Skill Optimizer

Skill Optimizer — Automated Iterative Skill Improvement

Scope: proactive, background, multi-round. For reactive in-conversation edits after a single feedback moment, use self_improve instead. Use this when the user asks to systematically improve a skill or when self_improve has been applied several times and the skill still has recurring gaps.

What this does

Runs a background optimization loop for any skill:

1. Reads the current skill
2. Generates realistic test scenarios
3. Evaluates each scenario against the skill, in the worker's own reasoning
4. Synthesizes targeted improvements
5. Edits the skill in-place
6. Repeats up to max_optimizations rounds, then stops

Each round sends a progress notification. Stops early if a round produces no changes (skill converged).

Interactive intake (when the user triggers this)

Step 1 — Confirm the skill exists:

manage_skill(action="read", name=skill_name)

If missing, tell the user and stop.

Step 2 — Collect parameters (ask if not provided):

• skill_name — the skill to optimize
• requirements — what the skill must reliably do (success criteria, in plain language)
• max_optimizations — rounds (default: 3)
• scenarios_per_round — test cases per round (default: 3)
• agent — whose seat runs the loop, and therefore who evaluates the scenarios (default: "condor")

Step 3 — Build and start the delegation:

delegate(
  action="start",
  agent="{agent}",
 
  timeout_sec=1800,
  task=<see template below, with all params interpolated>
)

Step 4 — Tell the user it is running in the background and END YOUR TURN.

Delegation task template

Interpolate all {placeholders} before passing to delegate:

You are running an automated skill optimization loop. Follow these steps exactly.

Target skill: "{skill_name}"
Requirements (what success looks like): "{requirements}"
Max optimization rounds: {max_optimizations}
Scenarios per round: {scenarios_per_round}
Evaluating agent: "{agent}"

===== OPTIMIZATION LOOP =====

Repeat the following block for round = 1, 2, … up to {max_optimizations}.
Stop early if a round produces no changes.

--- Round start ---

1. READ the current skill (do this at the start of EVERY round — a previous round may have changed it):
   manage_skill(action="read", name="{skill_name}")
   Save the body as `current_body`.

   Steps 2-4 are YOUR OWN reasoning, not tool calls. You are the evaluating agent —
   you are running in a background session of "{agent}", with its identity, memory and
   playbooks — so there is nobody to ask: `delegate(action="ask", agent="{agent}")`
   is your own slug and is refused as a self-ask. Do each step in your own head and
   write the result down before moving on, so a later step has something concrete to
   work from. (If a scenario genuinely needs ANOTHER domain's judgement, that is a
   real peer — `delegate(action="ask", agent="<other-slug>", task="...")` blocks and
   returns its answer.)

2. GENERATE test scenarios:
   Acting as a skill tester, read the skill and write {scenarios_per_round} concrete,
   realistic test scenarios. Each scenario is a specific user request an agent would
   need to handle using this skill. Number them. Be adversarial — include edge cases
   and ambiguous inputs. Hold them against: Requirements: {requirements}

3. EVALUATE each scenario, one at a time, in order:
   For each scenario N, imagine a user just said it to you. Follow the skill's guidance
   as closely as you can to handle it. Then write down:
   (a) What steps the skill directed you to take
   (b) What worked well
   (c) What was incomplete, ambiguous, or forced you to improvise OUTSIDE the skill
   (d) Specific gaps or missing steps
   Keep every evaluation report; the next step reads all of them together.

4. SYNTHESIZE improvement:
   Acting as a skill editor over the evaluation reports, identify recurring gaps and
   improve the skill body. Rules: make surgical edits only — add missing steps, clarify
   ambiguous ones, remove what caused failures. Do NOT bloat the skill. Do NOT change
   the description, when_to_use, or name fields — only the body. Produce the COMPLETE
   improved body (full markdown, ready to write).

5. APPLY the update (only if body meaningfully changed):
   If the improved body differs from current_body:
     manage_skill(action="edit", name="{skill_name}", body=improved_body)
     changed = True
   Else:
     changed = False

6. NOTIFY round result:
   send_notification(
     f"🔄 *Skill optimizer* — `{skill_name}` round {round}/{max_optimizations}\n\n"
     f"Scenarios tested: {N}\n"
     f"{'✏️ Skill updated — ' + one_line_summary_of_changes if changed else '✅ No changes — skill already handled all scenarios'}\n\n"
     f"{'Starting next round…' if more_rounds_and_changed else 'Stopping — skill converged.' if not changed else '✅ Optimization complete!'}"
   )

7. If more rounds remain AND changed: asyncio.sleep(30), then continue loop.
   If NOT changed: stop (converged).

===== END LOOP =====

FINAL NOTIFICATION:
send_notification(
  f"✅ *Skill optimization complete* — `{skill_name}`\n\n"
  f"Rounds completed: {rounds_done}/{max_optimizations}\n"
  f"Total scenarios tested: {total_scenarios}\n"
  f"Outcome: {brief summary — what kinds of gaps were found and addressed, or 'skill was already solid'}"
)

Rules

• Read the skill at the start of every round — not once before the loop; previous rounds change it
• Never change description, when_to_use, or name — optimizer touches only body
• Never delete the skill — only edit
• Skip a scenario you cannot evaluate — log it in the notification but continue the round
• Stop early on convergence — a round with no edits means the skill is stable against this scenario set; stop rather than burning more credits
• Budget awareness — 3 rounds × 3 scenarios ≈ a few minutes of reasoning; well inside the 30 min worker budget
• One optimization at a time — don't start a second loop on the same skill while one is running

Example trigger

"Optimize the run_a_grid skill. It should handle edge cases around narrow ranges and high-volatility markets. Run 3 rounds with 3 scenarios each."

→ Read skill → delegate the loop → end turn.