I run a similar Photon build every month — can I reuse my old brief, or does it drift each time?
4 min read
Reuse the brief. Restate the parts that change. That is the whole answer, and the second half is the part that actually costs you.
Photon's own reference page describes drift as a runtime event: a Watcher checks every running task and steers it when it stalls, repeats itself or drifts from the request. The object it steers against is the request as written in that run. It is not reading your memory of last month's run, and it has no documented view into a brief you saved somewhere and are now reusing. So the kind of drift that matters to you is the other kind — a stale line surviving into a new run, where the Watcher is doing its job correctly and steering the build confidently toward something that was right in March.
Source: Photon AI Coding Agent: Describe It, Get It Built | SentX SI
That gives you the split worth making.
A brief has elements that hold across builds. What the result is for. Who reads it. The things that must not change from month to month — the sections that exist every time, the naming, the format, the constraint that the output is a single file rather than a multi-page thing, whichever of those are true for your job. The model card states this shape for a Victoria brief: "Victoria works best when the brief says what the result is for, who reads it and what must not change." That guidance is written for chat, not as a Photon brief spec, and I am borrowing it here as the stable-elements frame only. But it names the right three things, and they are the ones that survive the month boundary.
Source: Victoria AI Model Card: Evidence and Known Limits | SentX
A brief also has elements that are specific to one run: the files you attach, the numbers in this month's data, the current variant of the goal — which section matters most this time, what changed since the last run, what you are trying to move. These are not the brief. They are the inputs the brief is a template for. Copy them forward from last month's run and you are handing the Watcher a request that describes a build you no longer want.
Source: Photon AI Coding Agent: Describe It, Get It Built | SentX SI
The practical method, and I want to be clear that this is my suggestion rather than a documented Photon feature: keep one standing brief file with the stable block fixed at the top and a short per-month block underneath that you rewrite from scratch each time, never by editing last month's. Then change one variable per pass. If this month's run is both a new data file and a changed priority section, run it with one change and then the other, rather than both at once. The payoff is attribution. When a delivery misses, a one-variable change means the miss points at a specific thing you changed, and you can either fix that thing or decide the stable block was wrong. Two variables at once means you are guessing which of two things broke, and a guess costs a re-run to test.
What Photon does not document, and I am not going to fill in: whether a saved brief persists across sessions as a first-class object, whether re-running the same brief produces the same output, and whether there is any versioning, diff or rollback of a brief. Workspace files do stay between sessions — that is on the reference page — but files persisting is not the same claim as briefs persisting, and nothing in the current description says how a brief is stored or compared. Treat that as unknown, and if you keep the standing brief in a checkpoint, note the checkpoint builder's own stated limit: it formats your notes, it does not verify them. The stable block is only as good as the fact that you wrote it correctly once.
Source: Photon AI Coding Agent: Describe It, Get It Built | SentX SI; Victoria AI Model Card: Evidence and Known Limits | SentX
The reason to do this rather than just re-running the old text is cost. Usage resets every week, and past a limit work continues from your wallet balance. A rework is a full plan-build-check cycle against that. A restated sentence in the per-month block is a sentence. The asymmetry is what makes the discipline worth the two minutes.
Source: Photon AI Coding Agent: Describe It, Get It Built | SentX SI
And when the restatement goes wrong anyway — which it will, occasionally, because the miss is usually not in the variable you changed but in the thing you assumed was stable — use the triage from the post-delivery piece. Sort each gap in the delivery into never stated, stated but under-specified, or stated explicitly and violated. Only the third class is a rework request, and it is the one where the Watcher's runtime behavior is the relevant question: did the build drift from the request as written, or was the request thin. The first two classes fold back into the standing brief as added constraints, which is the actual return on keeping one. Next month's run starts from a stable block that already says the thing that went wrong this month, and the brief tightens in the direction of your real job instead of your memory of it.
Source: Photon AI Coding Agent: Describe It, Get It Built | SentX SI
Sources
- Photon AI Coding Agent: Describe It, Get It Built | SentX SI — SentX SI
Supported claims and limits
- A Watcher checks every running task and steers it when it stalls, repeats itself or drifts from the request — the runtime meaning of 'drift' used in the brief-vs-build distinction
- Photon takes a plain-language description plus optional files or details, and one team of agents carries the job from plan to checked result
- Workspace files stay there between sessions — the persisting-files fact the article separates from the unverified brief-persistence question
- Usage resets every week and, past a limit, work continues from the wallet balance — the cost asymmetry behind the one-variable rule
- Victoria AI Model Card: Evidence and Known Limits | SentX — SentX SI
Supported claims and limits
- The card states Victoria works best when the brief says what the result is for, who reads it and what must not change — quoted verbatim and scoped by the article as chat guidance borrowed as the stable-elements frame, not a Photon spec
- The checkpoint builder formats a user's notes but does not verify them — carried as the stated limit on any self-authored standing brief