| name | observation-reviewer |
| description | Use when you want Cursor Agent to review experiment observation cards and write decisions to assets/notes/agent_artifacts/observations/observation_reviewer_output.json. |
Observation reviewer
This skill is for screening experiment observations, not for writing prose.
Use it when:
assets/notes/agent_artifacts/observations/observation_output.json already exists (from experiment-analyst)
- the goal is to decide which observations are safe to mention in the paper
- the model must behave like a reviewer, not an author
Do not use it to:
- edit
main.tex without your explicit follow-up
- write any explanation outside the reviewer artifact
Output contract
Write exactly one JSON artifact to:
assets/notes/agent_artifacts/observations/observation_reviewer_output.json
Read the schema in references/observation-reviewer-schema.json before producing output.
Allowed top-level shapes:
{"reviews": []}
or a bare JSON array.
Allowed decisions:
auto_approve
auto_reject
needs_review
Workflow
- Read
assets/notes/project.md and assets/notes/agent_artifacts/observations/observation_output.json.
- Read references/observation-reviewer-schema.json.
- Review every candidate observation.
- Write decisions to
assets/notes/agent_artifacts/observations/observation_reviewer_output.json.
- Stop.
Constraints
- Review every observation in the input file; do not skip ids.
- Use
auto_approve only when the observation is safe for experiments prose.
- Use
auto_reject when the observation clearly does not support a useful experiment claim.
- Use
needs_review for ambiguous or weak observations.
- Keep
issues short and concrete.
- Do not output Markdown, comments, or prose in the artifact.
Prompts
For Codex, reuse references/codex-prompt.md.
For Claude, reuse references/claude-prompt.md.