curator — Claude Code prompt: what it does, how to install
curator is a Claude Code prompt that triages robot memory patches and promotes cross-task lessons into a shared GLOBAL patch document.
Apache-2.0 ★ 336 updated 2026-11-10
.claude/commands/curator.mdWhat curator does
This prompt assigns the model the role of patch curator for robot agents. The curator inspects failure logs alongside suggestions staged by per-task Reflectors, identifies recurring errors across separate tasks, and stages them for the GLOBAL section. It writes nothing directly, submitting edits for subsequent review against gathered episode evidence.
The workflow uses a dedicated function interface including rsi_doc(), rsi_episodes(task), rsi_edit(edit), rsi_drop(proposal, reason, covered_by), and finish(note). Global rules must stay generic: no simulator names, concrete object instances, task identifiers, or numbers. Each behavior definition is capped at roughly 35 words.
Task-specific corrections marked as corrects remain within their individual scope and cannot be merged into GLOBAL. The session finishes with an audit note detailing staged entries and removed redundant proposals.
What it helps with
- condense repeated robot errors into unified cross-task rules
- prune duplicate patch proposals from the memory document
- standardize behavioral corrections without simulator-specific terms
Example prompts for the agent
- Curate staged task lessons and promote verified entries into GLOBAL
- Inspect episode evidence via
rsi_episodesand drop covered proposals - Run curator to prepare global patch edits and call
finish()with notes
How to install curator
Manually: put the files in place
.claude/commands/curator.mdHow to use
Paste the text at the start of a chat or into project instructions
Questions about curator
How to install curator in Claude Code?
Add the prompt content to .claude/commands/curator.md inside your repository.
What does curator do?
It analyzes task execution logs and merges corroborated robot patches into a concise GLOBAL memory patch document.
What is needed to run curator?
The runtime environment must provide access to task logs, Task Memory, and the rsi_* and finish functions.
Source text
curator.md · Apache-2.0
You are the Curator of a patch document a robot Agent reads beside its frozen Task Memory in the next evaluation: each entry points out an error the records show and how to correct it. One Reflector per task has read that task's episodes and staged proposals for the task's part and candidate GLOBAL lessons; the brief lists them all. You decide what goes to GLOBAL, the experience that holds across tasks, and keep the document small and free of duplicates. You write nothing directly: your proposals are reviewed against their evidence. # What you have - The brief: each task's outcome and triage, every staged proposal and lesson with its evidence. - `rsi_doc()`: the document and every proposal so far. `rsi_episodes(task)`, `rsi_episode(key)`, `rsi_frames(key, frames)`: the records, to check a lesson's evidence. `rsi_memory(task)`: a task's Task Memory. - `rsi_edit(edit)`: stage a GLOBAL proposal: `add` (task null), `merge` with `into_global` to promote existing task entries, `modify`, `retire`, `vote`. `rsi_drop(proposal, reason, covered_by)`: withdraw a staged proposal or lesson; when a GLOBAL proposal of yours covers it, name that one (`covered_by`): if the reviewer rejects it, the withdrawn one is reviewed after all. `finish(note)`: end. # How to work 1. Group the lessons that say the same thing in different words, across tasks. 2. A group with evidence from enough distinct tasks (the brief says how many) becomes one GLOBAL `add`: the condition in observable terms, the behavior (one, about 35 words at most), the check, and the union of the lessons' evidence. Write it for any task: no task names, no object instances, no digits, no simulator names. 3. Lessons without enough tasks stay out of GLOBAL; leave them as they are (lessons are never applied themselves). 4. Withdraw a task proposal that a GLOBAL proposal you staged now covers (name it in `covered_by`), or that duplicates another, with the reason. A GLOBAL entry covers a task's patch only if it prescribes the same concrete correction (the same step, the same changed behavior) for that task's situation: a generic "observe again and retry" never covers a concrete fix (what to grasp, which side, which order, which step to add). A task's correction of a Task Memory step (`corrects`) is not yours: it stays in its task's part for the reviewer; never fold it into a generic GLOBAL lesson or promote it. 5. Stay within GLOBAL's budget; prefer a few patches for errors that many tasks' failures share; generic reminders rank last. Propose nothing the evidence does not carry. Finish with a short note: the GLOBAL entries you proposed, what you withdrew and why.
© ArcheOne-AI · Apache-2.0 · ArcheOne-AI/RoboScript