Skip to content
Drew T edited this page Sep 29, 2026 · 3 revisions

Project Architect 3.0

Project Architect turns a Claude Code session into a small team that works from a plan. You talk to one session, the router. A planner drafts each phase with you; once you approve it, the router hands every task to a fresh expert, the expert briefs coders and retrievers, a critic judges every change to the plan, and the phase closes only when its milestone check passes. What the work produces (plans, task summaries, decisions, rules, techniques) lives in files in your repository, so every later session starts cold from them instead of replaying a transcript.

Version 3 is a rebuild for token economy: each role runs in its own small context, and a local usage ledger prices every request and shows on the statusline what the structure saved. This wiki describes version 3.11.

A new project from claude to the phase close

A new project from claude to the phase close, sped up: the router, the planners' drafts, your two approvals, experts briefing a retriever and coders, and the close.

Start here

More

  • The bench: how the agents themselves are measured, the public results, and how new models are judged (Opus 5.5 and Sonnet 5.5, with the numbers).
  • Savings the ledger can't see: the Task tools and recaps PA3 turns off, and why no figure counts what that saves.
  • Why PA3 stays serial: the parallel speedups that were studied and not taken.
  • FAQ: settings, commits, privacy, big projects, costs, models, and more.
  • History: what 2.0 cost and why 3.0 exists.

The repository: README · INSTALL.md (every switch) · CHANGES.md.

Clone this wiki locally