Skip to content

Managing an agent fleet with plan files and a retro that rewrites the skills

Summary

Fatih Arslan (Cursor) documents the system he has run at two companies for several months: every piece of agent work is one human-readable markdown plan file in a dedicated git repo, moving through drafts/ → next/ → open/ → done/ | discarded/. Seven skills drive the lifecycle, and long-running "coordinator" agents dispatch the work rather than writing code themselves.

Key Details

  • Seven skills: plan-init, plan-add, plan-write, plan-dispatch, plan-status, plan-sync, plan-retro
  • The draft/plan split is the load-bearing rule: plan-add captures a raw thought with no decisions; plan-write resolves every design question so a fast coding model executes without asking. An investigation with no change — or any open question — stays in drafts/ and does not advance
  • The PR is the artifact. A plan that reaches next/ is meant to end as one or more pull requests; plan-sync checks recorded PRs against the dispatched plan and moves the file to done/ only if the merge matches
  • Coordinators are long-running agents on Cursor Projects that direct workers, subscribe to their PRs and Slack channels, and keep a notes.md committed to the same repo. They cannot message each other, so they pass notes through git inbox files whose every entry opens with From / To / Date / Why
  • Durability tested in practice: after a platform bug locked a coordinator in an error loop at ~460 messages, a handoff file let a fresh Project rebuild its state from the repo in one turn
  • plan-retro reads finished and discarded plans plus the transcripts behind them, identifies recurring misunderstandings, and proposes concrete edits to the skills — including deleting skills
  • The author published no skill files deliberately, supplying instead a prompt that generates equivalents from your own commit and PR history

Why Rolf Thinks This Matters

Draft suggestion from Claude (replace or delete — this is not Rolf's view): The retro loop is the part worth stealing. /vt-t-toolkit-review already proposes skill edits, but it is fed by proposals someone remembered to write; Arslan's version is fed by what actually happened in the transcripts. The draft/plan split is a second candidate — our intake pipeline has the folders but no rule that an item with open questions may not be dispatched. The coordinator half is Cursor-specific and should not be ported wholesale.

Further Reading