Skilladam Optimize

SkillAdam Optimize

Use the skilladam_* MCP tools when available. Otherwise invoke the equivalent skilladam-product commands and parse their JSON output.

Build the task manifest

For a new run, research and construct the task signal in the current host session before calling SkillAdam. Do not start a nested agent or external task generator solely for this step.

1. Read the complete target SKILL.md and the user's usage intent.
2. Call skilladam_discover_history exactly once with skill_path, intent, and the durable output_dir. It automatically scans the current host's known local history formats, keeps only real user messages, redacts likely secrets and personal data, deduplicates them, and writes history_discovery.json. Review each candidate for actual relevance and evaluation feasibility. Turn a useful real request into a self-contained evaluation case by minimally normalizing it and, when necessary, embedding a small synthetic artifact that preserves the requested work. Record metadata.history_transform describing that normalization; never claim the synthetic artifact came from the user. Use source local_history and the candidate's exact returned source_ref; never invent a history reference. A completed search with zero accepted candidates is valid and must not be relabeled as a history-derived task.
3. Reuse concrete tasks the user supplied. Extract tasks from the currently visible conversation only when they represent actual Skill usage; use source conversation, redact secrets and personal data, and keep a precise turn reference.
4. When a recognized public benchmark would materially improve coverage, use the host's normal web/research tools to select relevant cases. Record the direct source URL and benchmark/license details in metadata. Do not invent provenance when no suitable source was found.
5. Fill the remaining coverage from the Skill itself. Produce exactly task_count distinct tasks, covering at least two concise lowercase capability slugs and at least one boundary, adversarial, or regression case. Prefer realistic end-to-end requests over shallow paraphrases.
6. Give every task a stable task_id, self-contained prompt, capability, difficulty, source, source_ref, and explicit evaluation. Allowed sources are local_history, conversation, user_provided, public_benchmark, and skill_generated, in that priority order when the source yields a genuinely useful task. Public benchmark references must be direct HTTP(S) URLs.
7. Use deterministic programmatic or reference evaluation only when the task has one uniquely correct answer, such as a number, fixed string, explicit JSON structure, or exact set of fields. Open-ended natural-language tasks that allow semantically equivalent answers must use a rubric_judge rubric with weighted observable dimensions; never score the complete response with exact or normalized_exact. This includes generated messages, reviews, rewrites, summaries, recommendations, and analyses. Use hybrid only for complementary components. Every evaluation, including each hybrid component, must include concise failure_feedback with the actionable correction or evidence to use after a failure. Never omit an evaluation or ask SkillAdam to infer one. Correct a rejected manifest instead of weakening the requested count or evaluation goal.

Pass {"schema_version":"1","tasks":[...]} in task_manifest. Task text must not depend on files, tools, credentials, or context unavailable to the rollout model.

Start or resume

1. Obtain the source SKILL.md, a durable workspace-local output directory, and the user's usage intent. Use a path such as run/skilladam-optimize under the current workspace; never use a host temporary or session-state directory.
2. For a new run, call skilladam_prepare with the source skill_path, intent, durable output_dir, completed task_manifest, and compact: true; omit provider fields so SkillAdam reuses the platform's authenticated CLI. Pass provider configuration only for an explicit fixture or OpenAI-compatible override. Use skilladam_prepare_for_review instead only when the user explicitly requested confirmation or hunk selection before apply. Do not use the asynchronous skilladam_start* tools: the Claude adapter intentionally does not expose them because a non-interactive turn may stop polling before a worker reaches a terminal state.
3. In default mode, call skilladam_prepare again with the same output_dir and compact: true while the workflow state is ready or validating, until the existing stopping policy returns completed. Each call resumes the durable checkpoint. If a call fails, report the exact error; do not shrink task_count or weaken evaluation settings as a workaround.

Review a proposal

1. Read proposal.hunks from the structured result. Preserve every hunk_id.
2. Determine whether the user explicitly requested hunk review or confirmation before changes are applied. A general request to optimize, improve, patch, or modify the Skill does not opt in to review.
3. By default, skilladam_prepare auto-accepts every proposed hunk before the call returns. Do not issue a second selection, omit or reject hunks, pause for confirmation, or override the gate based on your own assessment.
4. Only after an explicit review opt-in, use skilladam_prepare_for_review, present every hunk's target, header, and operations in original order, and wait at awaiting_review. Prefer a native structured selection UI that can map decisions back to exact hunk IDs; otherwise ask the user for accepted IDs, accept all, or reject all.
5. In review mode, never treat silence, timeout, disconnect, resume, or an unrelated positive comment as acceptance. An empty selection rejects all.
6. Submit exactly the selected IDs with a new, stable idempotency key. Do not use review results from a different proposal or infer a partial selection.

Validate and iterate

1. Trust only candidate_validation whose digest matches the selected candidate. Never reuse validation for the full proposal after a partial selection. Each iteration's current Skill rollout and selected candidate validation must keep the exact ordered mini-batch IDs recorded in the session.
2. Report the gate decision and stop reason concisely. The service writes the source Skill only when the gate accepts the selected candidate. After the gate, the service updates EIT/Momentum from ordered trajectory feedback, the patch, candidate feedback, and the decision, then updates the adaptive edit budget.
3. Continue synchronous skilladam_prepare calls through ready and validating until the workflow reaches completed; the same tool resumes both states, so do not call skilladam_continue.
4. If the source changed during review, preserve the external edit and report the conflict; do not overwrite it.
5. On completed, report final quality only from final_validation whose scope is heldout_terminal. baseline_validation and candidate_validation are optimization mini-batch evidence and must never be described as final or held-out. If terminal evidence is unavailable, report that exact condition; do not substitute another metric. Never feed final held-out evidence into patch generation, hunk selection, an iteration gate, or Momentum.
6. Use the compact response's iterations entries for per-iteration results. Do not request a full status response merely to reconstruct the final report.

Never edit, replace, or patch the source Skill with host file tools or shell commands during an optimization session. This remains forbidden when a proposal looks weak, the gate rejects it, or the iteration limit is reached. Only SkillAdam may write a selected candidate after frozen validation passes the gate; otherwise leave the source unchanged and report the session result.

The host owns task research and manifest construction only. Do not reimplement manifest validation, evaluation execution, patch application, gate logic, Momentum, edit budgets, checkpoints, or stopping policy in the host adapter.

Claude Code structured review

When the user explicitly requested confirmation or selective hunk review:

1. After skilladam_prepare_for_review reaches awaiting_review, present every hunk in original order. For each hunk, show its stable hunk_id, target path, header, and complete operations inside a fenced diff block. Do not summarize or omit the diff before asking for a decision.
2. Use Claude Code's AskUserQuestion tool to collect an explicit accept/reject decision for each displayed hunk. Keep the stable hunk_id in each question and batch questions only when required by the host tool's current limit.
3. Map the answers back to exact hunk IDs and call skilladam_submit_selection once with the final accepted set. Cancellation, silence, timeout, or an incomplete answer leaves the session at awaiting_review and must never trigger default acceptance.