Which agents work with it?

Any agent that reads skills from a directory. Claude Code, Codex and Gemini CLI read it natively; Cursor and VS Code connect over MCP.

The voice runs on your own machine, so there is nothing to meter. The first run downloads it, about 340 MB.

Yes. The skill is MIT-licensed and the authoring format is public.

gstack 1.91.157.0 cuts 20 KB of dead routing text from its Claude Code skills

50 minutes ago

gstack 1.91.157.0 for Claude Code: the "When to invoke" section is gone from every Claude skill file, about 20 KB across 56 skills, and voice triggers now route from the root router. Other hosts and --catalog-mode=full are unchanged. Release notes: https://github.com/garrytan/gstack/blob/main/CHANGELOG.md#1911570---2026-10-19 Commits: https://github.com/garrytan/gstack/commits/main

Ask

Ask about this presentation

Answers are generated from this presentation.

Show transcript

Lighter

gstack drops the “When to invoke” section from its Claude skills

## When to invokeremoved from every Claude SKILL.md

savedabout 20 KB across 56 skills

LighterWhyVoiceOther hostsYour turn

garrytan/gstack · CHANGELOG.md · 1.91.157.0

gstack's Claude Code skills just got lighter. Every skill file loses its When to invoke section. That cuts about 20 kilobytes of text across 56 skills.

Why

Claude Code opens a skill’s body only after invoking it, so that section never routed anything

Routing still comes from the one-line catalog description

and from the routing block in your CLAUDE.md

LighterWhyVoiceOther hostsYour turn

garrytan/gstack · CHANGELOG.md · 1.91.157.0

Claude Code only opens a skill's body after the skill is invoked, so routing text in there never picked a skill. Choosing still works the same way, from the catalog description and your CLAUDE.md routing block.

Voice

Voice triggers now live in the root router, one line per skill

- Voice trigger “speed test” → invoke /benchmark

21 lines in the root router, while 54 skills shed 5,139 tokens

LighterWhyVoiceOther hostsYour turn

garrytan/gstack · CHANGELOG.md · SKILL.md · 1.91.157.0

Spoken triggers moved too. The root router now holds one voice trigger line per skill, 21 in all. The router grows a little, and the skills together shed about five thousand tokens.

Other hosts

Only the Claude render changes

Other hosts keep the full description

--catalog-mode=full keeps it too

LighterWhyVoiceOther hostsYour turn

garrytan/gstack · CHANGELOG.md · 1.91.157.0

This trim is for Claude only. Other hosts keep the full skill description, and so does the full catalog mode. If you run gstack somewhere else, nothing changes for you.

Your turn

The full notes are in gstack’s changelog

CHANGELOG.mdgstack 1.91.157.0, October 19

Do you call gstack skills by voice, or by typing the command?

LighterWhyVoiceOther hostsYour turn

garrytan/gstack · CHANGELOG.md · 1.91.157.0

The details are in gstack's changelog. Do you call gstack skills by voice, or by typing the command?

gstack 1.91.157.0 cuts 20 KB of dead routing text from its Claude Code skills

Ask about this presentation

Chapters

Lighter

gstack drops the “When to invoke” section from its Claude skills

Why

Claude Code opens a skill’s body only after invoking it, so that section never routed anything

Voice

Voice triggers now live in the root router, one line per skill

Other hosts

Only the Claude render changes

Your turn

The full notes are in gstack’s changelog

Up next

gstack 1.91.158.0 lets you switch off the plan-mode preamble in every skill

gstack 1.91.156.0 gets the weekly /office-hours test running again, with no change to the skill

gstack's browse stops locking up on a busy tab, and /cso keeps a committed .env out of its history

gstack's /ship stops passing reviews that found bugs, and /cso runs on repos with submodules

gstack 1.91.150.0: /setup-browser-cookies now asks every question before it touches your browser

gstack now runs its eval cases on Codex every Monday, beside Claude Opus 5.5

gstack 1.91.146.0 merge trains let several ready PRs share one full eval gate

gstack 1.91.145.0 stops failing Opus 5.5 for harmless file reads in its /qa-only and docs-sync tests

gstack 1.91.144.0 now tests every skill change on Claude Opus 5.5, at a third of Fable's cost

gstack 1.91.143.0: docs-sync tests stop failing Claude's harmless joined reads

gstack's daily /cso report stops listing security findings it can't prove

gstack 1.91.138.0: its cheap PR screen lets fewer bad changes through than planned