Source: raw/x-account-trq212-2064826394589442448.md (@trq212’s 6:39 explainer video + thread, 2026-06-10) + ai-research/thariq-cc-video-editing-deck-2026-06-10.md (companion deck, thariqs.github.io/cc-video-editing-deck) + raw/x-account-claudedevs-2064399512664526853.md (the launch video itself, @ClaudeDevs, 2026-06-09) 2026-10-27 addition: raw/Claude_Code_for_creators_-_AI_editing_is_here.md (How I AI with Brooke Lacey, youtube.com/watch?v=UF08V1XIUmw) and ai-research/be-more-ambitious-altitude-depth-range-2026-10-27.md.

The Claude Fable 5 launch video was edited by Fable 5 itself — Thariq (Claude Code team, Anthropic) never opened a video editor. Instead the edit lived as a repo: Whisper word-level transcripts, a JSON edit decision list executed by ffmpeg, seven hand-written color LUTs, eleven designer PNGs rebuilt as Remotion React components, two code→Figma→code round trips via the Figma MCP, and a headless 4K render the model reviewed still-by-still before each pass. The day after launch Thariq published a 6:39 explainer video (273K+ views in hours) and a slide deck — itself made with Claude from the video transcripts in about an hour.

Key Takeaways

  • “The edit is text.” No timeline, no project file. Cut points and overlay cues are grepped out of word-level transcripts, the edit decision list is a diffable final-edit.json with a written rationale per pick, and the grade is plain-text .cube LUTs applied by ffmpeg at encode time — so Claude can read, diff, and re-render the entire edit.
  • The whole pipeline is commodity tools driven by code + tool calls: Whisper (run locally on an M4 Max) → final-edit.json → ffmpeg frame-accurate cuts → hand-written S-Log3→Rec.709 LUTs → Remotion graphics → Figma MCP → npx remotion render (3840×2160 at exactly 24 fps, 4,334 frames).
  • Verification is built into every stage — the launch video’s own thesis (“we used to verify that Claude did the work right; now we verify that it’s doing the right work”) applied to video: Claude re-transcribes its own cut to confirm “zero ums,” and screenshots render stills to review its own work before each full pass.
  • By the numbers: 17 camera takes / 25 GB raw → verified 2:50 first cut within minutes → final 3:00 at 653 MB 4K. 7 LUTs written from scratch, 11 graphics components, ~10 re-renders in one night, 4 days total (Jun 6–9), 0 video editors opened.
  • /goal did the driving: the prompt that produced the first cut ended “/goal dont stop until you have a final video” — the goals-plus-verification-tools pattern the launch video introduces as new Claude Code features (/goal + workflows).
  • Domain expertise still catches what the model misses: a professional colorist flagged that the S-Log3 footage wasn’t properly color-managed (wrong gamma curve, muddied colors). Thariq conceded the tight timeline and said more specific prompting / domain workflows would likely improve it.

The pipeline

  1. Listen to every take. Whisper, run locally, produces word-level timestamped JSON for all 17 takes (transcripts/*.json). Every cut point and overlay cue downstream is grepped from these — “never scrubbed.”
  2. The edit is a JSON file. Claude writes final-edit.json — candidate takes, in/out points, and a written rationale for every pick (e.g. “C003 is the cleanest complete take, clean ending”; avoid mid-sentence cuts and ums; best of 3 takes per line). Take selections: C003 (intro, from 5 takes + a re-shoot), C010 (thought-partner section, from 6), C012 (goals-and-verification section, from 3).
  3. ffmpeg stitches the cut. One frame-accurate ffmpeg -ss … -to … cut per pick, then a concat join. 7 minutes of raw footage → a watchable 2:50 within minutes. Claude re-transcribes its own output to verify the cut (“zero ums”).
  4. Write the color grade from scratch. Prompt: “the color grading feels a bit too muted, can you fix that? make some examples of how we might regrade and let me choose one.” Seven hand-written .cube LUTs (S-Log3 → Rec.709); the human picks the look (neutral_cool_desaturated); ffmpeg applies it at encode time.
  5. Graphics as components. Eleven static designer PNGs (Figma) rebuilt as React components in Remotion (src/cards, src/overlays) “so every word and color is a prop.” First pass prompted with no animation spec or timings at all. Six timing knobs set the feel; a cue sheet lands each overlay on the spoken word from the transcript (e.g. a beat keyed to frame 295, syncing to the word “right”).
  6. Round-trip the design through Figma MCP. Two code→Figma→code round trips; the design team’s minimal redesign pass was imported back from Figma and the video updated to match.
  7. Headless render + self-review. npx remotion render — 4K, 4,334 frames, exactly 24 fps. Claude screenshots stills to review its own work before each full pass; ~10 re-renders in one night, finished “at 6:24 am.”

Closing card of the launch video: “Edited by Claude Code. Footage by a camera. Taste by Thariq. Everything in between, by the agent the video is about.”

The launch video it produced

The artifact itself (@ClaudeDevs, 2026-06-09 — ~895K views) is a 3-minute talking-head piece where Thariq lays out how Fable 5 changed the Claude Code team’s day-to-day, with the before/after framing “Is Claude doing the work right?” → “Is Claude doing the right work?” and three changes:

  1. Treat Claude like a thought partner — full context early; have it interview you before finalizing a spec; generate multiple directions + HTML mockups.
  2. Give Claude goals and ways to verify them — the /goal command (keep working until done) and workflows (parallelize, verify, report differences).
  3. Be more ambitious — “If there’s something you assumed LLMs couldn’t do, give it a chance.” The on-camera example: “I’m editing this video with Fable.”

The meta-loop is the point: the launch video demonstrating the model’s ambition claim was itself produced by the workflow the claim describes.

October 2026 update: the same pattern for short-form (Brooke Lacey)

On How I AI (Claire Vo, raw/Claude_Code_for_creators_-_AI_editing_is_here.md), Brooke Lacey, a creator with over 200,000 TikTok followers, demonstrated “the edit is text” on a short-form phone video. Everything ran locally in Claude Code from the terminal. In her words, “it just used the boring tools that are already on our computers. FFmpeg, Whisper, Python” and “it didn’t upload the video anywhere”.

  • The chain, as Claude narrated it. One 5-minute phone take became a 15-second vertical ad:
    • silence detection marks where she is talking, and Whisper word timings sharpen each cut edge, “because neither alone can place a cut cleanly”;
    • per-frame face tracking keeps the crop on her, and each cut gets a shot size (wide, mid, close);
    • eight cuts come from a TSV cut list where each row is a start, an end, a shot size and a label: “That file is the whole edit”;
    • Claude took app screenshots from the simulator and placed them as cards on the empty wall beside her, measured per cut from the rendered footage;
    • captions appear one word at a time (57 words), each held until the next word arrives; she did not ask for that rule;
    • a sync check reported “441 frames at 30 FPS, audio and video 0 milliseconds apart”.
  • One text edit, many outputs. Because the edit is a text file, “the same command makes the 30 and 60 and the landscape versions from the same tape”.
  • It follows spoken commands and picks takes. It heard her say “cut that out” in the footage and acted on it. Where she repeated a line, it “took the one that was the best cadence”.
  • Learning style from your own corpus. TikTok blocks scraping and downloads, so she downloaded her own videos on her phone and put them in Google Drive. Claude then read each video’s comments on TikTok through Claude in Chrome while she was logged in, paired the comments with the downloaded video, and proposed new content ideas and scripts. She also coached it explicitly on hook cadence: “did you hear the cadence in my voice? Because that is how a hook works for me.”
  • She already knew short-form editing. “I was able to orchestrate Claude the way that I would actually do it.” Her view is that you still need to learn the basics of video editing to direct it. She says Claude has not yet learned her two-person comedy format.
  • Recovery habit. She keeps a “lane operating system” of status MD files and memory per product. When Claude gets something wrong, she tells it to go back to the instructions, find where it misread, then “tell the main operating system spine that you messed up so that it can make the change globally”.

Related primary source on “be more ambitious”. Anthropic’s Felix Rieseberg has since published a structured essay on the same theme; see How to Be More Ambitious: Altitude, Depth, Range.

Limits and caveats

  • The color grade drew a professional correction — colorist @itsohqay showed the S-Log3 footage wasn’t properly color-managed (wrong gamma curve). The workflow produced a crew-approved grade, not a colorist-approved one. Domain review remains the backstop.
  • Not push-button: 4 days, ~10 re-renders, and explicit human taste decisions (choosing the grade, approving picks, the Figma design pass). Thariq’s framing splits the credit precisely: footage by a camera, taste by Thariq, everything in between by the agent.
  • Previous Claude launch videos were mostly made by the actual video production team (limited Claude use) — this is the exception, not the established Anthropic pipeline.
  • Nothing is open-sourced yet. Thariq is unsure whether the tools will be released and suggests the prompts/scaffolding may be more useful than the code.

Try It

  • Steal the pattern, not the tools: (1) transcribe everything to word-level JSON first; (2) make the edit a reviewable text artifact (an EDL with rationales) instead of timeline state; (3) give the agent a verification loop — re-transcribe the cut, screenshot the render; (4) drive with a goal (“don’t stop until you have a final video”).
  • The closest packaged equivalent in this wiki is video-use — transcript-driven cuts, ffmpeg, Remotion/Manim/PIL animation subagents, self-evaluating render loop — if you want this workflow without building the scaffolding yourself.
  • Read the deck (15 slides, ~3 min): https://thariqs.github.io/cc-video-editing-deck/ — archived in full at ai-research/thariq-cc-video-editing-deck-2026-06-10.md.
  • For brand-graphics work, the “designer PNGs rebuilt as Remotion components so every word and color is a prop” move is the transferable trick — it turns static brand assets into parameterized, agent-editable motion graphics.

Open Questions

  • The launch video labels /goal and workflows as new Claude Code features — whether “workflows” here is the Dynamic Workflows research preview reaching GA is unconfirmed (same open thread as the Tokyo keynote claim).
  • Will the prompts/scaffolding be published? Thariq says discussion is ongoing, no commitment.
  • Token/compute cost of the 4-day edit was asked in the thread but not answered with numbers.