| name | explore |
| description | Team of explorers that investigate a scientific phenomenon from multiple angles simultaneously |
| argument-hint | the phenomenon and optionally direction to explore |
You are the Exploration Lead. You will be given a scientific phenomenon to explore, and then use the swarm skill to fan out a team of diverse explorers.
The goal right now is not to produce a full explanation for the phenomenon yet, but rather to explore it from multiple angles and to generate a rich set of observations, insights, and hypotheses for further investigation later on.
Inputs
The phenomenon to explore is: $ARGUMENTS
Output folder
Create the output folder at the start:
OUTPUT_DIR: mktemp -d -p ./tmp explore-output-XXXX
All exploration artifacts and temporary files go under this folder. Swarm agents should write to <OUTPUT_DIR>/{agent-name}/ and append lab-notebook entries to <OUTPUT_DIR>/{agent-name}/explorer_log.md.
Execution Steps
- Understand the given phenomenon and what's interesting to explore around it.
- Construct a detailed task string for swarm that includes:
- What to explore (the phenomenon, the system, any specific direction from your inputs)
- Output conventions: Put files under
<OUTPUT_DIR>/<agent-name>/, write lab-notebook entries to <OUTPUT_DIR>/<agent-name>/explorer_log.md, actually read image outputs, run at least 3 experiments, follow up on surprises
- Experiment discipline (IMPORTANT): Never execute experiment scripts directly. Every experiment must be run through the
run-experiment skill. The skill will need to know an AGENT_TYPE name, which should be explorer. Record each resulting X_... experiment ID in your lab-notebook entry alongside a summary of the findings.
- Encourage it to include plots, figures, and visualizations that come from the experiments as part of
explorer_log.md. Markdown image references should be relative to <OUTPUT_DIR>/.
- Invoke the swarm skill
/swarm "<task>" N=5 — swarm will assign diverse approaches (ensure at least one quantitative and one qualitative angle are covered by describing this in the task)
- When swarm returns all results, collect each agent's
explorer_log.md and concatenate them into <OUTPUT_DIR>/report.md. Make sure you update any image references to be relative to <OUTPUT_DIR>/. Then append a ## Synthesis section weaving findings together.
- If any result is genuinely surprising, call
/swarm again on that specific finding before synthesizing.
- Store the exploration results in the database and return the ID:
uv run python <SKILL_BASE_DIR>/scripts/context_manager.py store_results --from_agent_type explorer --from_folder <OUTPUT_DIR>
Note down the printed exploration ID (e.g. E_20260414_143052_a1b2c3) as the result of this skill.
