch; run it; then follow Verification.

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).

Family	Extensions	Expected GFM output	Decision cue
Word	.doc .docx .docm	#–###### headings, GFM tables, [^n] footnotes	Use when content extraction is enough; use documents when rendered layout matters.
PowerPoint	.ppt .pps .pot .pptx .pptm .ppsx .ppsm	slide titles as plain paragraphs, bullet lists, speaker notes as > blockquotes, GFM tables (PPTX/ODP; legacy .ppt flattens tables to text lines)	Need table fidelity? Prefer PPTX or ODP; legacy .ppt preserves cell text but not table structure.
Excel	.xls .xlsx .xlsm .xlsb	## <sheet name> heading + one GFM table per worksheet; number formats dropped (raw cell values)	Need displayed percentages, currency, or number formats? Prefer ODS; XLS/XLSX output is raw values.
OpenDocument	.odt .ods .odp	same document/slide shapes as DOCX/PPTX; ODS keeps formatted display values	Prefer ODS when spreadsheet display formatting is part of the meaning.
Rich Text Format	.rtf	same document shape as DOCX/ODT	Use for text extraction, not layout preservation.
EPUB	.epub	# chapter headings, GFM tables, internal anchor links	Use to read an existing EPUB; use epub to author or package one.
CSV	.csv	one GFM table; label-like first row promoted to header; delimiter sniffing; UTF-16 with BOM	Use for delimited tabular content; inspect delimiter and encoding when output looks wrong.
PDF	.pdf	headings + inline emphasis, but a lower-fidelity pipeline: tables flatten to text, footnotes and links degrade. Text-based PDFs stay local; scanned/image-only PDFs require explicit hosted OCR or another OCR tool	Use local mode by default; hosted mode uploads the whole PDF and has no page selection.

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.

Need	Command	Choose it when
Convert one file to small markdown on stdout	anydoc/scripts/anydoc convert <file>	The caller needs immediate content and does not need a saved artifact.
Convert one file to a markdown file	anydoc/scripts/anydoc convert <file> -o out.md	The output is large, must be reviewed later, or should be preserved as an artifact.
Convert many files to a directory	anydoc/scripts/anydoc batch <file1> <file2> ... --out-dir out/	The request is a bounded batch and per-file output/status is useful.
Show the tool and pinned CLI version	anydoc/scripts/anydoc info	You need to confirm the executable and version before troubleshooting or reporting an environment issue.
Raw pinned CLI, one document	npx -y @firecrawl/anydoc@0.2.4 <file> [-o out.md]	The wrapper is unavailable; preserve the pinned CLI and its documented semantics.
Raw pinned CLI, read stdin	cat data.csv | npx -y @firecrawl/anydoc@0.2.4 - --format csv	Bytes already arrive on stdin and the format is known; keep the producer pipeline separate from the converter.

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 you need to...	Load	It answers
Choose a format or predict fidelity	references/formats.md	Supported families, output shapes, and caveats such as raw XLSX values, ODS display values, legacy .ppt table flattening, and PDF degradation.
Select flags, stdin syntax, output behavior, or version details	references/cli-reference.md	Verbatim help, accepted options, stdin rules, stdout/stderr behavior, pinning, and runtime requirements.
Classify a failure or decide whether to retry	references/errors.md	Exit codes, exact error vocabulary, no-OCR/encryption boundaries, and the next route.
Choose a single-file, stdin, batch, vault, or large-output recipe	references/workflows.md	End-to-end recipes, safe output handling, per-file failure routing, and resource-limit behavior.
Verify a documented upstream or fixture claim	references/sources.md	Source URLs, access dates, fixture provenance, and the verification basis for documented claims.
Shape the final evidence report	references/report-examples.md	Complete success, expected-failure, and fidelity-boundary reports to imitate after following Verification.

When not to use

Use this routing table before reaching for a conversion command:

User's request	Reach for	Why
Generate, edit, inspect rendered layout, or validate a PDF/Word/Excel/PowerPoint artifact	documents skill	anydoc extracts existing document contents to Markdown; it does not author, preserve rendered layout, or validate artifacts.
Package or author an EPUB	epub skill	anydoc reads an existing EPUB to Markdown but never writes or validates an EPUB container.
OCR a scanned or image-only PDF	Local OCR tooling, or AnyDoc hosted OCR after explicit authorization	Local mode reports the OCR-required error without uploading; hosted mode sends the whole PDF to Firecrawl Parse.
Scrape HTML or other web content	A web-scraping skill	HTML is not a supported anydoc input.
Transcribe binary media such as images, video, or audio	A media or transcription tool	Embedded images become alt text; anydoc cannot transcribe media.
Preserve pagination, fonts, templates, or rendered layout	A document/layout tool	The only output contract is GitHub-Flavored Markdown.
Convert a password-protected file	An unencrypted copy from the document owner	anydoc has no password or decryption option.

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

Condition	Do not	Next
Scanned or image-only PDF / OCR-required error	Retry unchanged or upload implicitly	Use local OCR, or explicitly authorize and run --ocr hosted --allow-hosted-upload; page selection is unavailable.
Encrypted or password-protected document	Guess a password or retry unchanged	Request an unencrypted copy or owner-authorized re-export.
Unsupported, malformed, or resource-limit error	Guess a parser or claim partial success	Match the exact error in references/errors.md and follow its bounded route.
Exit 0 but expected structural markers are absent	Report success from the exit code alone	Inspect the output shape and source fidelity before reporting completion.

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.