In my last article, a reader's comment exposed a hole in SKILLmama's design, and I shipped Phase 1.5 to fix it. This time, nobody had to point anything out to me. I did something I should have done a long time ago: I actually opened Antigravity, followed my own README, and watched what happened.
It didn't work.
The Recording
SKILLmama supports four agents — Claude Code, Claude.ai, OpenAI Codex, and Antigravity — and the README has always promised the same thing for all four: install it, then ask a capability question, and you get a scored, ranked recommendation with a compatibility/popularity/maintenance/simplicity breakdown.
I had never actually verified that claim for Antigravity. So I opened it, typed:
find me the best vector database for a Python project
And got back a plain, generic answer. No stack scan. No constraint question. No scoring table. Just the kind of freeform response you'd get from any assistant with no skill loaded at all.
SKILLmama wasn't running. It wasn't even trying.
The Wrong Guesses
My first assumption was reasonable: the npx skills add CLI I'd been telling people to use must be installing to a path Antigravity doesn't check. I went and read the CLI's own source (vercel-labs/skills) to find the real path. Found what looked like the answer. Copied the skill file there by hand. Restarted Antigravity. Tested again.
Still nothing. Same generic answer.
Second guess. Maybe I had the wrong directory entirely. I tried another path that seemed plausible from the CLI's agent config. Restarted again. Tested again.
Still nothing.
Two wrong guesses in a row are usually a sign you're debugging the wrong layer. I was reverse-engineering a tool that manages the skill, when I should have been reading what the agent itself actually documents about where it looks.
The Real Answer Was in the Docs I Hadn't Read
antigravity.google/docs/skills — the actual, official documentation for how Antigravity discovers skills. I hadn't opened it once before this.
It states, plainly, two locations:
Project scope: <workspace-root>/.agents/skills/<skill-folder>/
Global scope: ~/.gemini/config/skills/<skill-folder>/
Neither of my two guesses matched. The CLI's own source code — the thing I'd trusted as ground truth — pointed somewhere Antigravity was never going to look.
I copied the skill to the real path, restarted Antigravity one more time, and asked it directly: "Which skills are installed?"
SKILLmama — AI-Native Capability Discovery Engine
There it was, listed under Global & Built-in Skills, sitting right next to Antigravity's own native commands. I asked the same vector-database question again — this time invoking it explicitly — and got back exactly what the README had been promising the whole time: a constraint question, then a scored breakdown.
It worked. It had just never worked the way I'd been telling people to install it.
Closing the Loop
Once I had a real repro, I checked whether anyone else had already run into this — and found an open issue on vercel-labs/skills asking for the exact fix Antigravity actually needs. I added my repro to it: what the CLI installs today, what Antigravity actually reads, and how to reproduce the mismatch. Not my repo to fix, but worth leaving evidence for whoever does.
The README now leads with the manual install to the correct path — confirmed, not assumed — and only mentions the CLI as a secondary option with an explicit warning about where it currently goes wrong.
Bonus: While I Was In There
Chasing down the Antigravity bug surfaced two more things worth fixing in the same pass:
Scoring got more honest. The Compatibility score used to be inferred purely from your detected stack. Now, before scoring, SKILLmama checks your actual environment — does the required env var exist in .env.example? Is the CLI on your PATH? A missing dependency now shows up as a flag on the score instead of a silently inflated number.
The four adapters had quietly drifted apart. codex/AGENTS.md and antigravity/PROMPT.md still carried leftover wording from an earlier two-stage architecture that had since been dropped — meaning the instructions Codex and Antigravity were actually running had subtly diverged from the canonical pipeline in skillmama/SKILL.md. Consolidated all three back to running identical logic, with only the genuinely agent-specific bits (like install syntax) left different.
What Changed
All of this is live in v1.4.3. For Antigravity specifically:
mkdir -p ~/.gemini/config/skills/skillmama
curl -sL https://raw.githubusercontent.com/Magithar/SKILLmama/main/skillmama/SKILL.md \
-o ~/.gemini/config/skills/skillmama/SKILL.md
Restart Antigravity, then invoke it explicitly: SKILLmama find me a vector database for this project.
github.com/Magithar/SKILLmama — Apache 2.0.
The lesson wasn't really about Antigravity. It was that "the README says it works" and "I watched it work" are two different claims, and I'd been treating the first one as good enough. It isn't. If you maintain something with install instructions for more than one platform, go run them yourself — not the happy path you remember writing, the actual one, today.
Top comments (0)