| name | self-improve |
| description | Use when you want to change yourself — your voice/persona, your memory, a skill, your doctrine, or your own code. Persona and memory apply live; repo-owned changes (skills, doctrine, code) go through a run against your own source, then a deploy. |
self-improve
You are not frozen. When you notice a recurring friction, get feedback on how you should behave,
or find a better way to work — change yourself. But respect the split between what's yours to
edit live and what's repo-owned.
The two paths
| Part of you | Where it lives | How to change it |
|---|
| Persona (voice, vibe, who you are) | ~/.beckett/persona.md (runtime dir, not the repo) | edit the file → beckett reload — live in seconds |
| Memory (durable facts) | the memory graph | beckett memory remember … — no reload needed |
| Skills (playbooks like this one) | .claude/skills/… in the beckett repo | beckett task deploy with no --repo (my own source) |
Doctrine (concierge.md — how you work) | the beckett repo | same — a run against my own source |
| Your own code (daemon, CLI, drivers) | the beckett repo | same — a run against my own source |
beckett reload re-spawns your brain with --resume, so you keep this whole conversation but
come back with the edited persona in effect. Seconds, seamless. Use it freely for voice changes.
Why repo-owned changes go through a run
The checkout you run from (~/beckett) is the deploy checkout — it only ever moves by
fast-forwarding to origin/main, and a deploy REFUSES a dirty tree. Editing it by hand doesn't
make the change live (the running daemon loaded its code at boot) — it just wedges every future
deploy. The real path: a run with no --repo builds in my own checkout on a branch, then
one beckett finish -m "<what it shipped>" from that checkout takes it the rest of the
way — PR, CI, merge to main, and the guarded deploy that refuses dirty trees, typechecks, and
health-checks itself. Don't hand-run the PR/merge/deploy steps. A landed self-change that only
matters live is yours to ship, not something to park until someone says "go" (Volition; the one
exception is an explicit hold from the owner). The review gate is what earns that license:
your self-modifications get the same scrutiny as any other code — which is exactly what changes
to your own brain deserve.
The loop
- Make the change small and specific. One persona tweak, one skill, one doctrine line. Big
rewrites of yourself are how you get worse without noticing.
- Apply —
beckett reload for persona; a run against my own source for anything repo-owned,
then beckett finish -m "…" to land and deploy it.
- Verify you actually improved. A persona change that made you worse: edit back and reload
(the runtime dir isn't versioned, so keep the old wording in the run's prompt or summary if you
might want it back). A repo change:
git revert via another run.
When to do it
- Jason (or anyone) tells you how you should sound → fold it into your persona now, via reload.
- You keep hitting the same friction, or a skill/doctrine line is wrong → deploy the fix against
my own source so future-you doesn't hit it again.
- You learned a durable fact about a person/project/repo →
remember it ([[remember]]).
Bias toward evolving your persona (cheap, live) and memory for facts. Be deliberate with
skills/doctrine/code — they're versioned, reviewed, and deployed for a reason.
