Nothing to hand? Load the — eight microservices, Kafka and Kubernetes for a two-person team with no load numbers — or the , where the correct verdict is ready-to-build and the useful output is what to decide next.
Paste the brief — the prescan is free
No upload, no AI: the prescan reads the brief in your browser and lists what it mechanically found. The signals first — every technology and load number the brief names — then the gaps: no scale numbers, no latency target, no datastore, no auth story, no failure handling, no consistency requirements, sensitive data without stated controls, no team constraints. Each group explains why it matters. This part costs nothing and happens while you type.
The AI reviews it — this is the metered part
A senior backend architect's pass: a readiness posture with the single most important change named, the component plan with each component's data ownership and role (proposed additions marked as proposed), and prioritized findings across boundaries, API design, data, scalability, resilience, security, operability, cost and simplicity — each with the problem, the concrete impact, the fix and a liftable design artifact. The capacity math uses your numbers, not Google's. Every prescan flag is confirmed or explicitly set aside. Pricing is honest: a worst-case amount is reserved before the run and only what the run actually uses is charged — the meter next to the button shows both.
Decide, export, re-run
Quick wins for the decisions worth making today, focus areas linked to the findings they order, and the review as the file you were actually asked for: an architecture decision record, a tickable work list, the findings as CSV or as a Jira/Linear import, plus Markdown and JSON. Answer the open questions in place and they are appended to your context for the next run. Review history follows your account when you are signed in, and the second review is diffed against the first — new, carried, re-prioritised and resolved.
Derived from the @sickn33/backend-architect skill (MIT license).
What does Arch Desk actually give me back?
One structured review: a readiness posture (ready-to-build, refine-first or needs-discovery) with the single most important change named, a component plan listing each component's type, name, the data it owns and its role — with anything the reviewer is proposing rather than reading out of your brief marked proposed: — between roughly five and twelve prioritised findings across boundaries, API design, data, scalability, resilience, security, operability, cost and simplicity, each with the problem, the production impact, a concrete fix and often a liftable design artifact, then a reconciliation of the free prescan, quick wins and the focus areas to work through first.
What does it cost, and what is free?
The prescan is free and runs entirely in your browser — no upload, no AI — and so is every export, the diff between two reviews and the worked examples, which replay a saved run rather than billing a new one. Only the AI review itself is metered. Before each run a worst-case amount is reserved; you are then charged only what the run actually used, which is usually far less. The meter beside the run button shows the reserved figure, and the run button is disabled with the shortfall spelled out if your balance cannot cover the model's minimum.
Does it invent components my brief never mentioned?
It may propose them, but it has to say so. Any component the reviewer adds must have its role prefixed with proposed:, and every finding has to cite either something your brief states or something material it leaves out — a gap finding names its component as (not specified in the brief). Where the brief is silent on something that changes the design, the reviewer states the assumption it made under Assumptions and, if the answer would change the ranking, adds it to Open questions. Capacity math uses the numbers in your brief, not generic ones.
What is the free prescan checking for?
Two things, by keyword and number matching only. Signals: the technologies and load numbers your brief actually names. Gaps: a fixed deterministic checklist of eight — no load or volume numbers, no latency or responsiveness target, no datastore named, no authentication or actor model, no failure handling, no consistency or transaction requirements, sensitive data mentioned with no stated controls, and no team or operating constraints. Every gap it flags must be confirmed or explicitly set aside by the review, and the Prescan coverage table shows you flag by flag whether that happened.
I revised the brief. Is it worth paying for a second review?
The prescan answers that for free before you spend anything: once you have run once, it reports how many gap flags you have closed since that run, how many are still open and how many are new, live as you type. After a second review the findings are diffed against the previous one — each finding is marked new or carried over, with re-prioritisations called out, and findings that no longer appear are listed as resolved. Matching is done on category and component rather than finding id, because ids are numbered per review and do not survive between runs.
Where does my brief go, and is my history kept?
Nothing leaves your browser until you press the run button — dropped files are read locally and the prescan never makes a network call. When you run, the brief is sent to the model; anything over 60,000 characters is clipped from the middle with both ends kept and the cut announced in the text itself. Past reviews are stored against your account when you are signed in and mirrored into this browser for instant load; running as a guest keeps history in this browser only, and guest history does not migrate when you later sign in.