Self Improve

Self-Improve — Real-Time Reflex

Scope: reactive, in-conversation, single-turn. Fires immediately after a feedback moment — a correction, a stated preference, a missed step. For proactive systematic improvement of a skill (multi-round scenario testing loop), use skill_optimizer instead.

Run this checklist immediately after any feedback moment. Do NOT accumulate and review at end of session (token-expensive and easy to forget).

Checklist

1. Is this a user preference or stable fact?

• e.g. "always show controllers before asking for params", "use 3 months as default backtest window"
• → manage_memory(action="write", ...) — one memory per fact, type=preference
• Memory is YOURS ALONE — keyed by (assistant, user), so what you write here is invisible to the chat and to every other agent, and there is no way to publish it. Write it anyway for your own work, but if the preference should govern how everything behaves, say so to the user — a cross-cutting rule belongs in a shared skill (below), not in one agent's memory.

2. Is this a workflow pattern or missed step?

• e.g. "when asked to backtest, list available controllers first"
• → First: manage_skill(action="list") — check if an existing skill covers it
  • If yes: manage_skill(action="read", name=...) to confirm, then manage_skill(action="edit", ...) to update it
    • If inherited: true: the skill is shared and read-only. create a local skill with the same name to shadow it, or ask the user to have the chat edit the published version.
  • If no: manage_skill(action="create", ...)
• Before writing the body, read skill_authoring — it covers anatomy, scoping, routine linking, and companion files.
• Whose library? Decide before creating: yours (no agent), a domain agent (agent="slug"), or all agents (shared=True). See skill_authoring for the full scoping rules.

3. Is this a one-off or session-specific detail?

• → Skip it. Don't pollute memory/skills with ephemeral context.

When to escalate to skill_optimizer

A single reactive edit (step 2 above) is right for most feedback moments. Escalate to skill_optimizer when:

• The user asks for systematic or automated improvement ("make it handle edge cases", "iterate on this skill")
• The same gap has surfaced more than once across conversations
• The skill is complex enough that a single edit may introduce new blind spots

skill_optimizer runs a multi-round scenario-testing loop in the background — read it before triggering.

Rules

• Save in the moment, not at the end
• One memory = one fact (no bundling)
• Always check existing skills before creating a new one
• Decide the OWNER before writing
• Shared skills land in every agent's context — publish deliberately