| name | ffmpeg-docs |
| description | Search and fetch official FFmpeg documentation from ffmpeg.org: ffmpeg, ffmpeg-all, ffmpeg-filters, ffmpeg-codecs, ffmpeg-formats, ffmpeg-protocols, ffmpeg-devices, ffmpeg-bitstream-filters, ffmpeg-utils, ffmpeg-scaler, ffmpeg-resampler, ffprobe-all, ffplay-all, general, faq, platform, developer. Use when the user asks to look up an ffmpeg option, find filter parameters, check a muxer or demuxer option, search ffmpeg docs, get the official description of a codec flag, find out what an ffmpeg flag does, or verify ffmpeg syntax against the real documentation.
|
| argument-hint | [query] |
FFmpeg Docs
Context: $ARGUMENTS
Quick start
- Find a filter / option: → Step 2 (
search --query <term>)
- Read the full section for a filter: → Step 3 (
section --page ffmpeg-filters --id <anchor>)
- Grab an entire doc page: → Step 4 (
fetch --page <name>)
- Prime cache for offline use: → Step 5 (
index)
When to use
- User asks "what does
-X do in ffmpeg?" or "what are the parameters for filter Y?"
- Need to verify a flag name exists before recommending it (prevents hallucinated options).
- Need the exact option table for a muxer / demuxer / protocol / device.
- Need to cite the canonical ffmpeg.org URL in a response.
- Before writing any non-trivial ffmpeg command, check the current doc for the filter/option you're about to use.
Step 1 — Know the page catalog
The script only works against a fixed list of known ffmpeg.org doc pages. Get the list:
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py list-pages
Common picks:
| Question | Page |
|---|
| "What does filter X do?" | ffmpeg-filters |
"What are the options for -c:v libx264?" | ffmpeg-codecs |
| "What HLS muxer options exist?" | ffmpeg-formats |
"What does srt:// accept?" | ffmpeg-protocols |
| "How do I capture with avfoundation?" | ffmpeg-devices |
| "What bitstream filters are there?" | ffmpeg-bitstream-filters |
| "ffprobe output fields?" | ffprobe-all |
"Expression syntax (eval, if, gt)?" | ffmpeg-utils |
| "Scaler flags?" | ffmpeg-scaler |
| "Resampler options?" | ffmpeg-resampler |
Read references/pages.md for the full catalog with descriptions.
Step 2 — Search first (this is the default)
When the user names a filter, option, muxer, or codec, search across all pages:
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py search --query "tonemap" --limit 5
When you already know which page, scope the search to it (faster, less noise):
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py search --query "hls_time" --page ffmpeg-formats
Output format for each hit:
--- <page>:<line> — <nearest heading>
<canonical URL with anchor>
<snippet with ±3 lines of context>
Use --format json for machine-parseable output when chaining into another tool.
First run downloads the page (~1–2s). Subsequent runs hit the local cache (~/.cache/ffmpeg-docs/) — instant.
Step 3 — Read one section in full
When the search hit points to a specific filter/option and you want the whole block (option list + description), use section:
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py section --page ffmpeg-filters --id tonemap-1
--id accepts either:
- An anchor id printed in search results as
[§xxxx] (e.g. tonemap-1, Muxer-Options).
- A heading keyword — the script falls back to the first heading matching the string.
Output is the section from its heading down to the next same-or-higher-level heading.
Step 4 — Fetch a whole page
Only when you need to dump the entire page (rare — usually overkill):
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py fetch --page ffmpeg-filters
Pair with --format json for structured handoff.
Step 5 — Prime the cache (optional)
For reliable offline lookups or before a burst of queries:
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py index
Fetches every known page, stores text-extracted versions in ~/.cache/ffmpeg-docs/. Run once; re-run only when you need fresh docs (new ffmpeg release, upstream doc edits).
To override the cache location: export FFMPEG_DOCS_CACHE=/path/to/dir.
To clear: uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py clear-cache.
Gotchas
- Never recommend an ffmpeg flag without searching first. Half the problem this skill solves is preventing hallucinated options. If
search --query "<flag>" returns zero hits, the flag doesn't exist on that page — don't claim it does.
- Cache is keyed by page name only, not by ffmpeg version. If the user is on a specific ffmpeg version and the latest docs don't match, re-fetch with
--no-cache or clear the cache.
- The text extraction is lossy for complex tables. The script converts ffmpeg.org HTML → text (headings,
dt/dd option pairs, code blocks preserved; complex multi-column tables flattened). If a search hit looks incomplete, open the URL printed in the hit header and read the original page.
- Anchors inside search results have the form
[§anchor-id] (a literal § sentinel). When passing to section --id, drop the [§ / ] brackets — just use the raw id.
- Some ffmpeg.org pages are MASSIVE (
ffmpeg-filters is ~2 MB of HTML). Searching the full page is fast because it hits the cache; re-fetching with --no-cache takes a few seconds. Don't fetch the whole page into the conversation — search with --limit or section are almost always better.
- Filter names are case-sensitive in ffmpeg but search is case-insensitive. The search will match regardless; the CLI invocation you recommend must use the exact case.
ffmpeg-all / ffprobe-all / ffplay-all are single-file concatenations of the per-topic pages. They're slower to search (big files) and duplicate content. Prefer the topic-specific page (ffmpeg-filters, ffmpeg-formats) unless you explicitly want everything.- Libav* pages (
libavutil, libswscale, libswresample, libavcodec, libavformat, libavdevice, libavfilter) are C API docs, not CLI docs. Skip them for command-line questions — they rarely contain flag/option info.
- This skill does not read doxygen (
/doxygen/trunk/…). It's scoped to the curated manual pages only.
Examples
Example 1 — "What are the options for the tonemap filter?"
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py search --query "tonemap" --page ffmpeg-filters --limit 5
Pick the hit with heading ## [§tonemap-1] …, then:
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py section --page ffmpeg-filters --id tonemap-1
Cite the URL printed in the hit header when responding.
Example 2 — "What are the valid hls_segment_type values?"
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py search --query "hls_segment_type" --page ffmpeg-formats
Example 3 — "What does -movflags +faststart actually do?"
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py search --query "faststart" --page ffmpeg-formats
Example 4 — "Does the sr filter still exist or is it deprecated?"
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py search --query "sr" --page ffmpeg-filters --regex --limit 3
Use --regex with anchored patterns (^sr\b) when a short term has many false positives.
Example 5 — "What color expressions does drawtext accept?"
uv run ${CLAUDE_SKILL_DIR}/scripts/ffdocs.py section --page ffmpeg-utils --id Color
Troubleshooting
Error: unknown page: foo
Cause: The name isn't in the catalog.
Solution: Run list-pages to see valid names. Common mistakes: using filters instead of ffmpeg-filters; using protocols instead of ffmpeg-protocols.
Error: urlopen error [SSL: CERTIFICATE_VERIFY_FAILED]
Cause: System certificate store is out of date (mostly macOS with older Python installs).
Solution: Install/update certifi, or run /Applications/Python\ 3.x/Install\ Certificates.command. As a last resort, set SSL_CERT_FILE env var to a valid CA bundle path. Do NOT patch the script to disable SSL verification.
Search returns zero hits
Cause: Term doesn't exist on that page, or you're searching a page the term isn't documented on.
Solution: Drop --page to search all pages, or try a broader query. Some options live in ffmpeg-utils (expressions, colors) rather than the obvious page.
Results look truncated / tables broken
Cause: Text extraction flattens complex HTML tables.
Solution: The search-hit header prints the canonical URL with anchor. Open it directly (WebFetch or browser) for the authoritative view.
Cache is stale after ffmpeg upstream update
Solution: clear-cache then index (or just fetch --no-cache --page <name> for a single page).
