| name | self-improve |
| description | Parker's self-improvement governance for the brand — the disposing half of the living layer. Weekly it curates captured reasoning traces and decides which dreaming proposals get promoted, human in the loop — including the doc-alignment proposals the research cycle files. It is the gate: dreaming and research propose, self-improvement disposes. The loop pipeline itself (roll-up, hypotheses, validations, re-validations) runs in /research-loops. Use weekly as a scheduled routine, or when asked to curate learnings, promote proposals, or self-improve. |
Self-improve — the governing half of the living layer
Dreaming and the research cycle generate; self-improvement governs whether anything becomes canonical. Nothing they proposed is applied on its own — context edits, skill changes, new workflows, and the research cycle's doc-alignment proposals all route through this pass with the human in the loop. The job is to preserve the reasoning traces that should change future behavior, not to remember every chat fragment.
When it runs
Weekly (or before a major prompt/skill refresh). Continuous trace capture happens during everyday work; this is the weekly curation that promotes carefully.
First run — scaffold if absent
If self-improvement/ does not yet exist in this repo, create it: reasoning-traces/[YYYY-MM]/, review-queue.md, patterns/INDEX.md, applied-changes.md. This is the routing layer that says what should change and why — not a replacement for brand memory or prompt docs.
The two jobs of the weekly pass
1. Curate reasoning traces
Review new traces in self-improvement/reasoning-traces/[YYYY-MM]/, cluster repeated themes, promote repeated or explicitly-approved rules into applied-changes.md, update the affected prompts / skills / context docs / brand surfaces, keep unresolved items in review-queue.md with a promotion condition, and mark stale or contradicted traces rejected or superseded. Promotion preserves provenance — cite the trace IDs that caused each change.
A trace is worth capturing when the user corrects Parker and explains why, says a rule should apply broadly, approves/rejects an output for a named reason, reroutes where content lives, edits a hypothesis or strategic read, or names a repeated failure mode. Not for routine status updates, typo fixes, one-off wording, or facts that belong in a brand context doc.
Use the narrowest correct route: a local output issue → fix the output; brand learning → running-notes/ or brand rules; a prompt/skill behavior → a trace + candidate on the affected skill; a system rule → a trace, and update CLAUDE.md only when the user explicitly asks or the rule is already clearly approved.
2. Dispose of dreaming proposals
Walk dreaming/proposals/pending/. For each, decide with the human in the loop:
- Promote → apply the change for that bucket (context edit, skill change, schedule, idea, or open loop), move the file to
dreaming/proposals/applied/, and log it in applied-changes.md tied back to the proposal.
- Dismiss → move to
dreaming/proposals/dismissed/ with the reasoning preserved (a rejected dream teaches what doesn't land).
A proposed schedule in schedules/proposed/ is promoted into an active schedules/[slug].md only on explicit confirmation — and arming its cron is then a setup-routines step. A proposed idea routes to the idea bank (harvest-ideas); a proposed open loop is captured in open-loops/ for the next /research-loops roll-up. A doc-alignment proposal from the research cycle (a validated finding waiting to be folded into a standing doc) is applied on the user's agreement with its validation provenance intact, or dismissed with the reasoning preserved.
Where the loop pipeline went
The research work that used to live here — the roll-up, advancing loops into hypotheses, running validations, and due re-validations — is its own standing routine now: /research-loops, weekly and mid-week, so findings are fresh by the time this pass curates them. This pass stays the governor: it disposes of the alignment proposals research files, and it curates what the week's findings taught into durable learning.
Hard rules
- Dreaming and research propose, self-improvement disposes — and the human stays in the loop. Do not silently promote an inferred "why" or incomplete history; ask for a context dump or confirmation before promoting.
- Do not treat one correction as a universal rule unless the user says it's global or the pattern repeats.
- Do not create new docs when an existing living surface can absorb the learning.
- Do not update a prompt or skill from a single trace without explicit user approval.
- Always connect an applied change back to the trace or proposal that caused it; preserve reasoning for rejected items too.
- Do not overwrite brand-specific rules with global rules; mark thin evidence as thin.
- Honor the brand hard rules on anything that touches creative or claims.
- Self-contained: in-repo surfaces only. The global product-signals stream (anonymized, cross-brand) lives with the factory, never in this brand repo. No factory paths at runtime.
Deliverable
Curated traces with promotions logged in applied-changes.md; dreaming and research proposals moved to applied/dismissed with reasoning, doc alignments applied with their validation provenance; and a short report of what was promoted, what's awaiting the user, and what was dismissed (and why).
