AI Documentary Production

First Use: Environment Setup

Images: By default, the user generates images with GPT; AI does not call image APIs; image_mode=agent requires explicit approval.
Environment: Python 3.10+, ffmpeg, ffprobe; contact_sheet.py additionally requires Pillow; no PyYAML dependency.
H3: If you haven't configured MiniMax H3, first check references/00_setup.md:
- Register/verify identity/recharge on the open platform;
- Create an API Key;
- Set the MINIMAX_API_KEY environment variable or use a key file;
- Align base_url / model name / resolution;
- Run a 4-second test first, then batch.
Keys should never be included in code, logs, or git.
If H3 is unavailable, switch providers or create an animatic first; do not pretend to generate.
If H3 is not configured, do not proceed to G5.

When to Use

User wants to create a character/theme short film, with duration determined by the brief;
30 seconds, 1 minute, 3 minutes are just examples, not hard limits;
Involves real people, historical materials, AI reenactments;
Needs to go from zero to a deliverable finished film;
Requires cost control, source retention, and credits writing.

Not Applicable

Multi-character portraits, interview-style feature films;
Commercial advertisements, pure product videos;
Full animation, full virtual production;
On-site documentaries requiring real filming.

Duration has no hard threshold; extra-long projects require additional asset management, but the workflow remains the same.

Must Ask Before Starting

Finished film duration, aspect ratio, frame rate, language, platform;
Who is the character, which part of the story, which facts cannot be wrong;
Budget ceiling, whether retries are allowed;
Who generates images (image_mode=human/agent), which video model, which editing software;
Whether narration is real voice or TTS; whether music is a separate track;
Whether historical photos/music licenses exist; whether an AI disclaimer is needed;
Delivery format and deadline;
Profile: short (approx. 15–60s) / standard (approx. 2–5min) / custom (as needed);
model_routing: hybrid / gpt_only / codex_deepseek;
Whether H3 is configured (Key/base_url/model/quota); if not, teach configuration first;
Opening card: none / static; how many seconds; whether it counts toward the target duration; title text to be finalized at G2.5.

Without a brief, do not proceed to research.

Duration and Profile

Duration is determined by duration_target_sec in project.yaml, not hardcoded.

profile	example duration	shots (example)	segments (example)	music (example)
short	15–60s	8–18	1–2	1 segment
standard	2–5min	24–60	3–5	3–5 segments
custom	user-defined	user-defined	user-defined	user-defined

These are estimates, not gates; 30 seconds and 3 minutes are just two common examples.

G4 Image Handoff

After prompts are completed, first remind the user to generate images, then work in parallel:

Send image_batch_request.md: shot number, filename, prompt location, quantity, priority; if opening card is enabled, 0-1_card.png goes into the first batch;
State clearly: "You can start generating images now";
Once the user starts, the agent does work that does not depend on images/video: subtitles, cue sheet, cost table, sources/credits, editing skeleton, QA for next batch of prompts;
Do not review images, generate video, or do final compositing;
After the user returns images, review them first; do not re-discuss already locked narration/shots;
Opening card is a static card, do not use H3; verify text character by character; if Chinese is unstable, switch to post-production overlay.

Do not silently work on other things and keep the user waiting; the image generation task must be handed off first. See references/13_parallel_work.md for details.

Model Routing

Default is hybrid:

Stage	Model	Task
G0–G3	Codex + DeepSeek	brief, research, narration, storyboard, prompts
G4	GPT	generate/review images; `image_mode` determines who generates
G5–G8	Codex + DeepSeek	tasks, costs, editing files, QC; temporarily switch to GPT when reviewing images

Use DeepSeek for text and files, GPT for pixels and visuals;
image_mode=human: user generates with GPT; image_mode=agent: Codex + image tools generates (requires explicit approval);
Write image_handoff.md before switching; project files serve as handoff credentials;
See references/10_model_routing.md for details.

9 Stages and G0–G8 Gates

Stage	Deliverable	Gate
G0 Brief	`project_brief.md`	Duration, budget, rights, platform locked
Research/Fact Check	`fact_check.csv`, `research.md`	G1 Facts locked
Script/Narration	`script.md`, `vo.csv`, `vo_text_review.md`, `voice_audition.md`, `vo_lock.md`, `vo_review.md`	G2 Narration locked (copy + voice + version)
Storyboard/Timeline	`beats.csv`, `shots.csv`, `segments.csv`, `timeline.csv`, `animatic.mp4`, `v0_review.md`	G3 Beats claimed + shot table + v0 confirmed
First Frames (per image_mode)	`first_frames/`, `image_review.md`, `image_handoff.md`, `image_batch_request.md`	G4 First frames locked
Video Generation	`h3_package/`, `clips/`, `tasks.json`, `cost.csv`	G5 Video locked
Rough Cut/Subtitles	`cut/`, `subtitles.srt`	G6 Picture locked
Music/Sound	`music/`, `cue_sheet.csv`	G7 Sound locked
Final QC/Delivery	`final.mp4`, `final_qc.md`, `credits.md`	G8 Delivery

Rule: Do not proceed to the next gate until the current one is passed; do not spend money.

Opening Card

Only ask at G0: "Do you want one, how many seconds, does it count toward the target duration"; title text is finalized at G2.5.

G0: opening_card: none|static; opening_seconds; opening_in_target;
G2.5: Main title/subtitle/font style; opening_card.md character-by-character confirmation;
G3: Written as a 0-1 static shot in shots.csv (static=yes, no narration), a CARD row in segments.csv (static=yes);
G4: Goes into the first image batch, filename first_frames/0-1_card.png; verify text character by character when images come out;
G5: Do not use H3; export_h3.py skips static=yes; do not include it in the H3 task table;
G6: Static hold in editing; can apply a slow 1–2% push-in;
Opening card duration is deducted from the target duration or explicitly added to the total length, decided at G0.

The opening card is a "duration decision during project initiation", not a "decoration option when generating images".

Narration Review Order

Copy Review: Show the user the full vo.csv, timecodes, and fact notes; get user confirmation before doing voice;
Voice Audition: Generate 2–3 candidates for the same sentence, let the user choose A/B/C or record their own;
Full Scratch: Generate the complete narration with the chosen voice;
v0 Animatic: Only then create static frames + narration;
Do not proceed to the next step without confirmation at any step.

Narration First

This is the core sequence of the skill:

Finalize narration first, then shots, first frames, video;
One narration line matches 2–3 shots as the norm;
One shot per line is a special case (short reaction, slogan, single-shot montage), not the default;
Narration stands independently as vo.csv, do not fragment by shot;
Do not generate TTS per shot; generate per sentence/paragraph, place the entire segment on one audio track;
Each sentence carries one piece of information; every 1–2 sentences contains at least one hard fact (year/person/number/action/cause-effect);
Poetic imagery must have a concrete anchor, a chain, and be grounded; "door/seed" can be a motif, not an automatic error; anchorless slogans are empty words; explain cause and effect first, then poetic language;
After writing, run narration_lint.py, then read once to finalize;
When editing, use J-cut/L-cut: picture cuts, narration continues;
Leave 0.3–1.0 seconds of breathing room between sentences; do not force a 5-second narration onto a 5-second shot; it is best to bear the breathing room with a short shot without narration, not inside a shot with narration;
Breathing room should only be placed between narration sentences; cut points must fall within the breathing room, not inside a sentence with narration;
After narration is locked, do not directly modify the old version; first run vo_diff.py, only change affected files;
One narration segment can span multiple shots; a normal shot should have at most one narration segment; a long_take=yes shot can contain multiple narration segments, but each must be completely within the shot;
The breathing room window between adjacent narrations must contain at least one shot cut point; sentence start/end can fall inside the shot, not required to align with the cut point; when there is a range, use the larger breathing room;
After finishing shots.csv / timeline.csv, first produce a v0 animatic (static/black frames + narration), explicitly remind the user to listen for breathing room between sentences; confirm before proceeding to first frames/video;
Show the copy first, then determine the voice, then make v0: without user confirmation of copy and voice, do not generate TTS/animatic.

First the narration, then the visuals; not 40 shots first, then cram a sentence into each one.

File Contracts

`shots.csv`

id,story_segment,scene_id,duration,shot_size,camera,scene,action,expression,vo_ids,anchor,first_frame,video,status,cost,notes,long_take,static

story_segment is the narrative segment number, not the H3 generation segment; generation segments are in segments.csv under generation_id;
scene_id is the scene number, must match segments.csv;
duration in seconds, supports decimals;
status only uses todo/image_ready/video_running/video_ok/conditional/retake/final;
vo_ids can be empty; a normal shot should have at most one vo_id; a long_take=yes shot can write V01|V02;
long_take=yes: one continuous long take, allows 2+ narration segments, but each must be completely within the shot;
static=yes: static opening card/insert card, do not use H3; must have a first_frame, use a static=yes segment in segments.csv;
This is the source of truth for visuals; narration is in vo.csv, linked via vo_ids.

`vo.csv`

vo_id,start,end,text,shot_ids,notes
V01,0.0,4.2,"A complete narration line","1-1|1-2","4.2s; 0.8s breathing room after"
V02,5.0,6.2,"Next line","1-3","1.2s; 0.8s at the end"

start/end are absolute seconds;
One narration line can span 2–3 shots;
One shot per line is only used in special cases;
Leave 0.3–1.0 seconds of breathing room between sentences;
One shot_id can belong to only one cue; one cue can cover multiple shots;
Subtitles are generated from vo.csv, not bound to shots.

`beats.csv`

beat_id,story_segment,kind,shot_id,vo_id,text,notes
B01,1,action,1-1,V01,Calculating behind the counter,
B02,1,action,2-1,V02,Putting the paper into an envelope,

A beat is the smallest unit of story, not a musical beat;
Each beat is exactly claimed by one shot_id;
Each non-static shot has at least one beat; static=yes opening cards can have no beat; each narration segment has at least one beat;
One narration can span multiple beats, repeating the same vo_id on each row;
Beat order must be consistent with shot order; missing, duplicate, or out-of-order beats are caught by validate.py.

`segments.csv`

generation_id,scene_id,cut_order,shot_id,start,end,duration,keyframe,static,notes
S01,1,1,1-1,0,5,5,first_frames/1-1_first.png,no,single-cut segment
S02,1,1,1-2,5,9,4,first_frames/1-2_first.png,no,cut 1 within segment
S02,1,2,1-3,9,15,6,first_frames/1-3_first.png,no,cut 2 within segment

generation_id = one H3 generation; a segment ≤15 seconds, does not cross scenes;
scene_id must equal the scene_id of the same shot in shots.csv; scene change requires a new segment;
cut_order = the cut order within the segment, starting from 1 sequentially;
keyframe = composition anchor image for this cut; cut 1 pins at 0.00 seconds, subsequent cuts pin at the cut point;
static=yes = static card, not for H3; must be a single cut; export_h3.py will skip;
Each shot can only appear in one generation;
Each cut's start/end/duration must equal the cumulative start and end in shots.csv; the entire film must have no gaps and no overlaps;
validate.py catches overruns, misalignments, gaps, and duplicates.

`h3_prompts.csv`

shot_id,prompt,soundscape,notes
1-1,"Maintain the first frame; the pen tip moves slowly, the lamp flame sways gently.","Sound of paper and pen, distant wind.",

export_h3.py groups by segments.csv and stitches each cut's prompt into a multi-image alignment prompt;
Outputs h3_package/<generation_id>/prompt.md + manifest.json;
Does not call APIs; generation is still executed by the H3 task table.

Narration Versions

vo_versions/vo_v1.csv
vo_v1_scratch.wav
vo_lock.md

Before locking, do a read-through/TTS, measure actual duration;
After locking, only add new versions, do not directly modify the old version;
To change narration, first run vo_diff.py, then modify affected files;
See references/12_change_control.md for details.

`project.json`

Machine configuration generated by init_project.py: profile/model_routing/image_mode/duration_target_sec/shot_target/fps/resolution/providers. project.yaml is the human-readable version; scripts read project.json.

`tasks.json`

[{"generation_id":"S01","provider":"minimax_h3","model":"MiniMax-H3","resolution":"768P","duration":9,"status":"todo","prompt":"h3_package/S01/prompt.md","first_frames":["first_frames_h3/1-1_first.png"],"cost_cny":0.0}]

generation_id corresponds to the generation segment in segments.csv;
duration must be an integer between 4 and 15;
prompt / first_frames are paths to the production package;
Failed tasks are not billed; retries are listed separately;
Old per-shot tasks can use id, but h3_tasks.json always uses generation_id.

Delivery Package

final.mp4
subtitles.srt
vo.wav
vo.csv
music/A-E.wav + cue_sheet.csv
credits.md
source_list.csv
fact_check.csv
final_qc.md

Hard Rules

Images are default generated by the user via GPT; AI does not call image APIs; image tools are only allowed when image_mode=agent and the user explicitly approves;
Video defaults to a single version; only add a second version if the first is unusable;
Generated video is only for ambient sound/dialogue, not for music/BGM;
768P source material may be 1344×768 (7:4), final delivery at 1920×1080;
First frames must be 16:9, otherwise image-to-video aspect ratio will drift;
Historical photos must cite sources; AI visuals must be labeled "reenactment";
Subtitles: 16–20 characters per line, max 2 lines, ≥100px left/right margins;
Each music segment should have 1.5–2 seconds of handle; low point is not silence;
Costs are calculated based on task logs, not from memory;
Before delivery, use ffprobe to verify resolution/frame rate/audio;
First finalize narration, then shots; default is one narration line spanning 2–3 shots; one shot per line is a special case;
Narration should not fill the entire shot; a 5-second narration paired with a 5-second shot will feel rushed; leave 0.3–1.0 seconds of breathing room between sentences;
Narration can be poetic, but poetry must have physical/action anchors; only the explanatory section requires one hard fact per sentence; do not mix the two standards;
After narration is locked, do not rewrite the entire project because of one changed line; first run vo_diff.py, then only modify affected files;
After G4 prompts are completed, first remind the user to generate images; only after the user starts generating should you work on parallel tasks that do not depend on images/video;
One narration segment can span multiple shots; a normal shot should have at most one narration segment; a long_take=yes long shot can contain multiple narration segments, but all must be within the shot;
The breathing room window must contain the shot cut point; do not require sentence start/end to align with the cut point; when there is a range, use the larger breathing room;
After the v0 animatic is produced, first remind the user to check the breathing room between sentences; do not produce first frames or generate video before confirmation;
The voice must be auditioned with 2–3 candidates and selected by the user; do not default to a specific TTS;
Narration defaults to "clear but not plain": first explain the character/event/cause-effect, then use concrete imagery and short sentences to create space; allow adding 1–2 sentences, but each must bring new information or a new image; do not pad with adjectives;
Each beat in beats.csv must be claimed by exactly one shot; each shot and each narration segment must have beats covering them; missing, duplicate, or out-of-order beats are caught by validate.py;
segments.csv: one segment ≤15 seconds; cut order within the segment is sequential, each cut has a keyframe, each shot appears only once; the start/end of each cut must equal the cumulative time in shots.csv; use export_h3.py to produce the H3 multi-image alignment production package;
Opening card: G0 sets duration, G2.5 sets text, G3 writes 0-1 static=yes, G4 generates image, G5 does not use H3.

Provider

config/providers.yaml is documentation; the machine configuration is providers.json (generated by init_project.py, validated by timeline.py).

Defaults: images=user_gpt, video=minimax_h3, narration=capcut_tts, music=local_midi_fluidsynth, editing=capcut, QC=ffmpeg.

The image provider is always human; this does not change with configuration.

Commands

python scripts/init_project.py --root my-film --profile short --duration 30 --shots 10 --segments 2
python scripts/timeline.py --shots shots.csv --vo vo.csv --vo-out vo_report.csv --project project.json --min-gap 0.4
python scripts/timeline.py --shots shots.csv --vo vo.csv --max-density 0.85 --allow-single
python scripts/narration_lint.py --text vo.csv --style lyrical --max-chars 28
python scripts/validate.py --shots shots.csv --vo vo.csv --beats beats.csv --segments segments.csv --project project.json
python scripts/validate.py --shots shots.csv --vo vo.csv --beats beats.csv --segments segments.csv --project project.json --strict-files
python scripts/validate_tasks.py --tasks h3_tasks.json --segments segments.csv
python scripts/selftest.py
python scripts/export_h3.py --shots shots.csv --segments segments.csv --prompts h3_prompts.csv --out h3_package
python scripts/timeline.py --shots shots.csv --vo vo.csv --animatic animatic.mp4
python scripts/vo_gap_report.py --vo vo.csv --audio vo_scratch.wav --out vo_gap.md
python scripts/vo_diff.py --old vo_versions/vo_v1.csv --new vo.csv --out vo_diff.md
python scripts/vo2srt.py --vo vo.csv --out subtitles.srt --max-chars 18
python scripts/contact_sheet.py --clips clips --out review --frames 6
python scripts/cost_report.py --tasks tasks.json --out cost.csv --default-model MiniMax-H3

Scripts: init_project.py, validate.py, selftest.py, export_h3.py, timeline.py, narration_lint.py, vo2srt.py, vo_gap_report.py, vo_diff.py, contact_sheet.py, cost_report.py. --animatic defaults to black frames if first frames are missing; --animatic-missing error can change this to error out.

validate.py handles structural contracts (schema / beats / segment cuts / coverage); G3 does not check keyframe files by default, G5 checks with --strict-files. timeline.py --vo handles timing (breathing room / density / long takes). Run both; they are not interchangeable.

Error Handling

Video failure: Check task status; failed tasks are not billed; switch to H3-Max or modify the first frame, do not blindly re-run;
Face changes/multiple hands: Regenerate or switch to profile/back view;
Action did not occur: Regenerate, do not fix in editing;
Subtitle overflow: Change style/break lines, do not shrink to unreadable;
Music jarring: Add pickup, expression ramp, handle; do not hard cut;
Export proportion wrong: Go back to project and set 1920×1080, do not stretch.

References

SOP.md: Human-readable workflow and checklist
references/: Stage details (including 00_setup.md, 09_vo_spans.md, 10_model_routing.md, 11_narration.md, 12_change_control.md, 13_parallel_work.md, 14_v0_animatic.md, 15_beats_segments.md)
providers/: Replaceable tools
templates/: Project card, shot table, prompts, review forms, cue sheet, credits, final QC

yinguozhang/ai-documentary-production

ai-documentary-production 是什么？

在你喜欢的 AI 中提问

文档