Victor

Vector embedder skill. Treats meaning as geometry, and tells you when geometry is the wrong tool.

npm run try

That prints two numbers for the same pair of strings:

left:    attribution for the lattice
right:   credited as Skye Wood
overlap: 0.0000   (shared tokens / max size)
hashed:  0.4113   (cosine of 64-d hashed 3-grams)

Zero shared words, and still related. That gap is the entire argument for reaching past grep. When the gap is not there, grep already won.

The part most embedding advice skips

Victor's first instruction is to not use Victor.

Always run a non-vector baseline. If overlap already ranks it, you did not need a model.

And before that:

Name the strings. If they must match exactly, stop. Use the string.

So: file paths, symbol names, ids, hashes, yes/no flags, secrets. Do not embed any of them. An exact match is a string comparison, and dressing it up as geometry makes it slower and less correct at the same time.

Measured case, from the session this skill came out of: git grep over a 581-file tree returns in 52 ms. Nothing about that wants an index. The place a vector earns its cost is the pair above, where the words do not match and the meaning does.

Ponytail order

Brute cosine over an array in memory. Then FAISS or USearch. Then a store. Skip the store if an array works.

The skill lists what actually makes and searches vectors, with the honest note attached to each: paid APIs are a model behind a wall, mixed models are mixed worlds, mixed dimensions are a crash, and an unnormalized dot product is not cosine no matter what you call the variable.

Layout

.cursor/skills/victor/SKILL.md    the skill
.cursor/skills/victor/victor.mjs  the probe. No install, no model, stdlib only.
tests/victor.test.js              four checks. npm test

victor.mjs is a hashed bag of character 3-grams. It is vector shaped and it is not meaning. It exists so you can see cosine behave before you download 384 dimensions of anything.

Sibling

ponytail-neckbeard is the laziness skill. Victor follows the same ladder: the cheapest thing that answers the question, and a runnable check left behind.

Near is near. That is all it knows.

MIT.