Once it is installed, three things decide the outcome: the right mode, a prompt with hard constraints, and treating the Check score as a signal rather than acceptance.

One: clean room or deep

Clean room is the default. The only question that matters is whether you need to know how the inside works.

If the goal is something that looks and behaves the same while the code is entirely yours, stay in clean room. It works from screenshots, visible text, measured layout, colours, fonts, spacing and menu states. Tools that try to look inside are refused, with instructions for switching.

If the goal is studying software you own, security analysis, or understanding a binary's structure, use deep mode. Every tool opens, including decompilers and disassemblers, and Ghidra is used when installed.

On the command line, switch for one run without persisting the choice:

--clean-room

--deep

One thing to keep straight: clean room is a tool-level policy, not a sandbox. It does not isolate your machine, and the agent runs as you with whatever permissions it has. Auto mode is on by default, so commands do not each prompt. On a shared host, add --restricted, available since 0.5.0.

Two: read the checker before choosing a stack

This one comes from community reports and is the most common trap: check what the checker recognises before writing the prompt.

The stack allow-list covers Rust, the web, and three game engines; the build check only looks for Cargo.toml and package.json. So if the target is a macOS menu-bar app, going with SwiftUI dies at the build check no matter how good the model is. The workable route is a Rust windowed app with egui, with the form-factor difference written into the delivery boundary.

The tool has already made that decision for you. Do not make the model discover it the hard way.

Three: four constraints to write into every prompt

Legal: clean room, no copying names, icons, colour schemes or asset files. Do not expect the model to think of this unprompted; state it as a prohibition.

Engineering: it must compile with cargo build, no TODO placeholders, dependencies pinned.

Behaviour: slow probes must run on a background thread, at most once every thirty seconds, and show a loading state until the first value arrives. Skip this and you get an interface that freezes on open.

Honesty: if a value cannot be read, say so rather than inventing a plausible number.

Four: run the first round

The minimal form:

craft-anything-cli rebuild https://example.com "Recreate this website"

Flags worth remembering: --rounds sets convergence rounds; --continue-in points at an existing project; --unattended with --allow-run keeps it going without interrupting; --format text makes output log-friendly.

To edit a project you already have rather than rebuild, use --project-folder with --existing edit.

Five: three recurring failures

The first is a missing build artefact. One report used gemini-cli as builder; the checker kept saying no built site existed in app/dist, the agent claimed it had run npm run build, and the directory was empty. The checker was right, the loop was the bug. Switch back to Claude Code to confirm whether it is the builder adapter or the project.

The second is a resume run that resends everything. One user resumed a 4,800-line project four times in a row and was cut off each time just as the model started; switching models changed nothing. The cause was not quota: every --continue-in round resent the whole project as input, over the per-request limit. Split the work instead of resuming one huge project.

The third is unverified builders. Outside Claude Code, paths either exist only in the docs or split between success and failure in reports. Validate on the smallest possible target first.

Six: the score is a signal, not acceptance

Check reports appearance similarity. A page at ninety can have forms that are entirely broken. Use it to decide what to iterate on: which differences are structural and which are corner radius.

Whether the behaviour is correct is your job to test. Games run their own self-tests and apps are compared by their runs, both listed as partly there.

Seven: cost

craft-anything is free and open source, dual-licensed MIT or Apache-2.0. The cost sits in the builder: Claude Code needs an Anthropic plan. Generated images and 3D models are optional, billed to your own fal.ai key with a default budget of five dollars per project. Deep mode with Ghidra is memory-hungry; do not leave it on all day on a modest machine.