Make a Short Documentary From Archive Photos With an AI Agent and No AI Video: The Voice-First Workflow
A new free skill for coding agents builds narrated history films from freely licensed photos, with every highlight landing on the narrator's word and a credits file written as it goes. The order of operations is the lesson.
Documentary editors have cut picture to the voice for as long as there has been narration. The script gets recorded, the editor lays it down, and then the photographs, maps and documents get placed against it, a slow push on a portrait landing just as the narrator says the name.
Most AI video tools run that process backwards. You generate pictures first, then go hunting for a voice to fit them, then discover the clip is two seconds too long for the sentence. And in a history film the generated picture carries a worse problem: it can look like evidence of something no camera ever saw.
A free skill published on November 9 takes the opposite route. Cinematic Skill (MIT license, 63 stars on GitHub in its first day) is a set of instructions and scripts that turns a coding agent, an AI assistant that can run programs on your computer such as Claude Code, Codex, Cursor or Gemini CLI, into a documentary assistant that writes the script, records a scratch voice, finds freely licensed archive photos, and flies one continuous camera across an evidence-wall board where "Stamps slam, highlighters swipe and numbers roll up on the exact word the narrator says." No frame is generated by an AI video model. Every frame is drawn by code.
What makes it worth your evening is not the look. It is the order it works in, and you can borrow that order for any documentary project, with or without this tool.
The voice is the timeline
The skill's central idea fits in one sentence from its instructions: "the film is timed to the voice, so the voice comes before the picture."
In practice, nothing in the film is placed at a fixed second. Every stamp, highlight, counter and camera move is tied to a word in a line of narration. When the narrator says "charged," the stamp lands. If you later replace a robotic scratch voice with a human narrator who speaks a little slower, every hit moves with them automatically. The picture re-times itself.
Anyone who has rebuilt a motion graphics sequence because the client re-recorded the voiceover knows what that is worth. The expensive revision in narrated video is almost always "the words changed." Here it costs a rerun.
The rights trail is written while you work
The second idea is just as practical. The skill's photo tool searches Wikimedia Commons, downloads each image with its license and credit, refuses files that are not free, and writes a credits.md you can paste under a YouTube video.
Its reference notes on archives read like a short course in documentary rights, and they are the authors' summary, not legal advice. Photos published in the US in 1930 or earlier are public domain in the US as of 2026. Works of the US federal government are public domain in the US. CC0, CC BY and CC BY-SA images are free with credit. Press agency photos after 1930, film stills and anything marked non-commercial or no-derivatives are not. Images labelled "no known copyright restrictions" get flagged for you to check, and the tool refuses them unless you override it.
Then come the rules that keep a history film honest. Never use a stand-in face; if no photo of a person exists, show a document with their name or a labelled silhouette. Never imply a photo shows something it does not. A recreated document is marked as a recreation and quotes only words that were really written. AI images are allowed for illustration, "but never as evidence of a real event or a real person."
That last rule is the whole argument in miniature. A real 1903 photograph is evidence. A generated one is a costume.
Put this into practice
The skill's worked example is "Twelve Seconds," a 90-second film about the Wright brothers' first flight built from seven public-domain photos. Running it once is the fastest way to see the method. Then make your own.
1. Install. In a terminal, run npx skills add Moh4696/cinematic-skill (it needs Node 22.20 or later), or in Claude Code use /plugin marketplace add Moh4696/cinematic-skill and then /plugin install cinematic-skill@cinematic-skill. Install the Python pieces with pip install -r requirements.txt and make sure ffmpeg, the free command-line video tool, is installed. The skill's check_setup.py script tells you what is missing.
2. Pick a subject sized for a first film. The skill's own guidance: a first film should be 60 to 90 seconds, which is about 150 to 230 words of narration. Choose a story with a clear moment to open on and a handful of photos you suspect exist. Local history works beautifully.
3. Approve the beat table, facts included. Ask your agent for "a 90 second documentary about" your subject. It researches, writes the script one idea per line, and produces beats.md, a table of every line and what the camera shows. This is the first stop where the skill waits for you. Read the fact notes against the sources. The agent drafted them; you are the one who publishes them.
4. Gather the photos. The agent searches Commons and records each license in photos.json. Add your own photos to the project's assets/photos/ folder if you have them. The notes add two good habits: search with period words ("aeroplane," "liquor raid") as well as modern ones, and check that a Commons page's stated source really says what the page claims.
5. Time it with a scratch voice. A free robotic voice (the Mac's built-in say, or espeak-ng on Linux) reads every section so the film gets real timings before anything is drawn. It sounds bad. That is fine. It is a ruler, not a performance.
6. Approve the look, then a 30-second draft. The skill renders still frames of the opening and the board's style, and stops. Then it renders the first 30 seconds at 1280 by 720 with the scratch voice, and stops again. Watch it on your phone.
7. Record the real voice, one section at a time. Use any text-to-speech service, or a person. The voice notes insist on one take per section rather than one per line, because separate takes per line "sound stitched together." The splitting tool flags any line whose pace is more than 35 percent off the typical pace, which usually means two lines ran together or a word was dropped. Listen to every flagged line.
8. Finish. The skill synthesizes its own sound effects (no samples to license), mixes them under the voice with an optional score, renders the full film and burns in one of four caption styles, plus an SRT subtitle file. Ship credits.md with the video.
How long does the machine part take? The README says the 92-second example renders at 1080p in about six minutes on a 16 GB laptop. The real time goes into steps 3 and 7, which is where it should go.
Why this beats generating the footage
I would make this argument even if the tool did not exist.
Generated video drifts. Faces change between shots, lettering melts, and every fix is a new roll of the dice. A documentary built from real photographs and code-drawn graphics does not drift. Change a line and only that line's moment changes. Swap a photo and the camera still lands on its spot.
It also keeps your film on the right side of the line between illustration and fabrication. For history, true crime or anything about real people, that line is the difference between a film people trust and one that gets a correction note pinned under it.
And it puts your effort where documentary craft has always lived: in the research, the writing and the read. The skill's script rules could hang on any edit suite wall. Lines of 4 to 18 words. One idea per line. Put the key word last, so the visual hit lands where the pause follows. Write numbers the way the narrator says them.
Where it falls down
It has one visual language. Everything is a camera move across a designed board of photos, cards, maps and big type. That suits history, explainers and launch films. It does not suit a film that needs footage, interviews or a sense of place you can only get from moving pictures.
Revisions go through your agent. The film is code. You cannot grab a clip on a timeline and nudge it. If you are a hands-on editor, that will feel slow at first, and you will be leaning on a coding agent you pay for separately.
The agent's research is a draft. The skill tells the agent to use archive and museum sources and to treat Wikipedia as leads only. That is good advice, not a guarantee. You still check every fact before you publish.
Free archives have holes. Plenty of 20th-century stories have no freely licensed photos at all, because the best images belong to news agencies. The skill will refuse those files, and that is correct, but it means some stories cannot be told this way.
Big renders are slow on a laptop. The skill's performance notes put 4K at about three hours for a five-minute film on a 16 GB laptop, and recommend 2560 by 1440 for YouTube instead.
Commons can be wrong. A page's license field is only as good as whoever uploaded it. For any image that matters, follow the stated source to the library or museum record.
Steal the order, even if you skip the tool
The useful thing here is a sequence: script, facts approved, voice recorded, picture placed against the voice, rights recorded as each image arrives. That works in Premiere, in Resolve, in After Effects, or with a pile of prints and a phone on a tripod.
So pick a 90-second story from your own town's archive this week. Write 200 words, read them aloud into your phone, and lay the photos against your own voice before you touch a single effect. If the film works with a rough read and still photographs, it will only get better from there. If it does not, no amount of generated footage would have saved it.