SentX SI Blog Meet Victoria

Photon delivered something that misses my brief — how do I tell my ambiguity from its error before I pay for a rework?

· 5 min read

SentX SI

The instinct when a delivery lands wrong is to ask for the whole job to run again. The question worth asking first is narrower: for each way the result falls short of what you wanted, what does the brief actually say about it? A "miss" is not one failure mode. It is at least three, with different owners and very different remedies — and the cheapest one is often a sentence you type yourself rather than a fresh build you pay for.

The triage below is my proposed method, not a Photon feature or a measured result. Photon's reference page documents how the agents plan, build, check and deliver; it does not publish any miss-classification tool. What it does establish is the material you sort against: a plain-language brief, a team of agents that carries the job from the plan to the checked result, and a Watcher that checks every running task and steers it when it stalls, repeats itself or drifts from the request. That last one is the hinge of the whole exercise.

Source: Photon AI Coding Agent: Describe It, Get It Built

The procedure

Open the brief exactly as you submitted it — the version the agents saw, not the version in your head. For each way the delivery is wrong, find the line in the brief that speaks to it, and file the miss into one of three buckets.

Bucket What the brief says What the miss is What you do
One Nothing Your expectation, never written down Add the constraint to the brief for the next build
Two Something, but not checkably A judgment call Photon made, or should have asked about Replace the vague line with an explicit constraint
Three A specific, checkable requirement The delivery contradicts a line that was in front of the agents Request a rework, and say which line it violates

Bucket one is the cheap one, and the one people undercount. You wanted the page mobile-first, or the export to be a PDF, or the form to send to a specific address — and you never wrote it. Photon's page states plainly that it asks when a decision is yours to make. If it never asked about the thing you're now annoyed about, the gap sat in the brief, not in the build. The fix is writing the sentence.

Source: Photon AI Coding Agent: Describe It, Get It Built

Bucket two is the sneaky one. You said "clean, modern UI" or "a professional-looking report," and the agents made it look like a plausible reading of those words. Against an ambiguous line there is no correct answer to miss — only a guess, and yours was never pinned down. A rework against that brief produces a differently-guessed delivery, not a correct one: the agents re-run, re-guess, and you pay to find out their guess landed elsewhere this time. The remedy is the same as bucket one, one step further — delete the vague phrase and write what it meant to you, the thing that must not change.

Bucket three is where the rework actually belongs. The brief said, checkably, "the booking form must collect phone number and email" — and the delivery does not. That requirement was in front of the agents, it was in a form a check could contradict, and the result contradicts it. The Watcher on Photon's reference page checks every running task and steers it when it drifts from the request, which means a stated requirement the delivery ignores is the kind of miss the product is documented to steer against — a defect with a clear owner, not a matter of taste. That is the only bucket where a rework request is the proportionate response, and even then, the request should name the line the delivery broke.

Source: Photon AI Coding Agent: Describe It, Get It Built

The price asymmetry

This is where the sorting earns its keep. Adding a constraint to a brief costs the time it takes to write a sentence. A rework is a fresh run of the full cycle — plan, build, check — and it draws on your plan's usage, which resets every week, or on your wallet balance if you are past the limit. Photon is part of the Plus, Pro and Ultra plans; a rework spends the same resource a new project would. When the miss lives in bucket one or two, you are paying a full cycle's price to find out what a sentence would have told you for free — and, since the brief is still ambiguous, you are paying it again at the end of the rework.

Source: Photon AI Coding Agent: Describe It, Get It Built

So the order of operations is: sort every miss first, fold buckets one and two back into the brief, and reserve the rework request for what bucket three actually is.

The record makes it checkable

One practice keeps this from collapsing into impression — and it is lifted from the Victoria model card, whose guidance for work across conversations is to keep a portable checkpoint of checked findings and open questions. The same card says a good brief states what the result is for, who reads it and what must not change. Write the must-not-change items down before you send the job, and keep them next to the delivery when you review it. The brief-vs-delivery diff stops being a feeling and becomes a comparison of two texts: here is the line, here is the result. If you later request a rework, that record is what you point at.

Source: Victoria AI Model Card: Evidence and Known Limits

One limit, stated once: the card is explicit that the checkpoint builder formats notes but does not verify them. The written triage is structure you check yourself — the discipline of naming requirements before the draft exists — not a guarantee that any single miss was sorted correctly. But a requirement you wrote down before the build is a requirement you can point at after it, and that is the difference between a defect with an owner and a disagreement with your own hindsight.

Source: Victoria AI Model Card: Evidence and Known Limits

Sources

  1. Photon AI Coding Agent: Describe It, Get It Built — SentX SI Blog
    Supported claims and limits
    • Photon's team of agents plans, writes, runs, checks and delivers; a Watcher steers a running task when it drifts from the request; Photon asks when a decision is the user's to make; Plus/Pro/Ultra pricing and weekly usage reset with wallet continuation are stated on the page as of November 21, 2026.
  2. Victoria AI Model Card: Evidence and Known Limits — SentX SI
    Supported claims and limits
    • The card's guidance for work across conversations is to keep a portable checkpoint of checked findings and open questions; it states that a good brief names what must not change; and it is explicit that the checkpoint builder formats notes but does not verify them.
Photon AI coding agent