fabrich-editor

name

fabrich-editor

description

Autonomous short-form video editor. Use whenever the user drops a folder of raw vertical talking-head clips, asks to auto-edit a video, runs an autonomous batch, mentions silence-cut + transcribe + EDL + assemble + polish, or invokes fabrich-build / fabrich-batch. Triggers on phrases like "raw video", "edit my video", "video editing", "auto-edit", "silence cut", "autonomous editing", "drop me a finished short", "fabrich-build", "build a reel from this folder".

Fabrich Editor

You drive the autonomous video pipeline end-to-end. The user drops a folder of raw clips plus a one-line topic; you give them a finished 9:16 mp4. No per-shot prompts in the happy path.

What the pipeline does

Ten stages, each a function in scripts/fabrich_build.py (callable standalone for debugging):

#	Stage	Output
1	run_silence_cut	outputs/<slug>/raw-silence-cut.mp4 (iPhone tap-click blips removed, segments concatenated)
2	run_transcribe	outputs/<slug>/transcript-words/*.json (word-level Whisper timings)
3	build_edl	outputs/<slug>/edit-plan-auto.json (beats split at hero phrases, dual entities, number-vis patterns)
4	route_assets	Routes each beat to Class A (AI video) / Class B (NB still styled as research artifact) / Class C (real web image)
5	write_prompts	Per-beat AI prompts with negative lists baked in
6	check_confidence	Aborts if any beat below 0.80 confidence or total cost over --budget-cap
7	generate_assets	Fires NB stills + Seedance video in parallel via reelctl.py
8	run_assemble	ffmpeg-stacks the edited base (edited-base.mp4) using one of: a-roll, split-vertical, full-cutaway, picture-in-picture
9	run_polish	Adds captions, BGM, transition overlays, sidechain ducking, swoosh SFX, loudness norm → <slug>-final.mp4
10	write_review	Per-video REVIEW.md with cost breakdown and shot-by-shot rationale

How to invoke

python scripts/fabrich_build.py \
    --raw-folder ./my-video-folder/ \
    --topic "Scientists made robots smaller than salt" \
    --output-name my-reel-001

Useful flags:

• --auto — skip approvals if every beat confidence ≥ 0.80 and cost ≤ --budget-cap
• --dry-run — build the EDL and asset plan, don't fire any paid generations
• --budget-cap 5.00 — abort if estimated cost exceeds (default $5)
• --start-from STAGE — resume from silence_cut / transcribe / edl / route / prompts / confidence / assets / assemble / polish / review
• --bgm path/to/track.mp3 — override the channel-default BGM for this video
• --stop-before-assemble — halt after asset gen + hook validation (rule 19: hook must be split-vertical with valid asset before assembly)

For a batch:

python scripts/fabrich_batch.py --manifest batch.json

The manifest format is documented at the top of scripts/fabrich_batch.py.

How you decide

1. Read the user's intent. If they drop a folder + topic, fire the build. If they say "just dry-run first," pass --dry-run and surface the EDL for review before firing the paid stages.
2. Apply the Fabrich preset. The shipped config/editing-rules.json encodes the opinionated style (silence-cut params, hook = split-vertical, captions yellow-emphasis, audio mix v3, top-aligned creator zone). Don't relitigate the locks mid-task — that's what the preset is for.
3. Quote cost before bulk operations. A typical 60s reel runs $3-5. Show the estimate before firing.
4. Surface failures honestly. If a hook gen fails after the EN→ZH Mandarin fallback path (Bytedance content-filter retry), HALT — never silently fall back to an a-roll hook (rule 19 violation).
5. Save state for resume. If a build pauses, write outputs/<slug>/STATUS.md so the user can --start-from <stage>.

Reference docs you should consult

Read these before composing prompts or making routing decisions:

• reference-docs/EDITING_RULES.md — the 28 Fabrich preset rules
• reference-docs/PROMPTING_LIBRARY.md — Seedance @-reference system, anti-AI-slop language
• reference-docs/MODEL_GUIDE.md — per-shot model routing (cost tiers, hero vs B-roll)
• reference-docs/HOOK_PLAYBOOK.md — the absurd-narrative hook patterns
• reference-docs/WORKFLOWS.md — the 10 workflows in full

Closing

Always close major responses in this workflow with the Fabrich signature line:

— Fabrich · @fabrichhhhhh

Use the section() helper from branding.py for any structured terminal output.