agentleFS
duncan-buildroomduncan-buildroom/profit-room-skills/skills/ig-channel-remix/SKILL.md

ig-channel-remix

Skill18 starsChanged 4 months agoMITView on GitHub

Mine any Instagram channel's viral format and render new videos in that format with a creative twist. Watches the account's top 10 most-viewed reels, focusing on the first 3 seconds of each (style, camera, realistic vs. cartoonish, the specific action taking place), synthesizes the recurring mechanics into a reusable Format DNA saved to a per-channel library (so future runs can recall it or generate more variations without re-mining), then applies a twist to the premise while keeping the proven camera/pacing/action mechanics intact (e.g. a cute-animal-rescue format becomes a mythological-creature-rescue format; a roller-coaster POV becomes a fantasy-world POV) and renders finished remix videos scene-for-scene with Higgsfield Seedance 2.0 in 15-second, 9:16, 720p clips, stitching multiple clips together for longer source videos. Trigger on /ig-channel-remix, "remix this Instagram channel", "reverse engineer this account's viral format", "steal and twist @handle's format", "what's this channel's formula and can we put a spin on it", "make more variations of @handle's format", or a pasted IG handle/profile URL plus "remix this channel" / "find their format and twist it". NOT for a single video 1:1 recreation with no discovery or twist and NOT for text-only hook/caption structure analysis.

The file

Show its rules and text

We split this skill into the rules it gives an agent

15 rules. About 0 are common, 8 are distinctive, and 7 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 7

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    A video-watching tool is optional: If you have a "watch"-style skill/plugin available it can stand in for the yt-dlp + ffmpeg …

    line 43
  2. Distinctive:

    Any living subject must read as alive, not a static prop: If the twisted (or original) subject is a person, animal, or mythical/fantastical …

    line 89
  3. Distinctive:

    Every mined channel gets a persistent folder:

    line 99
  4. Very specific:

    Before Discovery, check whether library/<handle>/format-dna.md already exists.

    line 113
  5. Very specific:

    If re-mining: run the full pipeline and overwrite meta.json/reels.json/analysis.md/ format-dna.md/frames/ in the library — but only …

    line 122
  6. Distinctive:

    If no library entry exists: run the full pipeline as normal and create the library entry at the end (see Step 11).

    line 125
  7. Very specific:

    If this exits non-zero (no IG session, private account, API shape change), fall back to any IG-reels MCP tool you have configured (e.g.

    line 144
  8. Distinctive:

    Read every extracted frame plus (optionally) the reel's caption for the 0-3s idea.

    line 158
  9. Very specific:

    Check twists-log.md (if the library entry already existed) and skip twists already tried recently unless the user asks to repeat one.

    line 193
  10. Distinctive:

    If the twisted subject has no usable real-world reference (a mythological creature, an invented setting), generate hero reference stills …

    line 200
  11. Distinctive:

    If a render clearly breaks (subject warps unrecognizably, motion incoherent), retry once with a tighter prompt, then keep the best result …

    line 251
  12. Distinctive:

    Only needed for multi-window remixes (source reel >15s) — concat windows for that reel in order:

    line 255
  13. Very specific:

    Write/overwrite library/<handle>/meta.json, reels.json, analysis.md, format-dna.md, and copy frames/reel_* (opening 0-3s only, not the …

    line 270
  14. Very specific:

    Update library/index.json with this handle's last_mined/reel_count.

    line 275
  15. Very specific:

    If the account's videos aren't format-consistent (the top 10 are a grab-bag of unrelated styles), say so plainly in format-dna.md rather …

    line 309

≈ “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

IG Channel Remix — mine a channel's viral format, twist it, render new videos

Point this at one Instagram account. It finds the account's top 10 most-viewed reels, watches each one's opening 3 seconds frame-by-frame, distills what's actually working mechanically (style / camera / realism / action), saves that Format DNA to a reusable per-channel library, then helps pick a creative twist that keeps those mechanics fixed while changing the premise — and renders finished remix videos scene-for-scene with Higgsfield Seedance 2.0. The deliverable is video, not a spec document.

Requirements

  • yt-dlp + ffmpeg/ffprobe on PATH — download reels and extract frames.
  • python3 with requests (and optionally browser_cookie3) — the discovery script.
  • Instagram access — the discovery script needs a logged-in IG session. It reads your own browser session cookies via browser_cookie3, OR you can set IG_SESSIONID (your account's sessionid cookie) in the environment. Never commit a cookie/session value. If IG access isn't available, the skill falls back to any IG-reels MCP tool you have configured (e.g. a vidIQ-style ig_profile_reels tool), at reduced reel counts.
  • Higgsfield — the higgsfield CLI on PATH (higgsfield account status should succeed; run higgsfield auth login if the session expired). Video renders always use seedance_2_0.
  • A video-watching tool is optional. If you have a "watch"-style skill/plugin available it can stand in for the yt-dlp + ffmpeg frame-extraction steps below; otherwise the inline commands are fully self-contained.

Tooling

  • Discovery — scripts/fetch_top_reels.py in this skill's directory. Pulls the account's real per-post play_count via Instagram's private API using a logged-in IG session (browser cookies via browser_cookie3, or the IG_SESSIONID env var), ranks by views, returns the top 10 reels. This is the validated method — do not reimplement IG scraping ad hoc.
  • Watch (download + frames) — download each reel once with yt-dlp, then pull frames for any time window with ffmpeg:
    # download the reel once (add --cookies-from-browser chrome for private/age-gated reels)
    yt-dlp -o "frames/reel_<NN>/source.%(ext)s" "<reel url>"
    # extract frames for a [start,end] window at 2fps, scaled to 1024px wide
    ffmpeg -ss <start> -to <end> -i frames/reel_<NN>/source.mp4 \
      -vf "fps=2,scale=1024:-1" "frames/reel_<NN>/window_<MM>/frame_%03d.png"
    (If you have a dedicated video-watching skill/tool, you can use it instead of these two commands — it produces the same thing: frames for a time window, plus an optional caption transcript.)
  • Render — the higgsfield CLI. Video model is always seedance_2_0 — never substitute another video model. Use nano_banana_2 only for hero reference stills if the twist needs a new subject. Always pass --wait so the command blocks and prints the result.
  • Assemble — ffmpeg to concat multi-window renders.
  • Library — ~/.claude/skills/ig-channel-remix/library/<handle>/, this skill's persistent memory of what it found and chose for each channel (see Library section below). If the skill is installed elsewhere, adjust that path to this skill's own directory.

Fixed render spec (non-negotiable)

  • Model: seedance_2_0, always.
  • Aspect ratio: 9:16, always — regardless of the source reel's native aspect ratio.
  • Resolution: 720p, always.
  • Clip length: exactly 15 seconds per render. This is Seedance 2.0's per-call ceiling, and we render every window at that ceiling rather than at the source's raw sub-15s remainder where avoidable — each clip should read as a full, deliberate 15-second scene, not a truncated fragment.
  • Multi-shot via timestamps, not multiple renders, within one clip: each 15s clip's prompt is bracket-timecoded with as many internal shots/cuts as the source reel actually has in that window (e.g. [00:00-00:04], [00:04-00:09], [00:09-00:15]) — Seedance handles the internal cuts natively from one prompt.
  • Multiple clips only when the source exceeds 15 seconds: N = ceil(source_duration / 15) windows, each its own independent 15s Seedance render, stitched together afterward with ffmpeg. A source reel ≤15s is one clip, no stitching.
  • Any living subject must read as alive, not a static prop. If the twisted (or original) subject is a person, animal, or mythical/fantastical creature, every render prompt must explicitly call for visible autonomic motion — blinking, breathing/chest movement, natural weight shifts, head turns, eye tracking/darting toward action or camera — never a stiff, frozen, or mannequin-like pose. This applies doubly to invented subjects (griffins, kelpies, etc.) since there's no real-world footage to fall back on for that liveliness — it has to be in the prompt.

Library — recall past mining, generate more variations without re-mining

Every mined channel gets a persistent folder:

~/.claude/skills/ig-channel-remix/library/<handle>/
├── meta.json          # last_mined (ISO date), reel_count, source ("cookie" | "mcp_fallback")
├── reels.json          # the top-10 (or fewer) reels mined, with play_count/url/caption
├── analysis.md          # per-reel opening-3s breakdown
├── format-dna.md         # synthesized reusable format spec
├── frames/               # opening 0-3s frames per reel (small, persisted)
└── twists-log.md          # one entry per render run: date, twist used, video count, output path

Plus one index: ~/.claude/skills/ig-channel-remix/library/index.json — a flat map of handle → {last_mined, reel_count} for a fast existence check without globbing.

Step 0 — check the library before doing anything else

Before Discovery, check whether library/<handle>/format-dna.md already exists.

  • If it exists: tell the user when it was last mined and how many reels it covered, then AskUserQuestion: reuse the saved Format DNA (skip straight to the twist step) vs. re-mine fresh (pull the current top 10 and re-watch, e.g. if the channel has posted a lot since). Reusing is the natural path for "make more variations of @handle" requests.
  • If reusing: load reels.json/analysis.md/format-dna.md/frames/ straight from the library, skip Discovery/Watch/Analysis/Synthesis entirely, and jump to the twist step. Read twists-log.md first and avoid proposing a twist already tried in a recent entry unless the user explicitly asks to repeat/iterate on one.
  • If re-mining: run the full pipeline and overwrite meta.json/reels.json/analysis.md/ format-dna.md/frames/ in the library — but only append to twists-log.md, never truncate it; past render history stays recoverable.
  • If no library entry exists: run the full pipeline as normal and create the library entry at the end (see Step 11).

Pipeline

1. Preflight

Confirm ffmpeg, ffprobe, yt-dlp, python3, and higgsfield are on $PATH (higgsfield account status should succeed — if session expired, ask the user to run higgsfield auth login). Slugify the handle and create the output tree:

~/Desktop/ig-channel-remix/<handle>/{frames/, refs/, clips/}

2. Discovery — top 10 by real views

Skip this step entirely if Step 0 chose to reuse the library.

python3 ~/.claude/skills/ig-channel-remix/scripts/fetch_top_reels.py --handle "<handle>" \
  --top 10 --out reels.json

If this exits non-zero (no IG session, private account, API shape change), fall back to any IG-reels MCP tool you have configured (e.g. a vidIQ-style ig_profile_reels, typically capped around 6 reels, popularity-within-recent-activity, no pagination). Note the fallback and the reduced count in the final report and in meta.json's source field — do not block the run over it.

3. Watch each reel's opening 3 seconds

Skip if reusing the library. Per reel in reels.json, download it once and pull the first 3s of frames:

yt-dlp -o "frames/reel_<NN>/source.%(ext)s" "<reel url>"   # add --cookies-from-browser chrome if needed
ffmpeg -ss 0 -to 3 -i frames/reel_<NN>/source.mp4 -vf "fps=3,scale=1024:-1" \
  "frames/reel_<NN>/frame_%03d.png"

Read every extracted frame plus (optionally) the reel's caption for the 0-3s idea. If a reel fails to download (deleted, private), skip it and note it in the final report — don't let one bad reel block the run.

4. Per-reel analysis → analysis.md

Skip if reusing the library. Grounded strictly in what's visible in the frames — never invent. For each reel record:

  • Style — POV selfie, static tripod talking-head, screen recording, b-roll montage, whiteboard/hand demo, meme-text-over-clip, etc.
  • Camera — phone front/back cam, DSLR/mirrorless, screen capture, drone, action cam, or AI-generated (no real camera involved).
  • Realism — live-action photoreal vs. animated/cartoon/illustrated/AI-stylized.
  • Action — one concrete line: what physically happens in frame 0–3s.
  • Secondary cues: aspect ratio, on-screen text/hook overlay present at 0s, cut count within the first 3s, first spoken/caption line.

5. Synthesize → format-dna.md

Skip if reusing the library. Cross-reel pattern synthesis with frequency counts, e.g.:

  • "8/10: static front-facing phone camera, no tripod visible"
  • "10/10: live-action, zero cartoon/animated openers"
  • "7/10: subject already mid-action at frame 0, no wind-up"

This is the mechanical spec that must survive the twist untouched: camera/framing, opening- beat structure, realism level, pacing. Write it as instructions, not just observations — this is what Step 9's render prompts pull from.

6. Twist concepts — check-in

Generate 3 twist directions that hold every format-dna.md mechanic fixed but swap the subject/premise into something novel. The move is: keep the camera, framing, pacing, and realism level identical; change what/who is on screen and what they're doing thematically. Examples: cute-animal-rescue → mythological-creature-rescue (same handheld rescue-footage camera, same rescue action, same live-action realism, subject is now a griffin/kelpie/etc.); roller-coaster POV → fantasy-world POV (same GoPro-style mounted camera and motion, same photoreal-vs-stylized call as the source, setting becomes a fantasy realm).

Check twists-log.md (if the library entry already existed) and skip twists already tried recently unless the user asks to repeat one. Use AskUserQuestion to present the 3 concepts plus a "my own idea" option, and confirm how many remix videos to render (default 3). This is the one required check-in — the twist materially changes the output and render credits are real money.

7. Reference stills (only if needed)

If the twisted subject has no usable real-world reference (a mythological creature, an invented setting), generate hero reference stills first:

higgsfield generate create nano_banana_2 --prompt "<twisted subject, consistent with
format-dna's realism level>" --aspect_ratio 9:16 --resolution 1k --wait --json

Save to refs/. If the twist keeps a real-world subject (e.g. fantasy styling of a real POV format), skip this — lean on the source frames' camera/motion plus prompt text only.

8. Full-video beat extraction for the reels being rendered → beats.md

The opening-3s frames from Step 3 are enough to define the Format DNA, but reproducing a reel scene-for-scene needs the whole video's beat structure, not just its opening. For only the source reels selected for rendering (the confirmed count from Step 6 — not all 10):

  1. Probe duration: ffprobe -v error -show_entries format=duration -of default=nk=1:nw=1 frames/reel_<NN>/source.mp4.
  2. Compute windows: N = ceil(duration / 15), windows [0,15), [15,30), ..., last window shorter only if duration isn't a multiple of 15.
  3. Dense frames per window:
    ffmpeg -ss <window_start> -to <window_end> -i frames/reel_<NN>/source.mp4 \
      -vf "fps=2,scale=1024:-1" "frames/reel_<NN>/window_<MM>/frame_%03d.png"
  4. Per window, write sub-timestamped beats grounded in the actual frames — character action, camera movement (static/pan/tilt/zoom/dolly/handheld/cut) and direction, cut frequency, on-screen text/style cues. This is the shot list the render prompt in Step 9 rewrites with the twisted subject — the shot count and timing stay the source's, only the subject/action content changes per the twist.

9. Prompt + render — Seedance 2.0, one 15s clip per window

For each window from Step 8, rewrite its beats with the twisted subject/action but preserve the shot count, timing, and camera moves exactly:

higgsfield generate create seedance_2_0 \
  --prompt "[00:00-00:04] <camera move from source beat> — <twisted action>
[00:04-00:09] <camera move from source beat> — <twisted action>
[00:09-00:15] <camera move from source beat> — <twisted action>
<realism/camera/pacing line pulled straight from format-dna.md, unchanged>
<if subject is living: natural blinking, breathing, weight shifts, and head/eye movement —
never stiff or frozen>
Recreate the camera work, shot timing, and pacing exactly; only the subject and setting are
different." \
  --start-image <original frame or generated ref> --image <...> \
  --duration 15 --aspect_ratio 9:16 --resolution 720p --wait --json

Cap references at ~9 per window (Seedance's documented limit) — pick the most representative (start, end, evenly spaced between). Bind renders by source-reel/window index, never completion order — renders finish out of order. Download to clips/remix_<NN>_window_<MM>.mp4 (or clips/remix_<NN>.mp4 directly if the reel is a single window).

If a render clearly breaks (subject warps unrecognizably, motion incoherent), retry once with a tighter prompt, then keep the best result and flag it — don't loop indefinitely.

10. Assemble

Only needed for multi-window remixes (source reel >15s) — concat windows for that reel in order:

ffmpeg -f concat -safe 0 -i <(for f in remix_<NN>_window_*.mp4; do echo "file '$PWD/$f'"; done) \
  -c copy clips/remix_<NN>.mp4

Re-encode (-c:v libx264 -c:a aac) if -c copy fails on codec mismatch. Single-window remixes need no concat step — the window's render is clips/remix_<NN>.mp4.

11. Report + update library → report.md

Summarize: how many reels were sampled/ranked, any fallback used (cookie method failed → MCP cap), any reel skipped, the chosen twist, and a link to every finished clips/remix_*.mp4. Print the final output tree.

Then update the library:

  • Write/overwrite library/<handle>/meta.json, reels.json, analysis.md, format-dna.md, and copy frames/reel_* (opening 0-3s only, not the full-video Step 8 frames) into library/<handle>/frames/.
  • Append one entry to library/<handle>/twists-log.md: date, twist chosen, number of clips rendered, path to this run's ~/Desktop/ig-channel-remix/<handle>/ output.
  • Update library/index.json with this handle's last_mined/reel_count.

Output folder (per run)

~/Desktop/ig-channel-remix/<handle>/
├── reels.json           # top 10 (or fewer, on fallback): shortcode, play_count, caption, url
├── frames/
│   └── reel_01/, reel_02/, ...              # opening 0-3s frames (Step 3, all 10 reels)
│       └── window_01/, window_02/, ...       # full-video dense frames (Step 8, rendered reels only)
├── analysis.md           # per-reel opening-3s: style / camera / realism / action / cues
├── beats.md              # per-window scene-for-scene beats for reels being rendered
├── format-dna.md          # synthesized reusable format spec (mechanics the twist must preserve)
├── twist-concepts.md       # the 3 generated options + which one was chosen
├── refs/                  # generated hero reference stills, if the twist needed a new subject
├── clips/
│   └── remix_01.mp4, remix_02.mp4, ...   # finished 15s (or stitched multi-window) remix videos
└── report.md

Rules

  • Never invent action, camera, or style not visible in the extracted frames when writing analysis.md/format-dna.md/beats.md. Fidelity to the source format is what makes the twist land — only the subject/premise should change, never the underlying mechanics.
  • The twist check-in is required, not optional — it's the one point where user judgment materially changes the output and where render spend begins.
  • seedance_2_0 / 9:16 / 720p / 15-second clips are fixed, regardless of source aspect ratio, duration, or resolution. Stitch with ffmpeg for sources >15s; never shrink the clip length to match a shorter source instead.
  • Bind renders by index, never completion order.
  • One retry per broken render, then move on and flag it.
  • Always save/update the library at the end of a run (Step 11) so a later "make more variations" request can skip straight to the twist step.
  • If the account's videos aren't format-consistent (the top 10 are a grab-bag of unrelated styles), say so plainly in format-dna.md rather than forcing a false pattern — a twist needs a real mechanic to preserve.

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.