Develop like 37signals

Develop like 37signals

Profiled as of 2026-07-16 · consent tier: self-published · full bibliography in references/sources.md. Cultures drift — if this is more than ~6 months old, re-run /dev-like 37signals to refresh.

Core principle

Fixed appetite, variable scope — decide how much time a problem is worth before you decide what to build, then build to that budget instead of estimating a spec [Shape Up ch.1]. Work happens in six-week cycles with a two-week cooldown between them, and all teams share the same cadence [how we work].

Principles

1. Fixed appetite, variable scope — decide the time budget before the solution [Shape Up ch.1]
2. Six-week cycles with a two-week cooldown, no exceptions across teams [how we work]
3. No backlogs — shape and bet, don't queue [Shape Up ch.8]
4. Hand over whole projects, not task lists; done means deployed [Shape Up ch.10]
5. Circuit breaker — cancel by default rather than extend a project past its cycle [Shape Up ch.8]
6. Show progress on a hill (uphill/downhill), not with estimates or percentages [Shape Up]
7. Vanilla Rails, server-rendered, minimal dependencies — fight hard before adding a gem or a JS package [vanilla Rails stack]
8. Manual, guided exploratory QA over exhaustive test-case matrices [all about QA]
9. Coding is craft, not overhead to delegate away — pair with AI, don't hand it the keyboard [coding should be a vibe]

Workflow

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

Before the cycle, shape the problem, state the appetite, and sketch the rough solution. During cooldown, place bets at a betting table; do not pull them from a backlog — "no backlogs" is explicit doctrine [Shape Up ch.8]. When you take a bet, own the whole project, not a list of tasks, and define done as deployed [Shape Up ch.10]. Track progress with hill charts (uphill = unsolved, downhill = just execution), not percent-complete or burndown [Shape Up]. Use a circuit breaker: cancel projects that don't ship within their cycle by default rather than auto-extending them [Shape Up ch.8]. During cooldown, fix bugs, plan cycles, and make the next bets [how we work]. As a two-person QA team, run manual, guided exploratory testing against ~100-item per-product checklists (not exhaustive test-case matrices), then run accessibility passes with screen readers and the home-grown BackstopJS visual-regression suite [all about QA].

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

Tensions

• Manual QA is a deliberate, celebrated practice — two people running guided exploratory passes and checklists, not automated end-to-end suites [all about QA]. That's a real bet against TDD/heavy-automation dogma; it works because their QA staff are experienced generalists embedded in the Shape Up cycle, not because automated testing doesn't matter — Minitest and fixtures are still standard practice in the stack [vanilla Rails stack]. Don't read "manual QA works for them" as "skip your test suite."
• AI posture is genuinely split by voice and by time. DHH is on record wanting to keep his hands on the keyboard and treating full vibe-coding handoff with suspicion [coding should be a vibe], while a year later the company shipped a fully agent-accessible Basecamp — revamped API, CLI, and a bundled skill for agent harnesses [agent-accessible]. Mimic the trajectory (cautious about full autonomy in the editor, enthusiastic about making the product agent-operable), not a single fixed stance.
• Shape Up's six-week/two-week cadence assumes a company that can hold "all teams operate on the same 6-week cadence" [how we work] — a ~70-person, cross-functional, deeply synced org. It doesn't obviously survive contact with much larger or less disciplined organizations without adaptation, which the book itself acknowledges (Appendix: "Adjust to Your Size") [Shape Up].

Want a reviewer/pair persona in 37signals's voice? See personas/37signals-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.