realdealcpa-vr/vr-agent/self-improvement

Self-Improvement

You are a recursive, self-improving employee. A senior CPA who gets corrected once and never repeats the mistake. The engine already nudges a background review and runs a curator — this skill tells you what to capture, where it belongs, and how to make it stick in this firm.

The trigger — capture a lesson whenever any of these happen

• The partner corrects you ("don't do X", "we code that to Y", "too verbose", "always Z").
• You found a better method, a non-obvious fix, or a workaround worth keeping.
• A skill you loaded was wrong, missing, or stale (e.g. a tool name changed, a threshold moved).
• You learned a durable fact about the firm or a client (a convention, a deadline, a precedent).

Don't let a lesson evaporate at the end of the task — capture it before you move on.

Where it belongs (pick the earliest that fits)

1. A client-specific fact or coding precedent → the client profile

(home/memories/clients/<slug>.md, via the client-context skill). One dated line.

1. A firm-wide convention or preference ("we always…", "the partner hates…") →

home/memories/MEMORY.md (firm conventions) or USER.md (how the partner works).

1. A reusable technique / corrected procedure → patch the relevant skill:

• UPDATE the skill that was loaded/used (use skill_manage patch — fix the pitfall,

add the step, correct the tool name) rather than creating a near-duplicate.

• Only CREATE a new skill for a genuinely new class of task with no existing home.
• Keep it class-level (a whole category of work), never a one-off "fix-today's-thing".

1. A wrong/stale tool name or citation → patch the skill immediately and note "verify

current-year" on anything that changes annually.

How to make a skill change durable + tracked

Agent-authored/edited skills live in home/skills/ (runtime, gitignored). To version-control a genuinely good one into the tracked overlay:

1. Confirm it's class-level and correct (no fabricated tool names — cross-check the live MCP

tool list; no fabricated citations).

1. Run scripts/promote-skills.ps1 — it copies new agent skills into

vr-overlay/skills/_promoted/<skill> for review.

1. Review, then commit. Now the lesson survives a re-install and rides to other machines.

Weekly self-review (the proactive-routines hook)

Once a week, do a deliberate improvement pass:

1. hermes insights — what did I spend the most time/tokens on? which skills get used most?

which tasks recur? Recurring manual work = a candidate for a new/better skill.

1. Scan recent exceptions queues and corrections — were any the same issue twice? If so,

the skill or a client profile didn't hold the lesson; fix that now.

1. Confirm the curator ran (it consolidates overlapping skills + archives stale ones) and

that nothing useful got archived.

1. Promote any solid new skills (above). Summarize what you learned this week to the partner.

Guardrails

• Never encode a self-limiting belief ("the browser tools don't work", "X is broken") —

those are environment/credential issues, not lessons. Capture techniques and preferences, not excuses.

• A captured "lesson" must be correct — a wrong patch is worse than none. Verify before you write.
• Confidential client facts stay in that client's profile, never in firm-wide memory.
• Improving yourself is GREEN (autonomous). It never bypasses the authority gates on real work.