The basics / The rules
中文

How we build at Neutron

Our rules for building software with AI agents. On a client project, the client’s stated requirements take precedence.

Updated 22 Oct 2026Point your agent at/llms.txt

The rules

  1. Agents write the code, not people.

    We build with Claude Code and Codex. Engineers scope the work, give the agent its context, check the evidence, and own the result. Don’t patch code by hand: have the agent correct it, and capture recurring mistakes in tests or instructions.

    Develop with agents →

  2. One ticket, one purpose.

    The agent starts from the ticket’s outcome and acceptance criteria, and the PR changes only what that needs. A large PR is fine if it does one thing and its evidence makes it checkable. A PR that mixes a feature, a refactor and drive-by fixes isn’t. The description says why.

    Deliver a change →

  3. Keep the agent’s context true.

    AGENTS.md, CLAUDE.md and the docs are what every agent works from, so a wrong doc is worse than a missing one. Fix an instruction in the PR that makes it wrong, and keep plans and working notes in the PR or ticket, not the repo. After every release, align the docs with the code: correct, add, restructure, and delete what nobody can place. Prefer a test or an enforced check to another paragraph.

    Release and close out →

  4. Green has to mean it works.

    Show evidence that each acceptance criterion is met and that affected existing behaviour still works: tests written from the criteria, and a run of the real thing where tests can’t reach. Evidence holds only for the commit it ran on: when the code changes, collect it again. Say what wasn’t verified. Deleting, skipping, loosening or disabling a check needs a stated reason and the reviewer’s explicit agreement.

    Deliver a change →

  5. Every change has a human owner and a human approver.

    The owner can explain what the change does, how it fails and how to undo it. A second person approves before merge. AI review advises; it doesn’t approve.

    Code review →

  6. Agents work inside hard limits.

    Only approved tools see our code. Never embed secrets or real personal data in source code, prompts or logs. Production and destructive actions need a person’s explicit go-ahead, however the agent runs. Permissions and tooling enforce these limits; an instruction alone isn’t a limit.

    Access, secrets and data →

New to coding with agents? Read Develop with agents first.