A Claude skill registry is the set of instruction folders an AI assistant loads by name when a task needs them. On September 19, 2026 I merged mine from 89 custom skills into 38, and kept every rule those 89 skills contained. This is the site’s record of that change, what a review found after install, and how the system grades now.
It follows the orchestrator layer post from August, written when the registry had 80 skills and a router in front of them. The router is still there. What changed is what it routes to.
What changed since the 80-skill registry?
The count crept back up. The August trim took the registry from 107 skills to 80, and six weeks of new work grew it to 95 folders, 89 of them skills I wrote. Trimming had not fixed what kept failing either. My failure log showed contracts split across two files: a skill that wrote a field another skill read, a check count quoted in three places with three different numbers, and pairs of skills that had to be installed together arriving one half at a time.
So this time the unit changed. Each of the 31 merged skills holds the unedited text of every skill folded into it, one per named mode, and routing names both parts, like verify:claim or run-control:loop. Seven skills stayed on their own. 80 old folders were retired, and a table in the router maps every old name to its new address so nothing typed from habit breaks.
Where are the full write-ups?
Two longer articles cover it from different sides. The mechanics, including the nine-check gate and the two blind spots it caught in my own tooling, are on DEV: I cut my Claude skill registry from 89 skills to 38 without deleting a rule. The reasoning, and why merging beat deleting, is on CoderLegion: Efficacy vs capability: why I merged my AI skills instead of deleting them.
This post adds what neither article has: the review that ran after install, and a graded report on the whole system.
What did the review after install find?
Five defects, all in routing or in my own tools, none in the merged content. The router had no route for 31 platform and vendor skills. Two short legacy names, run-loop and html-css, survived as plain words in 62 places because the rename pass skipped names under nine characters. Two skills advertised sub-modes their mode tables did not list. The gate that checks for old names had a blind spot in table rows. And a rebuild read my own merged skill back in as one of its own sources.
All five are fixed in a patch that is packaged but not yet installed. Three checks held without changes: every writer contract between skills still reciprocates, every skill-and-mode address resolves, and my standing instructions name only skills that exist.
How does the new system grade?
I ran a graded report card over the week of September 12 to 19, using a fixed five-axis framework and citing evidence for every grade. These are self-assessed, and they are regenerated each run rather than kept.
| Axis | Grade | Why |
|---|---|---|
| Shipping and proof | B+ | Prismwar went live on the Chrome Web Store with every gate re-run on the extracted package, but four finished items are still waiting on a push, a submission or an upload. |
| Code craft | A | Storage races fixed with per-key queues and zero escaped sabotage mutants, but one project’s tests were written by the same session that wrote its fixes. |
| Architecture and design | A | Hosting and scope decisions were made from measurements, though the homepage scene’s scope reversed within two days. |
| System thinking | A | Each repeat failure now has an owning skill, and the always-on skill text fell by half. |
| Learning velocity | B+ | Fixes land quickly, but the same failure class keeps turning up inside the tools written to fix it. |
The one target I took from it: nothing finished stays packaged and unshipped for more than seven days without a named blocker. The weak grade is delivery, not craft.
What did it save?
Skill descriptions load on every turn, so they are the number that matters. They went from 70,918 characters to 33,962, a 52% cut. My standing instructions to Claude went through the same treatment, from 6,837 characters to 3,881, by moving every rule a skill already enforces into that skill. For a sense of why a token count on its own is a poor measure of effort, Are tokens the new unit of effort? covers the Token Guard side of the same question.
What is not done yet?
The patch from the review still has to be installed. The install itself is not fully confirmed, because the file copy my chat sessions read lags the account, and one updated skill cannot be confirmed from its description at all. Six merged skills are over 80 KB and load whole when they fire, so the next step is one reference file per mode. And 64 section references in prose were broken before the merge and are still broken; the merge added none.
That last kind of drift has cost me before. The memory file was the bug is the story of a stale record that got a 966 KB zip built twice, which is part of why this review checked the installed skills against two sources instead of trusting one.
Every devlog entry is listed on the history page, newest posts are on updates, and the tools these skills help build are on the projects page.