boring-tools

agent
Guvenlik Denetimi
Basarisiz
Health Gecti
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 11 GitHub stars
Code Basarisiz
  • rm -rf — Recursive force deletion command in example/rebuild.sh
  • rm -rf — Recursive force deletion command in scripts/assemble-spans.sh
  • eval() — Dynamic code execution via eval() in scripts/blur-background.py
  • rm -rf — Recursive force deletion command in scripts/cut-silence.sh
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Agent-driven video editing where the edit is text. No timeline, no project file, nothing uploaded. ffmpeg + whisper + python, and a SKILL.md your agent reads.

README.md

boring tools

Claude edits your video. You talk, it works.

You don't cut anything. You don't open a timeline. You hand it a recording, tell it what you want the
way you'd tell a person and watch it happen in the terminal.

I shot one take on my phone and got twelve finished ads back. Three lengths, two aspect ratios,
captions burned in, brand cards on the wall beside my head. I haven't opened an editing app since
July.

Your directions stay plain. The current public setup still uses a terminal.

What comes out

One phone take produced twelve finished files across three lengths and two aspect ratios.

The example/ folder holds the real cut list, b-roll plan, measured edges and the rebuild script from that shoot.

Do this

A. Create a project and install the skill. This is a macOS tool. Apple Silicon enables the fast encode.

mkdir my-video-project
cd my-video-project
npx -y skills add brookejlacey/boring-tools -a claude-code -y
brew install ffmpeg whisper-cpp [email protected]
python3.12 -m venv .venv
source .venv/bin/activate
pip install -r .claude/skills/boring-tools/requirements.lock
python3 .claude/skills/boring-tools/scripts/download-models.py

The last command downloads three pinned models. It verifies each file before use. The speech model is about 488MB.

If you use the python3 that ships with macOS, the install fails: it is Python 3.9 and the pinned packages need 3.11 or newer. That's why the venv is made with python3.12.

If you run npx skills update later, it overwrites the skill folder, so any change Claude made inside .claude/skills/boring-tools is lost. Keep your own edits in your project folder.

B. Tell Claude what you want.

here's my recording: ~/Movies/take.mov
cut the dead air, give me a 15, a 30 and a 60, vertical and landscape,
burn in captions and put my brand cards on the empty wall to my right

Claude reads the installed skill, does the work and shows each result.

Things worth saying to it

Real notes that produced real changes. Say them like this.

  • cut the dead air
  • the cards are too small
  • put the phone beside my head instead of over my face
  • take that white plug off the wall behind my head
  • find my product's site and build the cards from whatever branding you find
  • my eyes jump between cuts, fix it
  • prove the audio still lines up

That last one matters more than it looks. Ask it to prove its work and it will write itself a check
that fails loudly next time.

Where this runs

Claude Code in a terminal. The desktop app is the same thing with a window around it, so that works
too. The browser version won't, because your footage isn't in it.

Setup downloads the listed model files. Normal video work makes no network request.

Everything else runs on your machine. The tools do not upload your footage, face, voice or transcript.


Appendix: what's actually happening

You can stop reading at the line above and this still works. The rest is for when something breaks, or
when you want to know why it made a choice.

The one idea

The edit is text. No timeline, no project file. Decisions live in a cuts.tsv holding a start, an
end, a framing and a name per cut. That compiles to a timeline.tsv snapped to the frame grid.

That is why a note works at all. An agent can read a text file, change one number and rebuild every
output. It can do none of that to a timeline blob, which is why every other approach ends with you
back in the app dragging rectangles.

You are not meant to edit those files by hand. They exist so the agent has something it can address.

The chain

Step Tool What it does
1 vad-spans.py Finds speech off the audio envelope. Run before transcribing, or silence makes the transcriber hallucinate.
2 speech-spans.py, word-spans.py Per-span transcription with word-level timing.
3 take_cutlist.py Groups repeated attempts at the same line, flags the last as the keeper. Proposes, never cuts.
4 match-beats.py Finds a scripted beat inside a long take.
5 face-track.py Face per frame so crops follow. Detection runs on square windows across the frame.
6 eyeline.py Holds eyes on one line across cuts of different tightness.
7 subject-bounds.py Measures the subject's real edge and top per cut. Replaced a typed constant that was 286px wrong.
8 assemble-cuts.py cuts.tsv in, timeline.tsv out, renders, asserts sync.
9 broll-screens.py Screen recordings. --column puts a phone in the wall beside the speaker.
10 wall-graphics.py Brand cards, scheduled off cut labels so one plan serves every length and survives a recut.
11 caption-burn.py One word at a time. Composites the other layers in the same pass so the job costs one encode generation.
12 wall-clean.py Removes a fixed object from a wall across a whole take by borrowing clean wall from the same frame.
13 verify.py Dimensions, frame count, audio against video, drift off the grid. Non-zero exit on failure.
14 check-sync.py Matches each cut's sound and picture back to the source and measures lip sync per cut. Catches a picture that lands frames late, which verify.py cannot see.

Extras: blur-background.py (needs pip install torch, a local RVM v1.0.0 checkout
and local weights),
clip_rank.py, tighten.sh, cut-silence.sh, assemble-spans.sh.

One rebuild.sh per shoot runs the chain in order. example/ preserves the real file formats and
one rebuild recipe. The source footage and private brand assets are not public, so it is a reference
rather than a runnable demo. verify.py runs against finished renders. It exits non-zero if a check fails.

Failure modes already paid for

Every tool carries its failed approaches in its docstring with the reason each one failed. An agent
handed a clean tool will re-derive the same wrong answers, so the wrong answers live next to the right
one.

  • Resolve phone rotation before any geometry. Metadata and pixels disagree until you do.
  • Run face detection on square windows, not the full frame. A landscape frame needs left, centre and right windows, or a speaker right of centre is missed.
  • Never seek with -ss before -i on phone HEVC. It lands one to three frames late while the audio is exact, so lip sync drifts up to 100 ms and every length check still passes. Seek early, then pick the frame on the source clock.
  • Transcribe per speech span. A long silence makes a transcriber invent sentences.
  • Rasterize caption text with Pillow. Most ffmpeg builds ship with no text renderer.
  • Never measure a subject edge against one wall reference. A room falls off in brightness across the
    frame, about 15 levels in mine. A naive detector calls that gradient a person and pins the edge at
    the frame border on every cut. Compare each column against a rolling baseline of the wall beside it: a
    falloff is a drift, a person is a step.
  • You cannot recover what sits behind hair by measuring over time. Median returns hair. Median over
    unsaturated pixels returns hair. Least-saturated returns a blown-out flyaway brighter than the wall.
    Borrow the patch from flat wall in the same frame instead, so it carries that frame's own grain and does
    not crawl.
  • HSV saturation calls a dark neutral colourful, which protects an object's cast shadow and leaves it
    floating in mid-air. Use undivided chroma.
  • One-word captions strobe unless each word holds until the next arrives when the gap is under 0.28s.
  • macOS ships bash 3.2, where an empty array under set -u is unbound, so a plain "${ARR[@]}" aborts
    a rebuild on the first vertical cut. Use ${ARR[@]+"${ARR[@]}"}.

Standing decisions

Measured, not preferences.

  • Layout numbers are measured off rendered footage, never typed.
  • Landscape puts screen recordings in the column beside the speaker. Vertical keeps the full-frame
    cutaway, because there the speaker spans the whole width and there is no wall.
  • Cards scale by one factor across type, padding, radius and shadow.
  • Captions: one word, white, black outline, 0.032 of frame height (--type-frac to change it). In a corrections file, _ joins two words into one caption slot, so INTACT → IN_TECH shows "IN TECH" for the misheard word's duration.
  • Every claim on a card has to be checkable. No install counts, no ratings, no invented numbers.
  • Hardware encode by default (h264_videotoolbox), measured faster at no visible cost.

License

MIT. Take it apart.

Yorumlar (0)

Sonuc bulunamadi