All 94 skillsQuality8 of 13
stryx:optimize
When a number should move and nobody knows which change wins: try, measure, keep the winners.
/stryx:optimizeInside its piece: 22 specks, one for each file on its shelf. Its twelve group-mates glow on the open book behind it. The light splits toward orange, the band of Quality.
In numbers
counted 2026-10-12 in the installed set- 10reference files
- 4helper scripts
- 22files in the skill
- 213KB on disk
- 6chapters
- 1agent works by this book
Use it when
in the skill’s own words- when a working system has a number that should move and the winning change is unknown
- so candidate changes
- parameters
- prompts or strategies must be tried against one repeatable measurement
- keeping only confirmed winners
- Triggers
- "optimize this against the benchmark"
- "run optimization experiments"
- "tune these parameters"
- "find the smallest memory limit that stops the OOM"
- "try approaches and keep what measurably helps"
- "cut test-suite or CI wall time without raising runner-minutes"
- "make this prompt cheaper but still good enough"
- clustering
- search relevance or summary quality scored by an LLM judge
- "run it overnight"
- an optimization spec YAML
- resuming an optimize/<name> branch
Not for
another skill does it- One known bottleneck→ stryx:performance
- An unknown cause→ stryx:debug
- A decided change→ stryx:execute
- A one-round race of 2-4 implementations→ stryx:orchestrate
- Competing drafts judged→ stryx:bakeoff
- Writing or running an eval suite for a prompt or model→ stryx:eval
On its shelf
22 files, 213 KB- references/10
- data/5
- scripts/4
- top level2
- evals/1
Inside the book
6 chapters, by title only- The rules that make the numbers trustworthy
- Core procedure
- Gates
- Stopping criteria
- Commands
- Talking to the user
Its helper
1 agent works by this book- stryx:perf-agentWhen an app needs a speed audit, or one named thing must get measurably faster.
Works with
the skills it names most- stryx:debug→Something is broken: a bug, an error, a crash, a failing test or wrong data.
- stryx:eval→When you must measure how a prompt, an agent or an AI feature behaves.
- stryx:orchestrate→When a job could use more than one agent: how to split it, staff it and merge the results.
- stryx:performance→Something is slow, laggy, too big or uses too much memory.
- stryx:ai-power→When you are not sure which skill fits, or you say "load aipower": it is the map of all skills.
- stryx:bakeoff→When two or more very different approaches must each be drafted before you can choose.
Claude loads it by itself when the moment comes; by hand, type the line at the top. It comes with the whole set: Installs all 94 skills, 32 agents and the brain program as one set. The skill’s own text is private source and is not shown here.