AI-assisted documentary studio · Two films finished · October 2026
Claude agents research the story and write the script. ElevenLabs gives it a voice, a score and sound. Blender draws the maps, graphics and 3D, and ffmpeg mixes the film. The producer approves the concept, the script and the storyboard before any money is spent, and watches every cut.
2 films, 3:52 and 4:43 ~$64 and ~$101 all-in at API prices ~2 h and ~2.5 h of working time ~$17 of Claude plan value for both
AI background
How winter ice became a global trade, and why Prague's harvest ended with a dam.
The proof of concept. Most of the pipeline was built while making it. 23 AI images, 2 archival photos, maps, infographics and 3D.
JWST keeps finding tiny red points of light in the early universe. What are they? The 2026 argument, explained.
Zero AI images. Real JWST imagery and data, spectra drawn from archive files, graphics and 3D visualizations.
s01a deepzoom00:03
Real · NASA, ESA, CSA, I. Labbe, R. Bezanson, A. Pagan · CC BY 4.0
AI images01c parallax00:22
AI · Leonardo, Nano Banana 2.1
s04e spectrum02:19
Data graphic · JWST/NIRSpec, RUBIES · DAWN JWST Archive
s03b map01:16
Map · Natural Earth data
s05c 3d02:45
Visualization · after Rusakov et al. 2026
s04c infographic01:39
Graphic · Blender
s06e 3d03:47
Visualization · Blender
s07g 3d02:56
3D reconstruction · from the 1913 photograph
Process · 11 stages · 2 approval gates
Each film is one folder and one storyboard file. Every stage reads the storyboard and caches what it makes, so a change re-does only the shots it touches. The producer is asked at the gates and for matters of taste. Everything else runs on its own.
Angle, length, audience and narrator.
Outputconcept.md
Producer decidesOne agent researches the facts, another finds real imagery and its licences. Every fact is sourced and tagged with a confidence level.
Outputresearch/notes.md · research/images.md (catalogue with licence and credit lines)
AgentsSections, timing, visuals per section, image policy and the music plan.
Outputlayout.md
Producer approvesNarration by section, a fact-check table and a list of phrases to avoid.
Outputscript.md
Gate 1 · producer approvesShots, techniques, sources and licences, camera, overlays, sound effects, music movements and silent passages. A validator checks it against the script and the trust rules.
Outputstoryboard.yaml · review page out/storyboard.html
Gate 2 · producer approvesNo credits are spent above this line
Downloads of real imagery, data and footage. AI images only where the storyboard allows them.
Outputassets/archival/ · assets/images/
Producer approves the download listOne text-to-speech call per section. Its character timings become the film's timeline.
Outputassets/voice/ · out/timeline.json
Voice audition firstOne piece with one movement per act, its lengths taken from the timeline. Cuts snap to the music's onsets.
Outputassets/music/ · out/beats.json
Runs on its ownBlender and Python renderers, one per technique. New techniques are built by agents working in parallel, each in its own files.
Outputshots/<id>.mp4
Runs on its ownDissolves, overlays (lower thirds, captions, credits, AI badges, titles over picture), ducked music, sound effects, loudness at −14 LUFS and an .srt file.
Outputout/<slug>.mp4
Runs on its ownFrames pulled from every shot, flicker checks, reading times and loudness. Then the producer watches. Fixes go back to stages 5–10.
Outputfixes and lessons for the pipeline
Producer watchesLong conversations get expensive, because every step re-reads the whole context. The repository docs carry everything over to the next session.
Research and new renderers go to agents. Each gets its own files, a clear brief and shared conventions. Their reports are checked before anyone believes them.
Concept, script, storyboard, downloads and voice, plus matters of taste. Nothing that costs credits happens before the storyboard is approved.
Contact sheets, frames pulled from the final film, an automatic flicker check. Rendering without looking cost a re-render twice.
Tools · prices as of October 2026
Every tool used on the two films, what it does here, and the settings that worked.
Claude Opus 5.5 at the default effort runs the production: research, layout, script, storyboard, all of the pipeline's code, the renders and the mix, and the checks. Agents take the independent pieces, such as research and new renderers.
All audio, through the API. Creator plan, $22 a month for 131,000 credits. The free plan has no commercial licence.
eleven_v4, about 14.4 characters a second. One request per section with character timestamps, which drive the cutsmusic_v2_5. One continuous score made of chunks, one per movement (3–120 s each). Roughly 900 credits a minuteWeb interface only, with no API. The producer logs in; Claude drives the page in a browser and reads each result back.
Headless EEVEE renders. Each shot is a self-contained job with absolute paths, frames, narration timings and the style palette.
Two techniques that run without Blender and write the same frames, so caching and previews work the same way.
Turns a still photo into a depth map, so the camera can move through a real picture without inventing anything.
Assembles the film from the rendered shots, the narration, the score and the effects.
Everything renders locally. Electricity for a film costs less than €1.
| Technique | What it does | Example on this page |
|---|---|---|
kenburns | Slow moves over a still; fit: contain keeps labelled plates whole | Braník, 1913 · six-dot mosaic |
parallax | Depth relief from a depth map, with a moving camera | Prague in winter (AI image) |
map | Antique-style maps from Natural Earth and OpenStreetMap data | Boston to Calcutta, 1833 |
infographic | Counters, bars, timeline, scale with a “region” mode, statements, checklist, mass drop, scales | 90,000 people · size to scale |
card | Title, newspaper, paper and end cards | Both title cards |
grid | A grid of real cutouts appearing one by one | |
spectrum | Spectra drawn from real archive data files (FITS) | “The Cliff” |
3d | Templates: ice plough, Braník cut-away, redshift, suspects, host galaxy, gas cocoon, scattering | Braník cut-away · cocoon · suspects |
Real imagery comes first. Every source is recorded with its licence before the storyboard is approved.
| Source | Licence and credit |
|---|---|
| NASA / STScI | Generally public domain; the credit line given on the page is used |
| ESA/Webb | CC BY 4.0: full credit line on the end card, changes noted |
| ESA (esa.int) | ESA Standard Licence: documentary and educational use allowed, commercial use needs permission. Footage music is never included |
| Wikimedia Commons | Checked file by file; licence and author saved with each download |
| DAWN JWST Archive | “Released for use by anyone”; the archive and the surveys are cited |
| Natural Earth | Public domain |
| OpenStreetMap | © OpenStreetMap contributors (ODbL) |
| Scientific papers | Only CC BY versions are adapted; paywalled figures are off-limits |
Costs · measured · prices as of October 2026
Measured from the service dashboards and from the token usage of every request in the session transcripts. Claude's figures are what the same work would cost at API prices. The films were actually made on subscriptions, which cost far less. Estimates are marked as estimates.
All-in cost per film, at API prices
Claude, main session Claude, agents ElevenLabs and Leonardo (estimate, upper end shown)
The Last Ice Harvest~$64
The Little Red Dots~$101
The Ice film's numbers include building most of the pipeline. The Red Dots film added deep zoom, footage, the grid, spectra, new graphics and the cocoon templates, built by seven agents.
| The Last Ice Harvest (3:52) | The Little Red Dots (4:43) | |
|---|---|---|
| Shots | 35: 23 AI images, 2 archival photos, maps, infographics, 3D | 32: 15 real and credited, 0 AI images, graphics, spectra, 3D |
| ElevenLabs credits | ~7,500 (including auditions) | ~7,000 (narration ~2,000, score ~4,300, effects ~700) |
| … as a share of the $22 plan | ~$1.25 | ~$1.20 |
| … at API or overage rates (estimate) | ~$1.50–2.50 | ~$1.50–2.50 |
| Leonardo credits | 1,536 (24 × 64) | 0 |
| … in money (estimate) | ~$2–4 | $0 |
| Claude at Opus 5.5 API prices (measured) | $58.52 (main $22.60 + 4 agents $35.92) | $99.37 (main $33.88 + 7 agents $65.49) |
| Claude as a share of a $100/month Max plan | ~$6 (upper bound) | ~$10–11 (upper bound) |
| All-in at API prices | ~$64 | ~$101 |
Both films together: 1,246 requests and about $158 at API prices. Most of it is the model re-reading its own context at every step.
The main lever is starting each film in a fresh session. At 800k+ tokens of context, every step re-reads about $0.16.
The next levers: fewer builder agents, now that the renderers exist, and Sonnet 5.5 for research and cataloguing agents.
Nothing was charged beyond these plans.
Claude Max
$100 a month
Both films used at most ~73% of one week's limit, about $17 of plan value. Extra usage was switched off.
ElevenLabs Creator
$22 a month · 131,000 credits
Both films used ~11% of the month's credits.
Leonardo
Existing plan · ~34.6k credits
~1.5k credits used, all on the Ice film.
A rough model for the next film, built from the measured numbers above. It is an estimate, and it says where it is guessing.
| Item | Expected |
|---|---|
| ElevenLabs | 7–10k credits (~$1.50–2.50) |
| Leonardo | 0–25 images × 64 credits (~$0–4); video clips 90 credits each, a few at most |
| Claude, API-equivalent | ~$40–80 in a fresh session reusing existing renderers (more if new techniques are needed) |
| GPU and electricity | < €1 |
| Active time | ~2–3 h, plus the producer's review time |
A studio would quote roughly $15,000–45,000 for a comparable 4–5 minute documentary with research, motion graphics, some 3D, narration and an original score. Motion graphics run about $3–8k per minute and 3D about $6–18k per minute, plus weeks of schedule. A freelancer might charge $5,000–15,000.
Cost of a 4–5 minute documentary · log scale, each gridline is ten times the last
The benchmarks come from vendor blogs, not quotes. The comparison is not like for like: our figures leave out the producer's own time, the hardware and the fixed monthly plans, and a studio's price pays for human craft, revisions and a schedule.
Time · from session timestamps
Clock span runs from start to finish, pauses included. Working time counts the main session and the agents, with overlaps counted once. Agent time adds up every agent's run, so it can exceed working time when agents run in parallel.
The Last Ice Harvest The Little Red Dots
Clock span
Actual working time
Agent time, all agents added up
Case studies · 2 films
Finished films are not changed after review. The lessons go into the pipeline and the rules for the next film.
The brief: the story of natural ice harvesting, leading to Prague.
The brief said Prague was the last place ice was harvested. The research did not support that: Plzeň's brewery harvested until 1987, and harvests still run in New Hampshire, Harbin and at the Icehotel. The script says instead that Prague's Vltava harvest outlived the Atlantic trade, into the 1950s, and ended because the Slapy dam (1954) stopped the river from freezing.
Lesson: research before the script, and flag any premise the sources do not carry.
eleven_v4; ElevenLabs music and effects.
AI image“Not bad! Actually quite good.”The producer, after the first cut
The whole pipeline: the ElevenLabs client, storyboard validation, the timeline from narration timings, the Blender job runner and its first five techniques, depth parallax, the mix, and the routine that drives Leonardo in the browser.
The brief: a 5-minute film on JWST's “little red dots”, compact red objects in the early universe that may be black holes wrapped in gas. AI images kept to a minimum. In the end there were none.
| What the producer saw | What changed |
|---|---|
| Too fast to read: the “infrared, where Webb looks” label appeared in the last seconds of its shot, and a note vanished too soon | A reading-time rule for every on-screen text, and the labels moved earlier |
| Awkward wording: “4–11% of its present age” | Rewritten |
| Misleading scale: “330 ly” looked like the size of the dot | A region mode for the scale graphic: a glow with a dashed boundary, labelled as the region the light comes from |
| Abrupt cut: the cocoon animation vanished when the next screen began | It keeps moving in the background, dimmed under the title |
| Unexplained jargon: “z” | A caption: “z = redshift · the bigger z, the further back in time” |
| Credits too short | Full CC BY lines on the end card |
| Subtitles in the picture | None burned in: the deliverables are the film and an .srt file |
fit: contain, and the new caption that explains “z”.s03d · kenburns · JWST image: NASA, ESA, CSA, STScI, Dale Kocevski (Colby College) · CC BY 4.0The deepzoom, footage, grid and spectrum techniques; the region scale; titles over picture; continuity between 3D shots; snapping cuts to the beat; one score made of movements; re-splitting cached narration without new voice calls; and the AI trust rules in the validator.
Rules · enforced by the pipeline
Real material comes first. AI is only ever illustration, always marked on screen, and never used for faces of real people or for scientific data. The storyboard validator refuses shots that break these rules.
A shot marked as evidence makes a factual claim through its picture, such as “this is the Braník ice house”. Evidence must be real, or a graphic of real data. The validator refuses AI there.
No fake old photos, engravings, newspapers, maps or paper figures, and no real people's faces. AI gets its own recognisable look: painterly or diorama, never “found footage”.
A real photo travels with the prompt as an image reference, and each AI shot has a fact sheet of details that must be right and things that must not appear. Anything wrong is regenerated or cut.
Show the source first and dissolve to the reconstruction, as the real dot dissolves into the cocoon. Use cue words such as “imagine” in the narration. Keep AI shots short.
To move a real photo, use depth parallax, which invents nothing. AI video is only for atmosphere such as snow, water or fog: a few clips per film, labelled “AI-animated”.
Every AI shot gets the same slightly desaturated look and the same badge in the top left, “AI reconstruction”.
The end card lists every AI shot, and a log records each one's time, tool, prompt, reference and fact sheet. Not built yet: C2PA content credentials on the final file.
of the runtime without words: music and pictures only, the longest pause 6–10 s. After the cold open, around the reveal, between acts, and a held ending.
of shots as plain Ken Burns zooms. Prefer parallax, footage, annotated plates, diagrams and 3D.
per word: how long every caption, label or title stays before its cut. A 12-word note needs about 4–5 s.
The narrator is directed per section: calmer for wonder, more urgent for tension, warm for the reveal, quiet at the end. Auditioned first.
Loudness of the final mix, with the narration clear over the music.
Every shot names its source, licence and on-screen credit. This is the six-dot mosaic from The Little Red Dots, abridged.
- id: s03d
narration: "Yet some pour out as much light
as an entire galaxy."
technique: kenburns
archival: littlereddots.tif
camera: {move: push_in, amount: 0.04,
fit: contain, margin: 1.2}
source:
type: real
credit: "NASA, ESA, CSA, STScI,
Dale Kocevski (Colby College)"
license: NASA / CC BY 4.0 (ESA/Webb)
onscreen: "JWST image: NASA, ESA, CSA,
STScI, D. Kocevski"
overlays:
- {kind: caption, text: "z = redshift ·
the bigger z, the further back in time"}Coming next · concept only
A game designed to warn against monopolies became one of the most defended monopolies in toys. The next film follows it from Lizzie Magie's 1904 patent for The Landlord's Game, to Charles Darrow's 1935 box story, to the Anti-Monopoly lawsuit that dug up the real history.
| Beat | Time | What happens |
|---|---|---|
| 01 Cold open | 0:00–0:40 | The box story everyone knows: 1935, the Depression, an unemployed man invents a game |
| 02 Rewind to 1904 | 0:40–1:40 | Lizzie Magie, her patent, and a game meant to teach about land and rent |
| 03 The folk game | 1:40–2:40 | Copied by hand, street names written in, the rules mutating |
| 04 1935: the product | 2:40–3:30 | Darrow sells it to Parker Brothers; Magie's 1924 patent bought for $500 |
| 05 The locks | 3:30–4:15 | The patent runs out in 1952; the name becomes the fortress |
| 06 The professor | 4:15–5:15 | 1973, Ralph Anspach's Anti-Monopoly, the lawsuit, the archives |
| 07 The ruling | 5:15–6:05 | 1982: the name is ruled generic. 1984: Congress changes the test |
| 08 Today | 6:05–6:50 | Game mechanics are free to reuse; art, text and names are not |
| 09 Close | 6:50–7:10 | Back to the irony, and a held ending |
Times are from the concept draft. The facts still need the usual research pass. MONOPOLY is a trademark of Hasbro; this studio and the planned film are not affiliated with or endorsed by Hasbro.