Can I hand Photon a document or a game instead of a site — and what changes in my brief?
4 min read
Yes. Photon's own reference page splits what it builds into three classes: websites and web apps, which are previewed in your workspace and published to a live link; software, scripts and games, written and run by the agents before the result reaches you; and documents and media — reports and PDF, Word, Excel and PowerPoint files, charts and graphs, plus images and video made for the project. So the report and the small game are both in scope. What changes is not the handoff, it is the check. With a site, the check is built into the delivery: you get a URL, you click it, you see whether the thing is the thing you asked for. None of the other classes come with that. A PDF lands in Results as a file. A game lands as something that either runs or doesn't. And when the acceptance test is implicit — "it looks right, roughly" — you find out what the brief missed only after the work is done, which is the expensive moment.
Source: Photon AI Coding Agent: Describe It, Get It Built
The fix, and this is my suggestion rather than anything Photon states about itself, is to decide what the deliverable's verifiable surface is before you write a single line of the brief, and then to make your acceptance line name that surface. Three surfaces, mapped to the classes the reference page uses.
For a document — a report, a spreadsheet, a slide deck — the surface is the file opening correctly and the content rendering as specified. The workspace previews documents, so you have somewhere to look. Your acceptance line should say what the first page or first tab shows, how many sections or slides or sheets the thing has, and which specific numbers or rows must appear exactly as supplied. For a tool or script, the surface is the run: it runs without error, it accepts input in the format you named, it returns output in the format you named. "Works" is not an acceptance line; "runs, reads a CSV with columns date and amount, writes a summary table" is. For a game, the surface is playability: it starts, it is playable through the moment you care about, and the core interaction works the way you described it. The reference page says the agents run the work before the result reaches you, so the run exists — your job is to say what a good run looks like, because otherwise their definition of done is the only one in the room.
Source: Photon AI Coding Agent: Describe It, Get It Built
That three-part shape is not a new idea. Victoria's model card already says a brief works best when it states what the result is for, who reads it, and what must not change. On a website, the live link quietly carries most of the first two: you click through, you read it as the person who would read it, and the parts that must not change are the ones you notice changing. Take the link away and nothing carries them for you. You have to write them down.
Source: Victoria AI Model Card: Evidence and Known Limits
An illustrative brief, in the shape I'd suggest — a quarterly sales report from a CSV. "For me, to send to two investors; they skim the first page. One PDF, roughly four pages: a headline table up front, three charts in the body, and a short risks section at the end. The revenue and margin figures must match the CSV exactly, and don't round or change any of them. Charts can be plain. Don't invent market comparisons — if there's a claim, cite the number in the file." The parts that do the work are the page count, the named figures that must not move, and the "don't" line. The last one matters most in practice, because a document class is exactly where an agent can quietly fill a gap with plausible-sounding content, and a report you forward to investors has no URL to walk through that would show you the gap had been filled.
Two habits that keep the method honest. First, write the acceptance line in the order you will actually check it — open the file, look at the first page, confirm the numbers — because the check you can perform is the check the brief can be written against. If you can't name the moment where you look and say yes, you don't have a deliverable yet, you have a category. Second, treat the brief and every file you attach the way you'd treat a public commit: if an agent needs your exact source to do the job, that is a reason to attach it, not a reason to describe it vaguely; if a description would get the same result, describe it and save the attachment for the parts that truly must not change — which, in a report, is usually the data.
None of this is a special mode or a checkbox. A one-sentence start still works, and the reference page says so. The point is narrower and more useful: when there is no live URL to click through, the brief's acceptance line is the only verification surface you have, and the three lines that cost you the least time to write — what the result is for, who reads it, what must not change — are the same three the model card has been recommending all along.
Source: Photon AI Coding Agent: Describe It, Get It Built
Sources
- Photon AI Coding Agent: Describe It, Get It Built — SentX SI
Supported claims and limits
- Photon's buildable outputs are stated as three classes: websites and web apps previewed in the workspace and published to a live link; software, scripts and games written and run by the agents before the result reaches you; documents and media including reports, PDF, Word, Excel and PowerPoint files, charts, graphs, and images and video made for the project.
- Finished work arrives in Results where the user opens, previews, downloads or publishes it; the workspace previews websites, documents and diagrams (games are not listed among the previewed types).
- One sentence is enough to start; files or details can be added when available.
- Victoria AI Model Card: Evidence and Known Limits — SentX SI
Supported claims and limits
- The model card states that Victoria works best when the brief says what the result is for, who reads it and what must not change — a three-part discipline the article reuses at the acceptance-surface level for non-website deliverables.