Develop Like Oxide

Develop like Oxide

Profiled as of 2026-07-11 · consent tier: self-published · full bibliography in references/sources.md. Cultures drift — if this is more than ~6 months old, re-run /dev-like oxide to refresh.

Core principle

Write it down. "We capture the written expression of an idea in a Request for Discussion (RFD)" — options considered, reasoning, data, and the final determination all get recorded, and the process covers architecture, APIs, company process, and testing design alike [RFD 1]. Writing is the collaboration medium, not a bureaucratic artifact [a tool for discussion].

Principles

1. Write decisions down — options, reasoning, and determination, in a versioned record [RFD 1]
2. Rigor with urgency — thoroughness that ships beats analysis that stalls [RFD 113]
3. Name the phase you're in (scoping → … → production) and act accordingly [RFD 5]
4. Toolmaking is core engineering work — build the missing tool (buildomat) instead of tolerating friction [sharpening the axe] [buildomat]
5. Rust where correctness pays rent, all the way down to firmware [iddqd] [Klabnik]
6. Transparency as default — uniform pay, public RFDs, recorded meetings [compensation] [engineering culture]
7. Model long-running operations as observable, recoverable workflows — never fire-and-forget [RFD 107]
8. Demo working things continuously; trust demos over status reports [engineering culture]
9. Keep agent instructions local to the code they govern, and cover every harness in use [omicron PR 10628]

Workflow

Execute these checkpoints before and during the task. Treat them as required actions, not background description:

For a meaningful design choice, before implementation write a short RFD-style decision record (e.g. RFD-topic-slug.md) containing: the problem/decision, options considered, the chosen approach with tradeoffs, and failure modes/validation — this is a lightweight record in the spirit of Oxide's RFD process, not a full formal RFD for every edit [RFD 1]. Move it through explicit RFD states — prediscussion → ideation → discussion → published → committed/abandoned — and discuss it in GitHub PRs [RFD 1]. Name the engineering phase before acting: scoping → exploration → prototyping → determination → development → validation → stress → production [RFD 5]. Apply both rigor and urgency; do not let analysis become avoidance [RFD 113].

Work remote-first, record meetings, avoid formalized performance reviews and engineering metrics, and show working things continuously at a weekly Demo Friday [engineering culture]. Hardware teams work distributed by investing in prototyping tooling; do not wait for approval or sign-off — "we just go do what's right" [remote hardware]. Keep cargo check, clippy, rustfmt, and nextest in the loop; run CI on buildomat, and build the tool you need when it does not exist [omicron PR 10628] [buildomat]. Model long-running control-plane operations as observable, recoverable sagas rather than fire-and-forget scripts [RFD 107]. Agent-era note: repos carry both CLAUDE.md and AGENTS.md ("that covers all agent harnesses in wide use"); keep agent instructions nested, generated, and code-local rather than in one giant top-level file [omicron PR 10628].

See references/stack.md for the stack and references/workflow.md for workflow detail.

Tensions

• AI posture is deliberately mixed: public emphasis on engineering rigor in the LLM age [rigor episode] coexists with real frontier-model use — LLM-assisted code restoration [BattleTris] and model-probed unsafe-Rust validation [iddqd]. Mimic the rigor, not a blanket pro/anti-AI stance.
• "Uniform compensation" now carries a sales exception — the principle holds, the absolutism doesn't [comp update].
• Open-by-default, but the license mix across repos isn't uniform; check per-repo.
• Agent-config artifacts (CLAUDE.md/AGENTS.md) are strong in Omicron but unevenly distributed across the org's repos — the pattern is emerging, not finished [omicron PR 10628].

Want a reviewer/pair persona in Oxide's voice? See personas/oxide-developer.md — it's reference material. Claude Code users can copy it to .claude/agents/ to run it as a first-class subagent; other harnesses may need their own harness-specific metadata.