The Legendary Team – A .Claude Engineering Squad That Actually Works

I started my serious AI journey in January 2025. By March I was already frustrated with Claude Code.

It’s brilliant at writing code, but it forgets everything between sessions, drifts off spec, and occasionally hijacks the conversation with its own ideas. Fixing the same schema three times in one day was driving me insane.

So I built a very small, very boring wrapper that turns Claude into something I can actually trust.

It’s called Legendary Team – one Python script you run once in any folder.

What it actually does (no hype):

• Boots a fixed team structure every session (chief + 3–5 specialists)

• Forces OpenSpec format for every task (no more “I’ll just write the code” rebellion)

• Remembers everything between sessions using a simple JSON memory file

• Catches drift instantly (e.g., missing 982-line Prisma schema flagged in <2 seconds)

• Runs open-source PR-Agent locally for automated code review (zero tokens)

• Requires explicit human sign-off before any commit

That’s it.

No magic, no “most advanced ever”, no 100-agent swarms. Just a handful of boring rules that make Claude behave like a reliable senior dev who never forgets and never ships broken code.

Real numbers from the last 6 weeks of daily use:

MetricBefore (plain Claude)After Legendary TeamAverage task time2–3 days3–6 hoursSessions ruined by drift~40 %0 %API cost per refactor$15–$30$0 (local fallback)Code needing manual fix2–3 hours<15 minutes

Repo (everything is public, everything works today): https://github.com/RegardV/LegendaryTeam_For_Claude

It’s deliberately tiny – <400 lines – because that’s all you actually need.

If you’re tired of babysitting Claude, clone it, run the script, and see for yourself.

No onboarding, no tokens, no cloud. Just a dev team that shows up every morning and remembers what you said yesterday.

That’s all I set out to build – and that’s exactly what it does.

Regard Pretoria, December 2025