Memory & knowledge

Skill asset management

Inventory, context cost, overlap and staleness — decide what to keep and what to retire.

Paid $15 list, $12 until 2026-12-31 · one skill, yours for good. The outline, the steps and the file open after purchase.

Background

Thesis: skills are assets, but assets with a holding cost — every skill's description enters the system prompt, so you pay for it once per turn. The management goal is therefore not "the more the better" but "every one gets used, and is worth the lines it occupies".

Why it is built this way

Classification comes first because the number that matters is the original-versus-installed split, and the answer goes into references/asset_ledger.md so the next audit reads it instead of re-judging. The marker on the first body line settles most cases, with one exception: the marker is a claim of authorship and the hash is evidence, so Get-FileHash runs against the candidate upstream first. Reversing that order misjudges. Four *-upload skills shared timestamps, a Chinese voice and isomorphic structure, and were byte-for-byte copies of an upstream MIT repository. Four markers were revoked.

The four cuts are ordered by how surely they can be judged, then by benefit, then by risk. Never-used cannot be measured, since there is no invocation log and a timestamp means modified. What can be measured is loaded: skills listed in the system prompt are the ones paying context each turn, while a cached skill that is not enabled only occupies disk. So cut one saves disk, cut two removes a family of descriptions, cut three merges two skills claiming the same action, and cut four shortens a long description without dropping the count.

Where it bites

• Judging authorship by creation time alone A batch built inside one hour gets labelled installed, or labelled original, as a whole. Mixed batches do exist: 32 skills in a single hour with self-built and installed side by side. The timeline only screens candidates; splitting a mixed batch needs the content criteria plus Get-FileHash against upstream, hash first.
• Installing on a matching SHA A hash that matches proves the copy is accurate and says nothing about what it does. The pre-install gate runs on its own: preflight-install.ps1 exits 2 for do not install, 1 for read every script-type file first, and the order is gate, SHA alignment, write to disk, ledger note.
• Starting from which skill was never used Sorting by file timestamp to find skills untouched for months returns a directory where everything looks in use, and the audit stops there. No invocation log exists, so the measurable question is loaded or not: diff against the system prompt list, then cut whole families, overlaps and overlong descriptions.
• A conflict hit read as duplication Five high-conflict classes come back and get merged by class, collapsing subtitle generation, subtitle translation and subtitle burn-in into one skill. The script only counts keyword co-occurrence and those are three layers. A first scan of 37 skills found five classes and three real consolidations.

When to use it

Treat skills as assets: audit what is original versus installed, compute the context holding cost, audit long-unused and overlapping skills, and decide keep/merge/archive/uninstall. Use when the skill library bloats, when every turn feels heavier, when you need to know 'which skills did I build myself', or when clearing out skills you do not use. Not for creating a single skill (that is skill-creator).