Skip to content

agent-skills-quickstart.md

Latest commit

 

History

History
141 lines (104 loc) · 4.57 KB

agent-skills-quickstart.md

File metadata and controls

141 lines (104 loc) · 4.57 KB
Raw

Agent Skills Quickstart

CheMLFlow skills are small operating manuals for agents. They teach an agent how to do CheMLFlow-specific work consistently: create single configs, review DOE files, audit analysis outputs, and avoid common manifest, row-count, scaler, and split-balance mistakes.

1. What is included

skills/
+-- chemlflow-study-runner/
+-- chemlflow-config-builder/
+-- chemlflow-doe-designer/
+-- chemlflow-analysis-curator/
+-- chemlflow-molecular-analysis/

Each skill has:

  • SKILL.md: concise instructions and trigger description
  • optional references/: details loaded only when needed
  • optional scripts/: deterministic helper checks

2. Use a skill in a prompt

Ask your agent to use the skill by path:

Use the CheMLFlow Study Runner skill in skills/chemlflow-study-runner to coordinate a local DOE run and audited analysis.

Use the CheMLFlow Config Builder skill in skills/chemlflow-config-builder to create one runtime config for a PGP random-forest baseline.

Use the CheMLFlow DOE Designer skill in skills/chemlflow-doe-designer to review config/doe_pgp.yaml.

Use the CheMLFlow Analysis Curator skill in skills/chemlflow-analysis-curator to audit pah/pah_analysis_6689856.

Use the CheMLFlow Molecular Analysis skill in skills/chemlflow-molecular-analysis to inspect this curated SMILES/pIC50 dataset and create a dedicated molecular EDA config.

The molecular-analysis skill is opt-in. Do not use it merely because an ordinary config or DOE dataset contains SMILES, IC50, or pIC50. Invoke it for an explicit request for molecule inspection, chemical-space projection, unsupervised clustering, activity-discontinuity analysis, or named molecular publication figures.

3. Run the helper checks

The config-builder skill is currently an operating manual, not a scripted checker.

Summarize generated DOE artifacts:

python skills/chemlflow-doe-designer/scripts/summarize_doe.py tmp/pgp_hpcc_analysis/pgp_doe

Run generated DOE configs locally:

python scripts/run_doe_local.py --doe-dir config/generated/my_doe --max-workers 1 --resume

Analyze local DOE outputs:

python analysis.py --backend local --doe-dir config/generated/my_doe --output-dir config/generated/my_doe/analysis_local

Audit analysis outputs:

python skills/chemlflow-analysis-curator/scripts/audit_analysis.py pah/pah_analysis_6689856

4. What the agent should check

For single-config work, the agent should inspect:

  • dataset shape: SMILES, tabular features, or non-molecular data
  • task type: regression or classification
  • curation/drop-row settings
  • feature/model compatibility
  • split mode, seed, scaler, and output paths
  • whether full K-fold CV should be handled by DOE fanout
  • Morgan vs RDKit assumptions, and random vs scaffold assumptions

For DOE work, the agent should inspect:

  • summary.json
  • manifest.jsonl
  • parent_manifest.jsonl
  • model, feature, scaler, and split compatibility
  • valid, skipped, and parent case counts
  • local vs Slurm execution backend

For analysis work, the agent should inspect:

  • report.json
  • all_runs_metrics.csv
  • all_runs_metrics_by_execution.csv
  • raw vs aggregated row counts
  • scaler, Morgan/RDKit, model, and split balance
  • failed or incomplete folds before discussing model performance
  • whether report.json says backend: local or backend: slurm

For optional molecular-analysis work, the agent should check:

  • that the user explicitly requested the optional analysis;
  • SMILES, identifier, property, and units compatibility;
  • that the work uses a dedicated dataset-analysis config rather than DOE children;
  • that publication figures are explicitly selected;
  • molecular EDA and publication manifest status, identity hashes, and artifact checksums.

5. Optional auto-discovery

Keep skills/ in this repo as the source of truth. If your agent supports automatic skill discovery, copy or symlink the skill folders into that agent's personal or project skill directory.

Example:

mkdir -p ~/.codex/skills
ln -s "$(pwd)/skills/chemlflow-doe-designer" ~/.codex/skills/chemlflow-doe-designer
ln -s "$(pwd)/skills/chemlflow-analysis-curator" ~/.codex/skills/chemlflow-analysis-curator
ln -s "$(pwd)/skills/chemlflow-config-builder" ~/.codex/skills/chemlflow-config-builder
ln -s "$(pwd)/skills/chemlflow-study-runner" ~/.codex/skills/chemlflow-study-runner
ln -s "$(pwd)/skills/chemlflow-molecular-analysis" ~/.codex/skills/chemlflow-molecular-analysis

After installing or symlinking skills, restart the agent session so it can reload available skills.