<< All versions

Skill v1.0.0

currentLLM-judged scan90/100
sakanaai/ai-scientist-v3/edit-program
──Details
PublishedOctober 22, 2026 at 08:50 PM
Content Hashsha256:d375e68d20ef0ee6...
Git SHAdf8313889a00
──Files
Files (1 file, 2.2 KB)
SKILL.md2.2 KBactive
SKILL.md · 37 lines · 2.2 KB

version: "1.0.0" name: edit-program description: Edit a program for next experimentation.


What the skill is about

Create the next program variant safely for iterative optimization experiments.

When to use the skill

  • After reviewing the latest evaluation results (or at the very beginning).
  • When you want to test a new idea without overwriting previous attempts.

What to do

  1. Find the next unused program index <idx> from {{RESULTS_DIR}}/program_<idx>.
  2. Clone the full previous program project (program_<idx-1>) into the new program_<idx> with the trusted clone utility:
  • ${TASK_PYTHON:-$WRAPPER_PYTHON} -I -m src.investigation.program_clone --from "{{RESULTS_DIR}}/program_<idx-1>" --to "{{RESULTS_DIR}}/program_<idx>"
  • The utility creates the destination and applies the program_spec.toml copy rules without carrying evaluator artifacts.
  • For every source after program_0, the utility refuses to clone until canonical evaluation and mirrored rubric artifacts prove the previous iteration is complete.
  • Do not use cp, cp -r, cp -a, rsync, or any other recursive copy. If the trusted clone fails, stop the attempt and report the error; do not fall back to another clone method.
  1. Edit code under {{RESULTS_DIR}}/program_<idx>/src/** (and update program_spec.toml if you change entrypoint or language).
  2. Keep all older program_<idx> directories unchanged.
  3. Never edit {{TASK_DIR}}/evaluate.py or {{TASK_DIR}}/capture.py.
  4. Before creating program_<idx>, ensure previous evaluated program_<idx-1> already has:
  • metrics.json
  • correct.json
  • program_<idx-1>/rubric_status.json
  • {{RESULTS_DIR}}/rubric_status.json mirrored to that latest evaluated artifact
  1. If latest rubric status has overall_pass: true, stop; do not create another program directory.
  2. Single-attempt evaluation cap: never create/evaluate beyond program_{{MAX_PROGRAM_EVALS_PER_ATTEMPT}}.

Example usage

bash
${TASK_PYTHON:-$WRAPPER_PYTHON} -I -m src.investigation.program_clone --from "{{RESULTS_DIR}}/program_2" --to "{{RESULTS_DIR}}/program_3"
# edit {{RESULTS_DIR}}/program_3/src/** (and {{RESULTS_DIR}}/program_3/program_spec.toml if needed)
All versions