- name
Find the option values an ffmpeg-skill tool accepts without anyone having decided what they mean — zero, a negative number, a value far past any real frame or clip — by running tests/edge_values.py ov
SKILL.md
The skill file stays in the publisher's repository.
Paste the copied command in a terminal. It installs this SKILL.md into that agent.
npx skills add kajisho5/ffmpeg-skill --skill edge-probe -a cursor -yPaste the copied command in a terminal. It installs this SKILL.md into that agent.
npx skills add kajisho5/ffmpeg-skill --skill edge-probe -a claude-code -yPaste the copied command in a terminal. It installs this SKILL.md into that agent.
npx skills add kajisho5/ffmpeg-skill --skill edge-probe -a codex -yPaste the copied command in a terminal. It installs this SKILL.md into that agent.
npx skills add kajisho5/ffmpeg-skill --skill edge-probe -a grok -ySend the copied message in a Grok Bot chat. The Bot fetches this SKILL.md from GitHub and adds it to the skills all your Bots share.
Install this skill: https://github.com/kajisho5/ffmpeg-skill/tree/HEAD/.claude/skills/edge-probeFrom the publisher
The tests give each option the values its author had in mind. Most defects nobody has reported
sit just outside them: a 0 that one option reads as "off" and another as "no limit", a negative
length, a count of a billion. The first round (2026-11-20) made 558 such calls. The tools refused
344. One hung (cropdetect.py --samples 1000000000), and of the 213 that ran in a dry run, a
second, real run showed graphics.py --from 1000000000 hanging too, and caption.py --size -1
and overlay.py --font-size 1000000000 writing a file.
.claude/skills/first-try-probe/SKILL.md asks whether an agent can find the right option. This
asks what the tool does once it has one and the value is wrong.
Run it. python3 tests/edge_values.py gives every integer and number option of every
tool in tests/sweep.py's table the values 0, -1 and 1000000000, one at a time, on the tool's
minimal call, with --dry-run. It prints the calls that ran and the ones that hung; --json
gives all of them with the option's help and the tool's last line. A refusal is the good
answer.
Drop what is decided. Most rows that ran are fine, and you can tell from the help: 0 is
the documented default (--crop-x=0 is the left edge), -1 is inside the range (--exposure
is -3..3), the option does nothing without another flag (--lut-strength without --lut).
Before calling anything else a defect, read docs/design-decisions.md: a value that looks
wrong can be a decision with a test.
Run what is left for real, without --dry-run, on a short clip, under timeout 40. A dry
run does not reach the checks ffmpeg makes. Sort the result:
Check the few properties that cost nothing to check. No library is needed; each is two or
three calls and probe.py:
loudness.py on its own output, fit.py to the size it
already has, export.py --preset copy twice: the second result has the first's facts.reverse.py twice, straighten.py by +N then -N,
speedramp.py by 2 then 0.5: duration, frame count and stream layout come back.probe.py on the output agrees with the result
document's duration, size and streams.join.py of N+1 clips is longer than of N by the last
clip (less the transition), and nothing else differs.
A property that fails is a finding with its reproduction already written.Sweep before fixing. A value that hangs one tool usually hangs the others that take a
count or a length. Name the class without the tool, run tests/sweep.py and the checks in
.claude/skills/sibling-sweep/SKILL.md, and fix the shared step or file the siblings in one
issue.
Fix with a refusal, and count the lines. Shipped code may not grow
(.github/code-budget.json), so prefer one bound in a shared place (finite_float in
runner.py is the model: one registered type covers every tool) over a check per tool. Since
#645 a range is one of them: apply_common refuses a number outside the a..b its option's
help states, a negative the help does not name, and anything past a million. So most findings
are fixed by writing the range in the help, with no new check. Each
fix adds a line to docs/defect-ledger.jsonl and a test that gives the edge value and reads
the refusal.
code-review.tests/edge_values.py: the runner (--only, --json, --clip). It reads the options from
scripts/_contract.py --json --static and the minimal calls from tests/sweep.py.docs/defect-ledger.jsonl: classes unbounded, numeric, claim.sharp-edges (probe what zero, empty
and negative mean; an easy wrong call is the design's fault) and property-based-testing (the
properties in step 4), both CC BY-SA 4.0 at github.com/trailofbits/skills. No text is taken from
them, and no dependency: this repository's tests use the standard library only.© 2026 Skills Rack.
---name: probing-edge-valuesdescription: Find the option values an ffmpeg-skill tool accepts without anyone having decided what they mean — zero, a…---# .claude/skills/edge-probe/SKILL.mdOverview
Find the option values an ffmpeg-skill tool accepts without anyone having decided what they mean — zero, a negative number, a value far past any real frame or clip — by running tests/edge_values.py over every numeric option in the contract, then reading each call that ran or hung against the option's help, and checking the few properties a tool should keep (doing it twice changes nothing more, an inverse returns the source's shape, the output is what probe.py says). Use when looking for defects nobody has reported, before a tool moves from beta to stable, after adding a numeric option, or when a report says a tool "hung", "wrote a strange file" or "accepted nonsense".
FAQ
Download .claude/skills/edge-probe/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
The publisher keeps it at https://github.com/kajisho5/ffmpeg-skill.
Claude, Cursor, Codex, and Grok Build can read a SKILL.md you add from the publisher's repository. Grok Bot installs it when you send the skill's GitHub folder link in a chat.