Long Recording to Short Video

What does Long Recording to Short Video do?

What Problem Does It Solve

One-click generation from long recordings often leads to late-stage discoveries of wrong segments, broken subtitles, music overpowering voices, or incorrect branding, resulting in complete rework. This workflow breaks editing into serial nodes with human approval gates: each node addresses only one type of problem, confirmation is required before proceeding to the next node, and issues don't accumulate across layers.

It does not aim for film-grade editing. The goal is to consistently achieve a publishable, passable final product, enabling users with zero editing experience to understand what they are reviewing at each step.

Node Overview

Stage	Output	What the Human Approval Gate Reviews	Input
Step 1 · Requirements & Asset Intake (P0)	"Final Video Requirements Sheet"	Are factual accuracy, authorization, restricted zones, and terminology complete and correct?	Raw video + User instructions
Step 2 · Segment Selection Plan (P1)	Structured list with timecodes	Are the essential points included? Has the filler been removed? Is the topic complete?	Confirmed requirements sheet
Step 3 · Rough Cut Without Packaging (P2)	Unpackaged spliced sample	Segment transitions, pacing, total duration	Confirmed structure list
Step 4 · Final Packaged Video (P3)	First official version	Subtitle line breaks, music volume, title and end card copy	Finalized rough cut
Step 5 · Highlight Clips (P4)	60–90 second independent clips	Can the clip stand alone without context?	Same raw source material

Process Tiers (Default: Light Process)

The workflow is divided into three tiers based on the intended use of the final video. Step 1 (P0) determines the tier. If the use is not specified, the default is L1 Internal Review, and the full heavy process is not automatically run.

Tier	Use Case	Depth of Intake & Packaging
L1 Internal Review (Default)	Member groups, internal reviews, not for public release	Minimal intake + light packaging: readable subtitles, privacy de-identification, no lead generation; branding (title/end cards, logos, official music records) can be omitted; structure list and verification records are simplified
L2 External Promotion	Public videos for external strangers, potential registrants	Full five-step field contract, title/end cards, subtitles, music, and factual accuracy cleanup
L3 Formal Recruitment	Formal external videos for recruitment, elections, fundraising	Adds per-person authorization records, brand standard review, commercially usable music, and pre-release review checklist to L2

• Non-negotiable baseline across all tiers: The subject appearing on camera must consent to being included. Privacy de-identification must be performed (others' privacy, sensitive medical/financial/fundraising amounts, contact information, and unauthorized segments must not be included). Subtitles must be readable. No lead generation, no collection of contact info, no posting of contact details. Do not write unverified brand slogans or titles.
• Tier upgrades (L1→L2/L3) must be explicitly triggered by the user's intention to release publicly. If L1 material is later intended for public release, return to Step 1 (P0) to upgrade the tier and supplement authorization, brand, and music requirements.

Execution Rules

1. Serial progression: Step 1 → Step 2 → Step 3 → Step 4 (P0→P1→P2→P3) have dependencies. Do not proceed to the next step until the previous step is confirmed. Step 5 (P4) is initiated separately after the main video is finalized. 1b. Tier-based trimming: Whether each step runs a full record is determined by the final video use tier (F0). The L1 light process can simplify intake fields, structure lists, and packaging records, but the tier baselines (authorization, privacy de-identification, readable subtitles, no lead generation/contact collection, no fabricated brand messaging) are not reduced.
2. Segmented prompts: Each step uses one prompt segment, carrying only the input required for that step. Do not compress all requirements into a single large prompt.
3. Layer-specific issue handling: Segment selection issues remain in Step 2 (P1), transition issues in Step 3 (P2), and packaging issues in Step 4 (P3). Do not modify the visual structure during the packaging stage, and do not discuss subtitle styles during the segment selection stage.
4. No fabrication of facts: Mark missing information as {{to be filled}}. Assign the pending item to the relevant step with its priority. Brand messaging must adhere to official guidelines.
5. Read-only first: Step 1 (P0) can scan visuals and transcripts to identify restricted zones, but must not perform segment selection or editing.
6. AI self-check for uncertainties: For uncertainties in audio or video, do not ask the user to listen or watch. The AI should perform secondary recognition, frame extraction, video understanding, and contextual comparison. If evidence is insufficient, adopt the conservative downgrade approach (do not use unverified quotes, do not keep jokes that cannot be substantiated). Human approval gates only review plans and final outputs.
7. Confirmation gate interaction format: For items requiring user decisions (overall plan confirmation, selection choices, method selection, pending items), consistently use dialog boxes: single-value mutually exclusive items use radio buttons, stackable items use checkboxes. Cards include options, rule-allowed objective defaults, the impact of not filling it in, and a fallback "Other (fill in)". Only factual supplements (new restricted zone entries, original quotes, terminology spellings) use free text fields with format examples. Do not generate tables for the user to fill in manually. Uncertainties in audio and video are handled by AI self-checks and not brought into dialogs.

Step 1 · Requirements & Asset Intake (P0, Finalized)

Prompt template: assets/prompt-p0-intake.md Field contract and adjudication rules: references/p0-field-contract.md Test cases and assertions: references/p0-test-cases.md

When executing Step 1 (P0):

1. First perform asset validation. If it fails, follow the stop branch.
2. Collect fields in this order: Asset facts (A) → Video content requirements (B) → Title/End card copy (C) → Content boundaries (D). Ask one question at a time. Enumeration fields use radio/checkbox dialogs (with options, defaults, consequences, and "Other (fill in)"); free text fields provide format examples.
3. After completion, run a completeness self-check and output the requirements sheet, glossary, risk list, and lock description.
4. The requirements sheet, once confirmed and locked, serves as the sole input for Step 2 (P1).

Step 2 · Segment Selection Plan (P1, Finalized)

Prompt template: assets/prompt-p1-selection.md Field contract and adjudication rules: references/p1-field-contract.md Test cases and assertions: references/p1-test-cases.md

When executing Step 2 (P1):

1. Use the locked requirements sheet and the time-coded transcript as input. First perform transcript preprocessing (coverage check, terminology calibration, fragment merging, risk marking).
2. For segments with confidence scores, select by level. For low confidence, mark as uncertain upon complete reading. AI performs secondary recognition and verification for key content. If numbering exists, register number and identity separately; if not, infer from context.
3. Output a segment selection structure list by role (hook/main/key quote/punchline/interaction/landing). Verify each segment for standalone intelligibility (self-sufficiency): arguments, evidence, and landing must be complete; examples, comparisons, and highlights must not be separated from their argument; non-self-sufficient segments should be supplemented with argument/landing or merged under the same argument.
4. Item-by-item check for consistency with must-keep highlights (D1) original quotes, exclude restricted zone list (D2), and remove segments with insufficient authorization.
5. Estimate net duration (actual speaking and interaction time). If exceeding or falling short, provide adjustment suggestions only; do not change the target duration.
6. The structure list proceeds to Step 3 (P2) rough cut only after human confirmation.

Step 3 · Rough Cut Without Packaging (P2, Finalized)

Prompt template: assets/prompt-p2-roughcut.md Field contract and adjudication rules: references/p2-field-contract.md Test cases and assertions: references/p2-test-cases.md Evidence execution protocol (anchor list / frame extraction selection / audiovisual cross-diagnosis / write-back fields): references/p2-evidence-protocol.md Tool limitations and verified dead ends (CLI restricted parameters, Windows parameter pitfalls, parsing conventions, environment baseline): references/tool-constraints.md

When executing Step 3 (P2):

1. Pre-check: structure list must be confirmed, timecodes must be locatable, materials must be readable. If not, return to the corresponding step. Only a clearly marked draft preview can be output without confirmation; no official rough cut.
2. Hard-cut splice in order of the final video sequence, leaving 0.3–0.5 second margins before and after. Cuts must land after a complete argument, Q&A, or reaction.
3. For segments requiring visual verification, perform frame extraction or video understanding. For segments requiring audio verification, use audio understanding or audio feature verification. Candidate segments should be sampled at anchor points based on semantic/audio/visual markers, with multi-point evidence collection. Do not rely on full-video encrypted frame extraction. For high points that cannot be substantiated, downgrade and leave a trace.
4. Fluency check: adjacent segments connect, same-argument segments are adjacent, highlights are attached to the main narrative, beginning and ending echo each other. Jumpy or floating segments return to Step 2 (P1).
5. Record AI time for locational, verification, and splicing subtasks. Human confirmation wait time is not counted. Time is only for reference; pass/fail is not judged by speed.
6. Output only the unpackaged rough cut. Subtitles, music, title/end cards, composition, and color grading are left for Step 4 (P3). The rough cut proceeds to Step 4 only after human confirmation.
7. After splicing, run a completeness self-check on the output: segment files exist and are decodable, final video decodes without errors, transition frames are spot-checked normally, duration and segment count match the structure list. Defective segments are re-cut and re-spliced before returning to the human approval gate.
8. Before outputting the preview, confirm the expected number of revisions. If it can be finalized in one round, output directly from the original high-quality file (master tape). If multiple rounds of revisions are expected, an optional low-resolution preview draft (proxy) can be used for rough cutting. Proxy is only for structure and transition judgment; after confirmation, return to master tape and output at high quality using the same timecodes. Final review of expressions and subtitles returns to master tape for verification. The tier of proxy and speed improvement are not promised unless actually measured.

Step 4 · Final Packaged Video (P3, Finalized)

Prompt template: assets/prompt-p3-packaging.md Field contract and adjudication rules: references/p3-field-contract.md Test cases and assertions: references/p3-test-cases.md

When executing Step 4 (P3):

1. Pre-check: rough cut confirmed, title method (C1-mode), end card method (C2-mode), term spelling table (D3), built-in subtitles and watermarks (A6), music preference (A5-3) all ready. Default title needs title card copy (C1), default end card needs end card copy (C2). Authorization checked per subject. If not confirmed, return to Step 3 (P2). Missing items are marked as pending to be filled; do not draft on behalf of the user.
2. Vertical composition: Re-compose according to the publishing platform and aspect ratio (B4). Subject centered slightly above, leaving safe zones for titles and subtitles. Avoid lower edges with props or clutter. Do not crop key actions or multiple people in frame. Do not stretch or distort.
3. Title/End card executed according to the locked method (default title, default end card; title options: no title, first sentence intro; end card options: own club end card template, no end card). Default title text is taken word-for-word from title copy (C1), with all three lines of titles included. First sentence intro does not create a separate title card; it only overlays one sentence from C1 at the beginning of the main video (approx. 5 seconds). End card does not include lead generation, contact info, or contact details. Slogans require pre-release brand review (R3).
4. Subtitles first compare to existing burned-in subtitles: if position/style/line breaks are essentially consistent, keep them. If enlargement, style change, or rearrangement is needed, evaluate erasing and re-burning. Phrase line breaks: maximum 2 lines per subtitle. Terms follow the term spelling table (D3). Uncertain or downgraded phrases are not put into subtitles. Do not overlay dual subtitles.
5. Background music (BGM) is created by the AI audio generation tool as original instrumental music (text-to-WAV) based on the video scenario. Do not search or download from existing music libraries. The workflow self-formulates generation prompts. Voice takes priority; BGM should not overpower vocals. Internal nicknames, jargon, and memes are cleaned up in the subtitle layer without deleting visuals.
6. After synthesis, run a completeness self-check (decoding, frame spot-check, duration comparison, no dual subtitle overlay), then submit to the human approval gate.

Step 5 · Highlight Clips (P4, Finalized)

Prompt template: assets/prompt-p4-clips.md Field contract and adjudication rules: references/p4-field-contract.md Test cases and assertions: references/p4-test-cases.md

When executing Step 5 (P4):

1. Pre-check: Main video finalized, selected segments locatable in the structure list, authorization individually confirmed for external release. If not finalized, return to Step 4 (P3). Missing items are marked as pending to be filled. Do not waive authorization or draft copy on behalf of the user.
2. Topics are taken from must-keep highlights (D1) and the confirmed structure list. One clip per highlight. Interactions only taken from segments where both parties' identities have been registered.
3. Single clip length: 60–90 seconds. If too short, supplement with arguments under the same topic or merge by topic. If too long, provide split suggestions. Do not pad to lengthen or compress multiple arguments into one clip.
4. Determine independence per clip: hook—argument—landing complete, references clarified, understandable without the main video. Clips are independently sourced; do not hard-cut from the main video.
5. Authorization, de-identification, terminology, subtitles, and music follow the locked conclusions from the main video. Template position follows the locked clip template method (C3-mode, default no template). Title text is taken word-for-word from clip title copy (C3). Title covers do not exaggerate, do not spoil other clips, and do not include lead generation or contact info.
6. After synthesis, run a completeness self-check, then submit to the human approval gate for independent viability review.

Verification & Self-Test

After modifying templates or contracts, run:

python scripts/selftest.py

The script checks field coverage, key mechanisms, contract consistency, and test case assertions. Only proceed to finalize the next step if all checks pass.

Boundaries & Don'ts

• Do not perform frame-by-frame fine editing, cinematic color grading, or multi-track special effects.
• Do not generate unverified brand slogans, titles, or event information on behalf of the user.
• Do not assume material is automatically cleared for external release. Mark unclear authorization as a risk and do not allow it to enter a publishable state.
• Do not skip human confirmation gates and output the final video directly.

Numbered Common Name Reference (Internal anchors preserved for verification and traceability)

Parentheses contain numbering used by the rule layer and self-test script. Read the common name first in the text. Numbering does not change assertions and instance traceability.

Stages: P0 Step 1 · Requirements & Asset Intake | P1 Step 2 · Segment Selection Plan | P2 Step 3 · Rough Cut Without Packaging | P3 Step 4 · Final Packaged Video | P4 Step 5 · Highlight Clips.

Intake Elements:

Number	Common Name	Number	Common Name
A1	Asset list	B1	Final video format and target duration
A2	Speaker name and title	B2	Target audience
A3	Event context information	B3	Desired audience action after watching
A4	Video, audio, and transcription status	B4	Publishing platform and aspect ratio
A5-1	Person appearance authorization	B5	Audience-content match
A5-2	Visual asset authorization	C1	Title card copy
A5-3	Music preference	C1-mode	Title method
A5-4	Segments forbidden/not for sharing	C2	End card copy
A6	Original video subtitles and watermarks	C2-mode	End card method
D1	Must-keep highlights	C3	Clip title copy
D2	Restricted zones and segments to remove	C3-mode	Clip template method
D3	Proprietary term spelling table	E1/E2/E3	Clip topics / Clip duration / Can it stand alone
F0	Final video usage tier	R3	Pre-release brand review

Common Jargon: Gate = human approval confirmation point; Lock = confirmed and finalized; Derivative = separately cut short clip; Downgrade = when evidence is insufficient, downgrade and not use that statement or footage; Master = original high-quality file; Proxy = low-resolution preview draft; Self-sufficiency = whether this segment can be understood by the audience on its own; Net duration = actual speaking and interaction time used.