Comparing gstack to my skill stack

Garry Tan posted about his gstack, a Claude Code skill framework, which naturally led to comparing it to my own skills.md stack. Here are the patterns worth noting.

Cognitive mode separation: gstack's core insight is that "planning is not review, review is not shipping."
Each slash command is a distinct "brain" optimised for one phase. We already do this with /dev-plan → /review-plan → /fan-out, but gstack takes it further with CEO review, eng review, QA, ship, and retro as separate modes.

SKILL.md template generation from source code: They have a gen-skill-docs.ts that reads code metadata and fills SKILL.md.tmpl placeholders. Generated docs are committed to git, validated in CI (Continuous Integration). I have an /update-docs skill that keeps the docs up to date with the latest code/plan changes, but Garry's implementation is a more robust workflow for a full-stack app.

Persistent browser daemon: The browse/component is a long-lived Chromium daemon with sub-second latency, @ref system using accessibility trees, and bearer token auth. We have Chrome DevTools MCP already (a new Chrome release uses your existing Chrome profile), so this is less relevant for us.

Three-tier eval system: Free static validation (parse commands, check schema), real Claude sessions via claude -p, and LLM-as-judge scoring. It is cost-conscious but thorough.

/ship skill with automated commit splitting: Automatically creates logical, bisectable commits ordered by dependency (infra → models → controllers → VERSION/CHANGELOG). This is sophisticated and something for us to consider.

/retro skill: Engineering retrospectives with per-person metrics, session analysis, streak tracking. Interesting for identifying areas for improvement. I think we currently ask Claude, "Based on our recent sessions what are the key improvements that we should make to our skills.md?"

Conductor integration: Parallel Claude sessions with isolated workspaces. Similar to our /fan-out but more infrastructure-focused.

Finally, given my existing skills.md (dev-plan, review-plan, fan-out, deep-review , content-draft, content-review, update-docs, rfc-finder, spec-compliance), the most impactful additions would be /retro. It is perhaps the most novel and the hardest to wrap my head around, but it may be worth prototyping against my current projects. The /ship skill is also very tempting. Automated commit splitting and changelog generation would give consistently formatted output across projects.

Will keep you updated!