edit

edit — THE master editing skill

📦 PLUGIN PATHS — read once. This skill ships inside the vibe-editing plugin. Every command below references ${CLAUDE_PLUGIN_ROOT} = this plugin's install directory (the folder that contains skills/, lib/_shared/, and vault/). Claude Code sets that variable for plugin contexts; if a shell you run doesn't have it, resolve it once at the start of the job — export CLAUDE_PLUGIN_ROOT="$(cd "$(dirname "$(command -v ... )")" ...)" or simply set it to the plugin dir you installed. You usually don't need to: every bundled Python script self-locates its own lib/_shared, sibling skills, and vault/ by walking up to the .claude-plugin/ marker (it also honors VIBE_PIPELINE_ROOT), so only the entry script path in each command has to be right. User-supplied assets (not shipped with the plugin): the licensed music library → put it at ${CLAUDE_PLUGIN_ROOT}/vault/content-skill-system/(1) Tik Tok/ or point VIBE_MUSIC at it; the ~11GB overlay/transition asset pack → ${CLAUDE_PLUGIN_ROOT}/assets/ or point VIBE_ASSETS at it. The core cut→reframe→caption→render→audit path needs neither. See the plugin's INSTALL.md.

Keys (optional — the kit runs key-free): transcription defaults to local whisper (no key, via long-form-ingest/scripts/transcribe_local.py). Set GROQ_API_KEY in ${CLAUDE_PLUGIN_ROOT}/config/keys.env ONLY for the faster cloud path; set ANTHROPIC_API_KEY for richest caption styling (else it falls back to the claude CLI). No key ships with the kit. Ingest (Step 0): /edit accepts a local file, a YouTube URL (fetched with yt-dlp), or a Google Drive link (via the footage-fetch skill) — a bare /edit <url> downloads first, then runs.

🛑 POST-MORTEM FROM Tier1 Q&A BATCH 2026-06-14 → 2026-06-16 — READ FIRST IF EDITING Q&A

8 rounds of revisions on an 18-clip batch before Operator was satisfied. ALL the defects were preventable with stricter pre-/post-build audit. Root causes + what to do every Q&A run, going forward:

Why the mistakes happened

1. I treated playbook non-negotiables as advisory — silently fell back to "simplest valid floor" (single-cam Speaker) instead of doing split-screen. The playbook EXPLICITLY listed split-screen as non-negotiable #2. FIX: non_negotiables_check.py runs as a HARD GATE. Defensive pre-justification in the delivery README ("happy to do X as a V2 pass") is the smell — if I'm writing it before the user complains, I'm cutting a corner.
2. I configured visual params (zoom, eye) once GLOBALLY and shipped without per-clip visual verification. Guests differ — Guest (short) needs zoom=1.15/eye=0.18; Guest (tall) hits the reframer's source-bound limit at the same settings; Guest (medium) worked at 0.95/0.10. There is NO single zoom that satisfies all guests. FIX: Step 9 now requires per-clip OPEN-FRAME extraction + visual face-position measurement BEFORE staging.
3. I shipped music tracks with slow build-up intros without profiling. øfdream 27s, øneheart 17s, Gods creation 25s, Labrinth 4s — all bled boring intro under the hook. FIX: STEP 7 of qa_playbook.md mandates RMS profile of every picked track; if beat-drop > 1s, --music-ss <N> is required. Audit gate G5 automates this.
4. I cut segments on Whisper's word.end labels without padding for soft consonants. "road." (-d) and "one." (-n) got mid-syllable clipped because Whisper's labels are 0.1–0.25s early on soft tails. FIX: G1 boundary-tail audit catches every flagged segment-end pre-build.
5. I cut cameras mid-completion. Cut to Speaker while Guest was still saying "yeah trying to figure out…" — the source-time jump reads as "muted guest mic." FIX: G2 guest-completion audit catches same-conversation cam-cuts that skip prev-speaker words.
6. I ended clips on mid-thought, not the natural button. Guest closed on "biggest, baddest one." instead of the punchier "So give the sickest one away." 1s later. FIX: G3 payoff-extension audit reads the 3s past EDL's last word for an imperative/connector-led button.
7. I optimized one clip at a time instead of running a batch-wide audit before each rebuild. Every revision round revealed defects I should have caught proactively. FIX: qa_prebuild_audit.py runs all 6 gates in <5s on the EDL before any 4K encode burns 5–8 min. Cheaper to fail at audit than at human review.

The Q&A discipline going forward (autonomous-safe)

Run this loop per clip, in this order: 0. Design the CUT with the CONTRAST HOOK + HOOK→PAYOFF→TENSION method (references/clipper_ai_prompt.md) — the DEFAULT, not a remembered option. WHAT-they-do→revenue hook; end on the portable PRINCIPLE, never a bare tactic/wind-down. Spine = references/QA_MASTER_SOP.md.

1. Design EDL + corrections.json + pick music
2. qa_editorial_score.py MUST EXIT 0 on the clip's spoken transcript (HARD GATE, not advisory — blocks plain/revenue-first hooks, tactic/wind-down endings, failed portability, sub-100-word skeletons). Compounding standard: gate, don't warn. [[feedback_compounding_standard_default_plus_gate_2026-06-12]]
3. qa_prebuild_audit.py MUST EXIT 0 before any encode (G1–G6)
4. Build via qa_assembly
5. Extract open frame at t=1.0s. If split-screen, measure guest face position in lower panel. If face center > 35% of panel height → rebuild with lower eye / higher zoom OR document the source-bound limit (tall guest in centered CCAM)
6. Head-trim if onset > 0.10s, reqc, stage
7. SF-sf-audit + scorecard-audit (post-build)
8. ONLY then declare clip ready

A clip that skips any of these steps is a clip that ships defective. The Tier1 batch took 8 rounds because I skipped steps. Don't skip steps.