Should Claude Build a Whole Feature at Once?
Build one layer at a time when you plan to read and keep the code. Let Claude run end to end when the scope is clear, the change is small and a test can judge the result. The choice depends on how much you need to review, not on what Claude can manage.
We're building [feature] in layers: [data], then [logic], then [API], then [UI]. Read [paths to related code] first. Do only the first layer now: [the schema and the functions that read and write it]. Add tests for it and run [test command]. When they pass, show me the command and its output, and list the function signatures the next layer will call. Then stop. Don't start the next layer until I say so.
- 1Datatests, commit
- 2Logictests, commit
- 3APItests, commit
- 4UIrun the app, commit
Try it now
- 1Send the prompt in Claude Code, naming your real layers and test command.
- 2When it stops, run
/diffto read the change, then commit it yourself. - 3Reply with the next layer, for example
Now do the logic layer against those signatures. Same rules. - 4For the UI layer, run
/verify, which has Claude build and run the app and look at the result. It worked if every layer has its own commit and test output you actually read.
Use it when
A feature that crosses layers in a framework you can't check by eye, and code you'll maintain. It also helps when the lower layers need your judgment, such as a schema you'll be stuck with.
Skip it when
Small, well-scoped work with a test that can judge it. Anthropic's docs say current models usually carry on through an agreed task, and that a check to pass is what's still worth adding, so stopping at every layer just slows you down there. It's also overkill for boilerplate you could regenerate cheaply.
Why it works
Capacity isn't the limit. Anthropic says Opus 5.5 is strongest on multistep work in a real repository, such as carrying a change through a large code base until its tests pass. The limit is you. Anthropic's help center makes the cost point directly: with a wrong 400-line diff, you pay once to generate it and again to redo it. Layers keep each diff short enough to read, and an error shows up in the layer that caused it. Each commit also gives you a point to return to that Claude can't lose. This split is a rule of thumb, not a vendor figure.
Example
For a comments feature, layer one is a comments table with createComment and listComments, tested with something like pnpm vitest run tests/comments.test.ts. You read that diff and commit it. Layer two adds a POST /posts/:id/comments endpoint against those two signatures. If the UI misbehaves later, you can look at the UI layer first, because the layers beneath it each passed their own tests.
Going further
Have Claude write each layer's signatures to a short file such as LAYERS.md, then start each layer by reading it. Anthropic's hooks guide warns that compaction summarizes the conversation and can lose important details, and a file survives that. Also, /rewind won't save you from everything. It restores files that Claude edited, not a database migration it ran. Commit after the data layer, and keep a way to roll that migration back. The migration advice is a rule of thumb, not a vendor figure.
Common mistakes
- Treating
/rewindas your history. Anthropic's docs say checkpoints sit alongside git rather than replacing it, a file only a Bash command changed isn't restored, and snapshots are kept for the 100 most recent checkpoints. - Relying on the words
stop and waitas a hard gate. A prompt is still Claude's call. If you need edits blocked, use plan mode, where they stay blocked until you approve. - Skipping the check between layers. An unverified layer means the next one builds on a guess.
- Not asking for the signatures between layers. The next layer then works from Claude's memory of the interface instead of the written one.
Sources
- Best practices for Claude Code
- Prompting Claude Opus 5.5, Claude Platform Docs
- Models, usage, and limits in Claude Code, Claude Help Center
- Checkpointing, Claude Code Docs
- Automate actions with hooks, Claude Code Docs
- Choose a permission mode, Claude Code Docs
- Commands, Claude Code Docs
Checked 6 November 2026. Products change often, so confirm details with the vendor.