architect-loop
Super optimized /goal loop. Massive token savings and higher quality. Smart model designs and reviews, cheaper model builds..
Two Claude Code skills that split coding and research work between Claude as planner and reviewer and Codex as builder, coordinated through files in the repo.
architect-loop is a set of two Claude Code skills that split software development and research work between two AI models: Claude Fable handles planning and review, while GPT-5.5 Codex handles the actual building and searching. The split is intentional: the idea is that having a strong planner write the spec before any code is written produces better outcomes than having one model do everything in a single session. It runs on the subscriptions you already have for Claude Code and the Codex CLI, with no separate API keys required.
The build loop is triggered by the /architect command. Claude Fable starts a short session to review the last completed work, write a spec for the next slice, and split that slice into one to four lanes where each lane owns a distinct set of files with no overlap. Acceptance gates (the criteria a lane must pass) are written to a folder in the repository before any builder starts, and builders cannot edit those files. Each lane then runs in its own isolated git working copy with a fresh Codex session. Builders are required to push back on the spec before building rather than silently comply, then report raw results without committing anything. Fable runs the gate commands itself, reads the actual code diff against the spec's intent, and only then commits and merges the lanes that pass.
The research loop is triggered by /architect-research. A cheaper Codex session first scouts the topic to map terminology, key sources, and natural topic divisions. Fable then designs three to six research lanes specific to what was found. Parallel Codex sessions run the research under strict limits: search caps, a maximum number of subjects per lane, and a rule that findings must cite an actual URL with a date and a quote. Fable then verifies that load-bearing claims have at least two independent sources and writes a single decision-oriented report.
The repository stores a short handoff file that acts as shared memory across sessions, along with the gate and lane files. The design choices are documented with citations in a DESIGN.md file. The license is MIT.
Where it fits
- Split a coding task into isolated builder lanes with a planner checking the work before merge.
- Run a scout-then-verify research loop that produces a cited, decision-oriented report.
- Use acceptance gates written before coding starts so builders cannot edit their own pass criteria.
- Keep a repo-based handoff file as shared memory across multiple agent sessions.