Source: raw/Fable_5.2_Is_INCREDIBLE_at_Video_Editing_WILD_RESULTS.md
Creator: Brock Mesarich (channel “Brock Mesarich | AI for Non Techies”, confirmed from YouTube metadata with yt-dlp on 2026-11-13; the transcript itself gives no name) · URL: https://www.youtube.com/watch?v=1fBhLGhQQ2M · Duration: not recorded · Platform: YouTube · Fetched: 2026-11-13
A creator who has had Claude edit all his shorts and long-form YouTube videos for about six weeks walks through the system: two slash commands in one Claude Code project folder, a playbook file that gains a rule every time he gives a note, a browser review page for timestamped comments, and the Higgsfield MCP for background, outfit and B-roll generation. He ran it on Opus 5.5 and has now moved editing to Fable 5.2 because Opus made cutting mistakes. He says it is not one-shot: the system took six weeks of notes to tune.
Key Takeaways
- Edit as code, no editing app. Everything runs in Claude Code from one project folder; Claude “writing the video as code and then turning that code into a video file.” He does not name the rendering library.
- Two commands:
/shortfor vertical videos and/editfor long-form. Each is a plain-English skill file: purpose at the top, “read my playbook” first, ordered steps (voice, facts, animations, cover, finish), and a “never do” list at the bottom where every item came from an earlier note. - The playbook grows from notes. The first version was “basically empty”: he gave Claude one video whose editing he liked and asked it to match the style. Each later note became a rule, often with his original note written next to it. His advice: you do not need his playbook, “just start giving notes and have Claude save them every single time.” The same pattern as Ajené Pinnock’s “save this to the skill”.
- Approve the cut before any visuals. Claude first returns a plain cut with retakes and mistakes removed. Building animations on a bad cut means redoing them, so he recommends this order to anyone copying the system.
- Fact-check while collecting assets. After the cut, Claude visits the live page for whatever he is talking about to check his claims and capture real pages, screens and logos for the edit.
- Long-form adds two checkpoints: (1) the cut, delivered with a 1-minute reel of only the join points so he can hear every splice; (2) for a new look, three or four short style samples before it builds the full 15 minutes.
- Per-subject styles. Claude builds the look from the subject’s website (a Claude video looks like Anthropic’s brand: cream background, pixel characters, Anthropic orange). Styles are saved in the project folder and picked by subject without being asked. For the Fable 5.2 video it built a new style from the release page.
- Review page with pinned comments. The final 4K file comes with a page showing the timeline as clickable section blocks. He pauses, comments, and the comment is saved to that second; quick buttons cover his most common notes (“cut this word”, “wrong visual”, “hold this longer”). Claude reads every comment and returns the next version.
- Time: a 15-minute long-form video takes about 45 minutes to edit.
- Why Fable 5.2: on Opus 5.5 he would get “a word was chopped off or a bad take was left inside of the video.” Fable is “about double the price for the same amount of work on paper,” but in his testing on similar editing skills it cost about 14% more than Opus 5.5.
The /short pipeline, in his order
- Type
/shortand drop in the raw recording. Claude writes its own checklist: read playbook, cut takes, check facts, build scenes, finish. - Claude listens to the whole recording and finds every flubbed or restarted line.
- It returns the cut with no visuals. He approves it.
- It checks facts on the live pages and grabs real screenshots and logos.
- It builds the visuals: an animated hook, an animation for each explanatory line, and real pages with the exact words highlighted as he says them.
- It adds captions, music and a cover image, cleans up the audio, and exports.
- It schedules the post on Instagram, TikTok and X. (He does not say which scheduling tool or connector does this.)
Higgsfield MCP: background, outfit and B-roll
He uses the Higgsfield MCP from inside the editing session so he never opens the Higgsfield site.
- Setup (about a minute, once): on higgsfield.ai click MCP, choose the Claude Code tab, copy the message, and paste it into Claude Code as a new message. Claude installs it and opens a browser window to sign in to your Higgsfield account.
- Background swap prompt he used: “change only the background to a bright modern studio and keep me, my face, and the movement exactly the same.” Claude uploads the clip and picks the model; his face and hand movement came back unchanged.
- Outfit swap: “change only my sweater to a black blazer over a white t-shirt and keep everything else the same.” He had it make 11 versions of one clip.
- His numbers: the swaps used Kling inside Higgsfield, 12 credits each, 4 to 8 minutes per clip. Generations draw on your Higgsfield plan’s credits.
- Limit: a few seconds per clip, so it works for hooks, shorts and B-roll of himself, not a full 15-minute video, unless you cut it up and stitch the pieces.
- B-roll for lines with no footage: paste a sentence and say “make me a 5-second shot for this line.” Claude writes the shot description from that sentence plus the context of the video it is editing, sends it to Higgsfield, and the clip returns in a couple of minutes. He only watches and decides whether to use it, and says he still prefers real footage when he can get it.
- In a short, he can ask mid-edit “put me in a different studio for the hook” and Claude handles it as a step.
Try It
- Start the playbook empty. Give Claude one video whose editing you like and ask it to match the style, then save every note you give as a rule in a playbook file the skill reads first.
- Split cut from build. Have your editing skill stop after the cut and wait for approval before it builds any graphics.
- Ask for a join-points reel on long videos: a short clip of only the splices, so you can check them in real time.
- Add quick-note buttons to whatever review page you use for your three most common notes.
- Test one Higgsfield swap on a 5-second clip before using it in an edit; check the credit cost on your plan first.
- Price the model switch on your own skill. Run the same edit on Opus 5.5 and Fable 5.2 and compare cost and the number of cutting errors, as he did.
Open Questions
- Creator identity is inferred, not stated in the transcript.
- The render stack. “Writing the video as code” fits tools like HyperFrames or Remotion, but he names neither.
- The 14% figure is his own measurement on “similar editing skills”; in his launch-week bake-off the per-video edit cost 5.78 on Opus (about 69% more for that task), and Opus won that test because Fable left a repeated “and”. The 14% matches his three-task total, not the editing task.
- Kling credit cost and model choice inside Higgsfield are his figures for his plan; not checked against Higgsfield pricing.
- How the posts get scheduled to Instagram, TikTok and X is not shown.
Related
- Ajené Pinnock’s Claude Code editing setup — the same note-to-skill loop and timestamped review app.
- Matt Wolfe’s short-form editor — another creator’s Claude editing pipeline.
- How Fable 5 Edited Its Own Launch Video — Anthropic’s edit-as-text version.
- Higgsfield MCP — the connector used for swaps and B-roll.
- Fable 5.2 vs Opus 5.5 head-to-head tests — his bake-off, where Opus won the editing task.
- Claude Fable 5.2 + Mythos 5.2 — pricing and routing.
- Grok Bot — the subject of the “Grok Bot style” edits.