Skip to content

Agents

doekit is hybrid: the same engine serves people and LLM agents. Agents reach it two ways, sharing one contract — facts from doekit, judgment from the agent:

  • Skill — the portable experiment-designer skill teaches the DoE loop (brief → recommend → evaluate → lab → ingest → analyze → interpret → decide → next) without inventing metrics. Below.
  • MCP server — the same loop as callable tools (recommend / evaluate / propose_and_decide) over the Model Context Protocol.

Skill package (copy these two files)

File Role
SKILL.md Workflow, gates, reply template. Keep filename SKILL.md.
reference.md Self-contained API cheat sheet

Do not require this index page inside the skills folder — only SKILL.md + reference.md. The package is self-contained.

Contract: agent owns process context and decisions; doekit owns rankings, efficiencies, fits, and reports. Always call the library and read to_dict() / summaries.

Install

Cursor

SKILL.md + reference.md  →  .cursor/skills/doekit-experiment-designer/
  • Project: .cursor/skills/doekit-experiment-designer/
  • Personal: ~/.cursor/skills/doekit-experiment-designer/
  • Never use ~/.cursor/skills-cursor/ (reserved).

Trigger with experiment / DoE / doekit questions, or @doekit-experiment-designer.

Claude

SKILL.md + reference.md  →  .claude/skills/doekit-experiment-designer/

VS Code / Copilot

Point custom instructions at docs/agents/SKILL.md and docs/agents/reference.md, or @-mention them in chat.

Session hygiene

  • No secrets in factor names, metadata, or reports.
  • Prefer ed.experiment(...) / Experiment.to_dict() for handoff; export run sheets with exp.export_csv.
  • For multi-session research, persist with ed.project(name) → wave → automatic-conclusions/conclusions.json (read gates; do not invent metrics).
  • Treat lab responses in HTML reports as sensitive when required.

Traceable workspace

experiments/experiment_project_<slug>/
  PROJECT.json
  waves/wave_001/
    doe-configuration/   # INPUT: experiment.json, design.json, thresholds.json
    data/                # run_sheet.csv (+ responses.csv after lab)
    results/             # evaluation.json, fit.json, next_runs.json
    reports/             # optional HTML
    automatic-conclusions/  # conclusions.json + .md (LLM/agent/human)
    metadata/            # provenance + checksums
    assets/              # researcher auxiliaries
  • Wave = one DoE cycle (not a lab row). Lab row ids stay in run_sheet.csv as run_id.
  • Agents should load conclusions.json (gate_board, rules, facts) and only paraphrase those strings.