SkillOpt-flavored offline training loop for any SKILL.md. Treats accumulated learn-rule corrections as training trajectories, proposes bounded patches via an optimizer LLM, gates each candidate against a held-out validation set built from the user's own past corrections, and ships only candidates that demonstrably improve the score. Inspired by Microsoft SkillOpt's ReflACT pipeline (rollout → reflect → aggregate → select → update → evaluate) adapted to pro-workflow's SQLite store. Use when a ski
git clone https://github.com/rohitg00/pro-workflow.git--- name: skill-optimizer description: SkillOpt-flavored offline training loop for any SKILL.md. Treats accumulated learn-rule corrections as training trajectories, proposes bounded patches via an optimizer LLM, gates each candidate against a held-out validation set built from the user's own past corrections, and ships only candidates that demonstrably improve the score. Inspired by Microsoft SkillOpt's ReflACT pipeline (rollout → reflect → aggregate → select → update → evaluate) adapted to pro-workflow's SQLite store. Use when a skill has accumulated 8+ learn-rule rows and the user wants the skill itself to get better, not just longer. user-invocable: true --- # Skill Optimizer Train an existing SKILL.md the way a deep-learning optimizer trains weights: via rollouts, gradient-like reflections, validation-gated acceptance. No model retraining; only the skill markdown changes. ## When to use Use this skill when: - A pro-workflow skill has accumulated 8+ learn-rule rows for it - The user reports the skill is "getting bloated" or "rules keep being repeated" - The user wants offline, budget-capped improvement over multiple sessions Do not use when: - Skill has fewer than 8 trajectories (nothing to learn from) - The user wants real-time edits (this is offline, single-shot) - No key has been explicitly configured for the selected provider ## Architecture (mirrors SkillOpt's six-stage loop) ```text rollout pull recent learnings from SQLite (existing learn-rule rows) reflect optimizer LLM analyzes a minibatch, proposes add/delete/replace patches aggregate vote-merge patches across minibatches select clip by LR budget (default: 3 adds, 2 deletes, 3 replaces per step) update apply selected patches to a candidate skill content evaluate evaluator LLM scores candidate against held-out validation items gate accept candidate only if weighted score >= current + acceptThreshold slow update at epoch boundary, consolidate accepted edits into a coherent rewrite ``` Failed candidates are stored in a rejection buffer and fed back to the next reflect step so the optimizer doesn't propose the same patch twice. ## Run it In a plugin session, use the `providers` MCP server's `run_provider_task` tool with `task: "optimizer"` and `args: ["--slug", "<slug>", "--budget-usd", "0.50"]`. Keys come from the plugin configuration dialog. Standalone CLI installations use explicit `PRO_WORKFLOW_*_API_KEY` variables. See [provider configuration](../../references/provider-configuration.md); never request keys in chat or retrieve existing machine credentials. ```bash /skill-optimize <slug> [options] ``` Options (all optional; sensible defaults shown): | Flag | Default | Notes | |---|---|---| | `--epochs N` | 3 | Outer loop count | | `--batch-size N` | 8 | Trajectories per minibatch | | `--minibatches N` | 2 | Minibatches per epoch | | `--holdout N` | 6 | Validation items reserved (max ~25% of trajectories) | | `--budget-usd X` | 0.50 | Hard cap; loop aborts when spent | | `--optimizer-model M` | `claude-sonnet-5` | Reflect + slow-update model | | `--evaluator-model M` | `claude-haiku-4-5` | Gate model (cheaper) | | `--max-adds N` | 3 | LR budget per step | | `--max-deletes N` | 2 | | | `--max-replaces N` | 3 | | | `--accept-threshold X` | 0.0 | Minimum score delta to accept candidate | | `--max-skill-tokens N` | 2000 | Hard cap on candidate length | | `--slow-every N` | 2 | Epochs between consolidation passes | | `--json` | off | Machine-readable output | Kill switch: `touch ~/.pro-workflow/STOP` aborts the loop between steps. ## Output - Candidate accepted → SKILL.md overwritten, hash stamp appended in HTML comment - Run details persist in `optimization_runs`, `optimization_candidates`, `optimization_patches`, `optimization_rejections` - Validation set persists in `optimization_validation` (reusable across runs) Inspect after: ```bash sqlite3 ~/.pro-workflow/data.db "SELECT id, skill_slug, initial_score, best_score, accepted_steps, rejected_steps, spent_usd FROM optimization_runs ORDER BY id DESC LIMIT 5" ``` ## Rules - Validation set is frozen at run start. Never re-derive from new corrections mid-run. - One candidate per step. No parallel branches. - Slow-update output is itself a candidate; it must pass the gate to replace the best. - The optimizer LLM and evaluator LLM may be different models. Mixing a strong optimizer with a cheap evaluator is the SkillOpt-recommended config. - If `spent_usd >= budget_usd` at any step boundary, the loop ends with `stopped_reason="budget exhausted"`. - Patches whose anchor is no longer present in the skill (because a prior patch in the same step removed it) are recorded as rejected with reason `anchor_missing`. ## Provenance Inspired by Microsoft SkillOpt (arXiv:2605.23904). The six-stage rollout/reflect/aggregate/select/update/evaluate pipeline, LR budget, rejection buffer, and slow / meta update mechanics are adapted to pro-workflow's existing SQLite + learn-rule data plane. No SkillOpt code is reused. "ReflACT" is not a SkillOpt term and is not used here; the loop is referred to by stage names only.
["Prepare your data: Ensure your SQLite store has a training table (e.g., `ticket_classifier_train`) with columns for learn-rule corrections, and a validation table (e.g., `ticket_classifier_val`) with held-out corrections. Run `sqlite3 skill_store.db \".schema\"` to verify.","Configure parameters: Replace [SKILL_NAME], [TRAINING_TABLE_NAME], [VALIDATION_TABLE_NAME], and [MIN_IMPROVEMENT_THRESHOLD] in the prompt template with your values. For threshold, start with 2% if unsure.","Run the loop: Paste the prompt into your AI tool (e.g., Claude Code) and execute. The optimizer LLM will generate candidate patches based on your training trajectories.","Validate and ship: Review the output. Only patches with improved validation scores (p<0.05) are shipped. Use the diffs to update your SKILL.md file or codebase.","Iterate: After applying patches, rerun the loop with new corrections to continuously improve. Monitor metrics in your SQLite store to track long-term performance."]
No install command available. Check the GitHub repository for manual installation instructions.
git clone https://github.com/rohitg00/pro-workflow/tree/main/skills/skill-optimizerCopy the install command above and run it in your terminal.
Launch Claude Code, Cursor, or your preferred AI coding agent.
Use the prompt template or examples below to test the skill.
Adapt the skill to your specific use case and workflow.
I have a SKILL.md file that defines [SKILL_NAME] with current performance metrics. Run SkillOpt’s offline training loop on my local SQLite store: 1) Extract all learn-rule corrections from [TRAINING_TABLE_NAME] as trajectories, 2) Generate candidate patches with an optimizer LLM, 3) Gate each candidate against the held-out validation set [VALIDATION_TABLE_NAME], 4) Ship only patches that improve the score by at least [MIN_IMPROVEMENT_THRESHOLD]%. Output the top 3 validated patches with before/after metrics and a diff of the changes.
### SkillOpt Offline Training Results for 'ticket-classifier' **Training Data:** 1,247 learn-rule corrections (2024-05-01 to 2024-05-15) from `ticket_classifier_train` **Validation Set:** 312 held-out corrections from `ticket_classifier_val` (F1-score baseline: 0.821) **Optimizer LLM:** claude-3-5-sonnet-20240620 **Min Improvement Threshold:** 2% #### Candidate Patches (Validated & Shipped) **Patch #1: 'urgent-tag-handling'** - **Before:** F1=0.821, Precision=0.812, Recall=0.830 - **After:** F1=0.843 (+2.7%), Precision=0.835, Recall=0.851 - **Diff:** Added learn-rule: `IF ticket.subject CONTAINS "urgent" THEN priority = "high" AND tag = "urgent"` - **Validation Score:** 0.843 (p=0.003 vs baseline) **Patch #2: 'spam-filter-v2'** - **Before:** F1=0.821, Precision=0.812, Recall=0.830 - **After:** F1=0.848 (+3.5%), Precision=0.852, Recall=0.844 - **Diff:** Updated learn-rule: `IF sender.email DOMAIN IN ["spam1.com", "spam2.net"] THEN classify = "spam" (confidence=0.98)` - **Validation Score:** 0.848 (p=0.001 vs baseline) **Patch #3: 'escalation-path'** - **Before:** F1=0.821, Precision=0.812, Recall=0.830 - **After:** F1=0.839 (+2.2%), Precision=0.829, Recall=0.849 - **Diff:** Added learn-rule: `IF resolution_time > 48h AND status = "open" THEN escalate = "true"` - **Validation Score:** 0.839 (p=0.012 vs baseline) **Summary:** 3 patches shipped (out of 12 candidates). Average improvement: +2.8% F1-score. All patches passed validation with p<0.05. Changes applied to `ticket_classifier_v2.1.0`.
skills-collection
Take a free 3-minute scan and get personalized AI skill recommendations.
Take free scan