old-coder vs three-man-team
Uncle Bob's strategy as a skill: the agent writes a SPEC you approve, then runs a gauntlet — tests, mutation, property-based, coverage, supply chain — and hands you an evidence report instead of code. — versus — A disciplined 3-agent dev process as context files — Architect plans, Builder builds the brief, Reviewer gates — running in one Claude Code session via subagents. Token-frugal by design.
Two ways to keep an agent honest without reading its code. three-man-team splits the work across Architect, Builder and Reviewer roles that gate each other; old-coder keeps one agent and makes it prove itself against a mechanical gauntlet.
| old-coder | three-man-team | |
|---|---|---|
| Stars | 667 | 944 |
| Forks | 51 | 122 |
| Language | Python | Shell |
| License | MIT | MIT |
| Last activity | 6 days ago | 2 months ago |
| Topics | skills, coding | skills, coding |
| Curated connections | 4 | 4 |
old-coder — the curator's take
The right skill when you've accepted you won't read the diff: you read two documents instead — a SPEC of concrete must and must-not examples before any code, and an EVIDENCE report with real numbers from one fresh run you can reproduce with a single command. Effort scales with risk, so a typo fix runs two checks while anything touching money, auth or concurrency runs mutation testing, property-based tests and hostile inputs. It's plain Markdown, so any instruction-following agent can use it. Read the honesty clause carefully: the gauntlet turns your spec into executable evidence, it cannot prove the spec is complete or authenticate its own checkers — the demo's own report caught an unsound mutation runner behind an earlier green state.
three-man-team — the curator's take
It's a process, not software: three markdown personas with strict handoffs (plan → brief → build → review → deploy) that target the classic solo-agent failure modes — scope drift, unrequested features, token burn. The five CLAUDE.md token rules ('is this speculative? kill the tool call') are genuinely good hygiene even outside the team. Everything runs in one Claude Code session via the Agent tool; works with anything that reads context files. NOT a framework, and not for exploratory hacking — the ceremony pays off on production work and taxes quick experiments. Know the two strings attached: a commercial 'Pro' waitlist, and Arch phones a version registry at session start to offer updates (auditable, consent-gated, but it's there).