fully-automated-prompt-optimization: Command for Claude Code

---

description: > Prune noncompliant synthetic examples and normalize placeholder data. TRIGGER when: user wants to clean up synthetic examples, remove bad samples, fix placeholder hashes, validate synthetic data quality, or align review CSVs. DO NOT TRIGGER when: user is creating new synthetic examples (use synthetic-samples), running evals (use eval-runner), or optimizing prompts (use optimization agent).

Synthetic Pruner

Overview

Prune noncompliant synthetic examples and normalize placeholder hashes while keeping review CSVs aligned to the remaining example directories.

Workflow

1. Review requirements
   • Use tenant-specific dataset requirements plus docs/references/synthetic-requirements.md.
2. Identify severe violations
   • Example default rule: email bodies with <= 10 words and no greeting/signature are severe violations.
   • Missing-body examples can be kept when intentionally included as hard cases.
   • Placeholder hashes like repeated A/B/0 values are realism violations; fix them rather than deleting examples.
3. Apply changes
   • Remove only the examples that match the agreed severity rule.
   • Update review CSV files in the examples root to remove deleted examples.
   • Replace placeholder attachments_sha256 with realistic 64-hex values in remaining examples.
4. Verify
   • Ensure CSVs reference only existing example directories.
   • Confirm no placeholder hashes remain.

Scripted Cleanup

Use scripts/synthetic/prune_synthetics.py for deterministic cleanup. Default mode is dry-run.

Examples:

python scripts/synthetic/prune_synthetics.py \
  --examples-dir tenants/<tenant_id>/datasets/synthetic_artifacts \
  --max-words 10

python scripts/synthetic/prune_synthetics.py \
  --examples-dir tenants/<tenant_id>/datasets/synthetic_artifacts \
  --max-words 10 \
  --apply

Guardrails

• Do not modify tenant source artifacts under tenants/*/source_artifacts/.
• Keep changes scoped to synthetic example folders and their review CSVs.
• If requirements or thresholds change, update this skill and document in auto-memory notes if needed.