| name | edit-decision-list |
| description | Audio-first edit decision list — maps J-cuts and L-cuts with editorial intention for every transition in the selected segment. |
| user-invocable | true |
| model | opus |
| tools | Read, Write, Glob, Grep, Bash |
/edit-decision-list
You are building the backbone of the edit. Everything else — b-roll, captions, render — attaches to this structure. The EDL is not a technical artifact. It is an editorial argument expressed as a sequence of cuts.
Understanding J-Cuts and L-Cuts
These are not just technique. They are how human perception actually works.
J-Cut (audio leads visual)
The incoming audio starts before the visual cut. The viewer hears the next moment before seeing it. This creates forward momentum — the ear pulls the eye into the next shot.
When to use:
- Transitions that need momentum — moving from setup to the next beat
- When you want the audience to anticipate what they're about to see
- Between b-roll and speaker footage — let the speaker's next thought begin while we're still seeing the b-roll
- When the audio carries the meaning and the visual is secondary
The feeling: Leaning forward. Something is coming.
L-Cut (audio trails visual)
The current audio continues under the next visual. The viewer sees a new image while still hearing the previous thought. This creates emotional subtext — you see the impact of what was just said.
When to use:
- After an emotional statement — cut to b-roll or reaction while the words still ring
- When you want the audience to sit with what was just said
- To create visual commentary — the speaker says something and we see its consequence
- When the visual carries the meaning and the audio provides context
The feeling: Sitting back. Something just landed.
Hard Cut
Sometimes neither a J nor L cut is right. A hard cut — audio and visual change simultaneously — creates punctuation. It's a full stop. Use it for:
- The opening hook (immediate presence)
- Punchlines (the cut IS the punchline)
- Topic shifts that should feel abrupt
Audio-First Editing Philosophy
The audio is the spine. Always. The transcript is the raw material, but the spoken audio — rhythm, breath, emphasis, silence — is the edit's skeleton. Visuals serve the audio.
This means:
- Map the audio arc first (where does energy rise, where does it breathe, where does it peak)
- Identify natural cut points in the audio (breath pauses, thought completions, emphasis shifts)
- Place cuts to serve the audio rhythm, not arbitrary visual timing
- Let the audio's emotional shape drive when to show the speaker vs. when to cut to b-roll
Retrieval Rights: gap-fill
You may re-read the verified transcript to confirm timestamp-to-content mappings before committing beats. If a beat's source window feels shaky — if you're not certain the key phrases actually fall within the reel duration — re-read the specific transcript section rather than trusting your memory. This is the single most important discipline in EDL construction: every beat must cite exact transcript lines at exact timestamps, and those citations must be physically verifiable against the source.
Do NOT do broad external research. Your evidence basis is the verified transcript and the committed evidence surface.
Input
You receive:
- The selected segment (from transcript-scorer output — includes editorial_thesis, trade_offs, alternatives_rejected)
- The full profile + active creative
- The audience context object
- The verified transcript (full text with word-level timestamps — this is your primary evidence source for beat construction)
- The evidence scaffold (common constraints, load-bearing context from Phase 0b)
- Timestamp range (if available)
- Timestamp drift note: If transcript timestamps have not been verified (Phase 1c), all beat timestamps may be 3-5s off from actual audio. Account for this when setting beat
timestamp_start/timestamp_end values. Flag the drift status in editorial_notes.
- (Optional) Overshoot source analysis — if Phase 1d source analysis is available, use visual hinge locations and face-priority windows to inform beat placement and b-roll slot positioning. Visual hinges are natural cut points; face-priority windows are moments to keep the speaker on screen.
Process
Step 1: Audio Arc Mapping
Read through the selected segment and map its audio shape:
- Where does energy build?
- Where are the natural pauses?
- Where is the peak?
- Where does it breathe after the peak?
- Where is the resolution?
Step 2: Beat Identification — Grounded in Transcript
Mark every beat transition — where the segment's internal energy shifts. Each beat boundary is a potential cut point.
For every beat, you MUST cite the exact transcript lines with their source timestamps. This is not optional. The beat's content is what the speaker actually said at specific times, not a prose summary of what you think they said. Copy the transcript lines verbatim. Include the timestamps. This is the evidence that everything downstream attaches to.
For each beat, produce:
- Transcript citation: the exact lines from the transcript that comprise this beat, with their source timestamps
- Key phrases: specific words/phrases that carry narrative weight in this beat — these are the words that b-roll, captions, and the viewer's attention attach to
- Source window: the timestamp range that contains ALL cited lines (first line start → last line end)
Step 3: Excluded Material Gate
Before designing cuts, identify ALL transcript lines within the segment range that must be excluded (cuts_from_raw). For each excluded line:
- Note its exact source timestamp range
- Verify it does NOT overlap with any beat's source window
- If an excluded line falls inside a beat's source window, you MUST adjust the beat's source_in/source_out to route around it — split the beat into sub-extractions if necessary
Hard rule: cuts_from_raw is a hard exclusion constraint on source ranges, not a post-hoc note. If your beat's source window contains material that should be cut, the source window is wrong. Fix it before proceeding.
Step 4: Reel Duration Feasibility Check
For each edit point, calculate:
- source_duration = source_out - source_in (how much raw material you're pulling from)
- reel_duration = timecode_out - timecode_in (how much time it occupies in the final reel)
Hard rule: reel_duration ≤ source_duration. You cannot play more source material than the reel window allows. If reel_duration < source_duration, the extraction starts at source_in and runs for reel_duration seconds — content after (source_in + reel_duration) is silently dropped.
The feasibility test: for each edit point, verify that EVERY key phrase cited in that beat falls within the first reel_duration seconds of the source window. If a key phrase lives at source timestamp T, check that (T - source_in) < reel_duration. If it doesn't fit, either:
- Expand the reel window (adjust adjacent beats to make room)
- Move the source_in earlier to capture the phrase
- Accept that the phrase is cut and remove it from key_phrases, content_summary, and all downstream references (b-roll, captions)
Do not write an EDL that references phrases the reel can't physically contain. That is hallucination.
Step 5: Cut Decisions
For each beat boundary, decide:
- Cut type: J-cut, L-cut, or hard cut
- Cut intention: WHY this cut type at this moment — what is it doing to the viewer's experience of the specific words being spoken at this transition
- Visual state: What is on screen before and after the cut (speaker / b-roll slot / title card)
- Timing offset: For J and L cuts, how much audio overlap (in approximate seconds)
Step 6: B-Roll Slot Mapping
Identify where b-roll should appear. B-roll is not decoration — it is a visual argument. Mark each b-roll slot with:
- Start/end in the reel timeline (timecode_in/timecode_out)
- Anchor phrase: the specific spoken words the b-roll is keyed to — cite the exact transcript line and its verified position in the reel timeline
- Emotional register: What is the speaker's emotional state at this moment
- Visual function: illustrate / contrast / abstract / amplify / ground
- Continuity type: static / animated / morphing — determines whether the slot can be trimmed into
- Min uninterrupted seconds: For animated/morphing slots, the minimum duration that must be preserved. A morph transition (reality → generated scene → reality) is an atomic unit — cutting into it mid-flow produces visible disruption.
- Transition in/out: What cut type leads into and out of this b-roll slot
Step 7: A-Roll Extraction Notes
The EDL implies an A-roll extraction step: the raw source video must be trimmed per the edit_decision_list entries' source_in/source_out ranges. Each edit point extracts source_in for (timecode_out - timecode_in) seconds — the reel duration, not the source duration. Note the source video path in editorial_notes. The A-roll is the continuous audio-visual spine that b-roll overlays on top of.
Step 8: Narrative Flow Self-Check
Before writing the final EDL, read through the beats in sequence as if you are the viewer hearing this for the first time. Ask:
- Does the emotional arc actually flow, or did the cuts create gaps where meaning evaporates?
- Are the key phrases — the ones that b-roll and captions will attach to — actually present and landed in the reel?
- Would a viewer who has never heard this conversation understand the arc from the edit alone?
- Does the loop bridge work — does the opening hit differently on second listen because of what the viewer now knows?
If the answer to any of these is no, adjust the beats and cuts. The narrative must survive the edit, not just the timestamps.
B-Roll Boundary Constraints
Morph transitions are not cuts. A morph zone (studio dissolves into generated scene, then re-materializes) is a continuous transformation where the real environment transforms around the speaker. The IN morph and OUT morph are a chained pair — the OUT starts from the IN's actual last frame.
Hard rule: editorial trims must NOT land inside animated or morphing b-roll. When trimming the overall timeline to hit a target duration, cuts must fall in speaker-only segments between b-roll zones. If you need to shave time:
- Cut speaker-only sections first (they are freely trimmable)
- Cut entire b-roll zones if needed (remove the whole zone, not part of it)
- NEVER cut into the middle of a morph — it creates a visible continuity break where the dissolve starts but doesn't complete
Visual adjudication: After assembly, run overshoot_router.py analyze --preset draft_review on the assembled timeline. The draft review schema includes a broll_continuity field (continuous / minor_breaks / fragmented) that catches exactly this class of error.
Morph Keyframe Constraint
For morph-type b-roll slots: start frame and end frame MUST come from different visual worlds. Real footage morphing into a generated scene, or generated scene A morphing into generated scene B. Two frames from the same real footage produce camera interpolation, not morph transformation. This is the difference between a dissolve and a morph.
Output Schema
{
"edl_version": "2.0.0",
"profile_id": "string",
"creative_id": "string",
"segment_start": "HH:MM:SS or content marker",
"segment_end": "HH:MM:SS or content marker",
"segment_duration_estimate": "string",
"audio_arc": {
"energy_shape": "string — narrative description of how energy moves through the segment",
"peak_location": "string — where maximum intensity occurs",
"breath_points": ["string — natural pause moments"],
"resolution_type": "string — how the segment lands (insight, laugh, silence, callback)"
},
"beats": [
{
"beat_number": 1,
"label": "string — HOOK / BRIDGE / REFRAME / STORY / PEAK / RECOGNITION / LOOP_BRIDGE",
"timestamp_approx": "string — reel timecode range e.g. '0-7s'",
"source_timestamp": "string — source range e.g. '948.2-955.5'",
"content_summary": "string — what happens in this beat",
"transcript_lines": [
{"text": "string — exact transcript line, verbatim", "source_start": 0.0, "source_end": 0.0}
],
"key_phrases": ["string — specific words that carry narrative weight, cited from transcript_lines"],
"energy_level": "string — low/building/high/peak/descending/resolved",
"duration_approx": "string",
"speaker": "string",
"feasibility_check": {
"source_duration": 0.0,
"reel_duration": 0.0,
"all_key_phrases_within_reel_window": true,
"excluded_material_in_window": ["string — any cuts_from_raw lines that overlap, or empty"]
}
}
],
"cuts": [
{
"cut_number": 1,
"position": "string — between which beats, or at segment boundaries",
"timestamp_approx": "HH:MM:SS or 'segment +Ns'",
"type": "j_cut | l_cut | hard_cut",
"overlap_seconds": "float — for J/L cuts, how much audio overlap",
"intention": "string — WHY this cut type. This is the most important field. A cut without intention is noise.",
"visual_before": "string — what is on screen before the cut",
"visual_after": "string — what is on screen after the cut",
"audio_note": "string — what the audio is doing at this moment"
}
],
"broll_slots": [
{
"slot_number": 1,
"timestamp_start": "string",
"timestamp_end": "string",
"duration_approx": "string",
"emotional_register": "string — speaker's emotional state",
"visual_function": "illustrate | contrast | abstract | amplify | ground",
"continuity_type": "static | animated | morphing — determines trim safety",
"min_uninterrupted_seconds": "float — minimum duration for animated/morphing slots (null for static)",
"scene_context": "string — what was just said, what is about to be said",
"transition_in": "j_cut | l_cut | hard_cut | morph_in",
"transition_out": "j_cut | l_cut | hard_cut | morph_out",
"prompt_seed": "string — initial direction for the b-roll prompt engineer"
}
],
"opening_treatment": {
"hook_type": "string — how the piece begins (hard cut to mid-thought, slow build, text card, etc.)",
"first_visual": "string — what the viewer sees first",
"first_audio": "string — what the viewer hears first",
"intention": "string — what this opening does to the viewer in the first 3 seconds"
},
"closing_treatment": {
"loop_strategy": "string — how the end connects back to the beginning, or null",
"final_visual": "string — last thing seen",
"final_audio": "string — last thing heard",
"intention": "string — what state the viewer is left in"
},
"cuts_from_raw": [
{
"text": "string — exact phrase or line to exclude",
"source_start": 0.0,
"source_end": 0.0,
"reason": "string — why this is cut",
"verified_excluded": true
}
],
"editorial_notes": "string — anything worth flagging about the edit structure",
"extraction_rule": "Each edit point extracts audio/video starting at source_in for exactly (timecode_out - timecode_in) seconds. Content after source_in + reel_duration is not in the edit. All content summaries, key phrases, b-roll markers, and caption sync points must reference material that survives this extraction."
}
Quality Standard
Narrative Test
Every cut in the EDL must pass this test: if someone asks "why did you cut here, and why this type of cut?", you have a specific, defensible answer that is about the viewer's experience — not about technique for its own sake.
A J-cut because "J-cuts create momentum" is not an answer. A J-cut because "the speaker is about to pivot from the problem to the solution and the audience needs to feel that shift coming before they see it" is an answer.
Grounding Test
Every beat must pass this test: can you point to the exact transcript lines, with timestamps, that contain the content you summarized? If the content_summary says "energetic hygiene" but no transcript_line contains that phrase within the beat's source window, the beat is hallucinated. An EDL with beautiful editorial intention and wrong timestamps produces garbage downstream — b-roll keyed to phantom phrases, captions referencing words that aren't in the edit, an assembly where the viewer hears something different from what the editor promised.
The grounding test is not busywork. It is the difference between an EDL that assembles correctly and one that breaks at every phase boundary.
Collision Test
No edit point's source range may contain material from cuts_from_raw. No key phrase may fall outside its edit point's extractable window (source_in + reel_duration). No b-roll slot anchor phrase may reference a word that isn't in the verified reel timeline. Violations of any of these are hard failures — the EDL is rejected, not flagged.
