Do you regularly optimize your claude with Claude?

I showed much of my work this week, and so I just wanted to drop this and share that I weekly make Claude go through itself and ensure we are working at the best architecture, and cost savings.

Especially after I have it explore the usability of various git hubs.

Once it understood what I do as the overseer, and the architectural process we want to follow, it looks at each github I give it and analyzes it in reference to what I do, as well as in reference to cost.

Here is an output example:

Done. Final pass complete.

What changed:

_config/BRAND.md — added AI-business-speak blacklist (leverage, streamline, empower, cutting-edge, seamless, robust solution, delve into, navigate the landscape). Em-dash limit already there.
CLAUDE.md — added two operator triggers: "red team [topic]" (destroy this plan) and "pre-mortem [topic]" (assume it failed in 6 months, top 3 reasons). Inline specs, no separate files needed. Use them on the June Challenge before June 8.
CLAUDE.md — added a new WHAT NOT TO DO rule: do not Read full files when you only need a slice. Use grep + Read offset/limit on any hot-path file >100 lines. Reading 644 lines of MAY-CALENDAR to extract 30 was the most expensive habit in the setup.
briefs/_TODAY-TEMPLATE.md step 2 — made the calendar slice-read explicit. Each "prep today" call now saves ~600 lines of read.
briefs/_MONTHLY-AUDIT-LOG.md — tracked as A10/A11/A12 with audit-history columns expanded.

Architecture verdict. ICM layers are right-sized. L0 (CLAUDE.md / CONTEXT.md), L3 (_config/), L4 (daily/briefs/sandboxes/CONTENT-STATUS) are properly separated. _GABEYOGA-BOOTSTRAP.md is portable kit, not auto-loaded — zero hot-path cost. _WORKER-RULES.md is correctly scoped to sandboxes. No restructuring needed. The leak was the read pattern, not the structure.

Concrete token wins now in place:

WinMechanismEstimated saving per fireCalendar slice-readgrep + offset/limit~600 lines per "prep today" (daily)Reference-only daily briefsDefault mode = link, not inline~30–50% per daily file (compounds on reads)CONTENT-STATUS date auto-stampPostToolUse hookOne Edit call per content update (~50 tokens)CONTEXT.md dedupeRemoved duplicate triggers~22 lines per session startGeneral slice-read ruleNew WHAT NOT TO DO bulletBounded — applies to any 100+ line hot-path file

The daily compound matters: every "prep today" now reads roughly 30 lines instead of 644 for the calendar plus summary + paths instead of full inlined post copy. Over a month, that's the meaningful cost reduction.

That's the full sweep. Real session end now. See you Sunday.

2 comments