agentleFS
kanishk-gangwarkanishk-gangwar/filmcraft-mcp/skills/recap-edit/SKILL.md

recap-edit

Skill4 starsChanged 10 days agoMITView on GitHub

Edit a folder of clips (and photos) into a beat-synced montage — a year recap, travel dump, trip highlight, "core memories" reel — matching a reference video's cuts, music, transitions, intro and colour, or cut to the beat of a song the user gives, built in FilmCraft (live in the app if wanted). Use this whenever the user shares a reference reel/TikTok/CapCut template or says "edit my videos like this", "make a recap/montage/highlight", "same style as this video", or hands over a folder of trip footage to cut to music ("cut my clips to this song") — even if they don't say "recap".

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Reference video: ask for the downloaded file.

    line 17
  2. Very specific:

    Write a plan JSON (references/plan-format.md), one shot per slot between the reference's cuts:

    line 54
  3. Very specific:

    Run python3 ${CLAUDE_SKILL_DIR}/scripts/plan_check.py plan.json --clips <folder>: it flags unused clips, overlapping repeats and in-points …

    line 62

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Recap edit: footage + reference → finished montage

The user judges the result against the reference, frame by frame, and against their own footage ("where is my waterfall shot?"). So the job is: copy the reference's structure exactly (cut timing, transitions, intro, colour, framing), fill it with the user's best moments, use every clip they gave you, and check the render before calling it done. Each step below has a skill with scripts; set FILMCRAFT_CLI="${user_config.filmcraft_cli}" (or the path to filmcraft-cli) for them.

1. Gather and confirm inputs

  • Footage folder and any photos folder (photo-title intros, stills).
  • Reference video: ask for the downloaded file. A web link to Instagram/TikTok is only good for a first look; frame-accurate timing needs the file. The reference's audio becomes the soundtrack when the user wants "the same music" (mute the clips: audio: false).
  • Song but no reference? Fine: cut to the music itself (see "No reference video" below).
  • Frame shape: match the reference (9:16 reel) unless the user asks for landscape. If the reference shows 16:9 footage letterboxed in a 9:16 frame, that is a style choice to copy too.

2. Analyse the reference → analyze-reference

Extract frames, detect cuts, flashes and the beat, and look at contact sheets of the intro, every transition and any section that's not a plain cut (stacked pictures popping in on the beat, split screens). Write down: cut frames, transition types at each, intro construction, caption style, ending, and the colour (shadows/highlights balance). Keep the reference's exact cut frames as your plan's cuts — it is what makes the edit "hit" the music.

No reference video, just a song

Find the beat by listening, and build the cut grid from it:

python3 ${CLAUDE_SKILL_DIR}/../analyze-reference/scripts/detect_beats.py song.mp3 --fps 30 --cuts 24 --every 2 --start 0

--cuts N prints the plan's cuts for N shots, one every --every beats (2 at ~120-160 BPM, 1 at ~70-100 BPM gives the usual 0.6-1 s shots), from the first beat on or after --start (the intro's last frame). Shape it like a reel would: longer holds (4-8 beats) in quiet sections and on hero shots, every beat or half-beat bursts in loud ones, a cut on each drops frame, and start the first shot on a downbeat. Put the song in the plan's audio. No reference means no colour or intro to copy: ask what feel they want (warm, moody, clean) or keep the grade light. If the script exits with code 2, the file has no clear beat: say so rather than cutting to noise.

3. Review all the footage → select-shots

Sample every clip, score it, look at the review sheets. For each clip note its best moments and, for portrait clips, the focus height that keeps the subject (heads!) in a landscape slice. Find hero shots for the long holds (cinematic, steady, golden light, a gesture); quick beats can take busier moments.

4. Plan the cut

Write a plan JSON (references/plan-format.md), one shot per slot between the reference's cuts:

  • Every clip appears. Users notice missing clips immediately. With more slots than clips, repeat only long clips at different moments; with fewer, use picture-in-picture/stacked sections.
  • Story order (usually chronological) with the strongest shots on the holds and the climax.
  • Clips that can't fill the frame well (selfies, tight close-ups, low-res portrait) go in as whole pictures in an overlay section instead of being zoomed full-frame.
  • Per-shot look by light (haze, golden, green, stage, night, indoor, sport) and denoise only on dark, noisy shots.
  • Run python3 ${CLAUDE_SKILL_DIR}/scripts/plan_check.py plan.json --clips <folder>: it flags unused clips, overlapping repeats and in-points past a clip's end.

5. Build the parts the reference needs

  • Photo-letter or footage-in-letters intro → photo-title (trace the reference's letters for an exact match). Bake the frames (import_media bake=true).
  • Light-leak on a big cut → python3 ${CLAUDE_SKILL_DIR}/scripts/light_leak.py <dir>, bake, place 8 frames before the cut, blend Screen (light_leaks in the plan).
  • Cream/white flashes, stacked pictures popping in on the beat → overlays in the plan.

6. Build it

python3 ${CLAUDE_SKILL_DIR}/scripts/build_from_plan.py plan.json --live 7878 --launch --pause 0.08 builds it live in a new FilmCraft window (the user watches it assemble); without --live it writes the project headless. Then connect_app(port: 7878, token: <printed token>) to keep editing with the MCP tools.

7. Grade → match-grade

Measure the reference vs your render and tune the overall Lumetri on the adjustment layer (and per-shot looks) until shadows, highlights, brightness and saturation line up.

8. Render and check → check-render

render(maxRenderQuality: true, bitrateMbps: 30), then check_render.py with the intro length and expected duration. Look at frames from every section. Report what changed, what you verified and what you couldn't (audio by ear).

When the user gives feedback

They will point at timestamps ("at 0:15 the pictures should stack", "heads are cut", "too zoomed", "looks grainy", "the font is different"). Re-analyse that exact span of the reference frame by frame before changing anything, fix it in the plan, rebuild, and re-check — don't patch blind.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.