Case: gstack
Read gstack as an opinionated multi-host workflow harness with specialists, power tools, QA, checkpointing, and release gates.
Key takeaways
- gstack reads a harness as an opinionated software factory: a workflow and command distribution layer, not a loose prompt library.
- Its signals include 23 specialists, 8 power tools, 10 AI coding agent hosts, team-mode auto-update, and checkpoint mode.
- The Think to Plan to Build to Review to Test to Ship to Reflect loop keeps review, test, and ship as separate stages.
- Browser QA and iOS live-device QA verify real behavior, while /learn captures domain knowledge back into the harness.
- Adopt the core loop first rather than deploying the full catalog, since opinionated workflows create friction when they miss the local domain.
gstack is useful because it shows a harness as an opinionated software factory rather than a loose prompt library.
This page is based on the repository state read on 2026-05-23.
Current Positioning
The README frames gstack as a set of specialists and power tools for AI coding agents across multiple hosts. The practical pattern is a workflow and command distribution layer.
Key current signals:
- 23 specialists and 8 power tools.
- Support for 10 AI coding agent hosts.
- Team mode auto-update.
- Browser QA and iOS live-device QA.
- Checkpoint mode.
/learnand domain skill capture./codexsecond-opinion style review.
Workflow
| Stage | Harness role |
|---|---|
| Think | Problem framing and hypothesis |
| Plan | Scope, approach, and execution path |
| Build | Implementation |
| Review | Code, design, security, and second opinion |
| Test | Browser/device verification |
| Ship | Release and docs |
| Reflect | Capture learning and update the harness |
Why This Is Engineering
gstack turns loose agent usage into a pipeline.
- Specialists reduce role ambiguity.
- Power tools make common operations discoverable.
- QA commands make verification explicit.
- Checkpoints make long work recoverable.
- Team mode makes updates distributable.
- Domain learning pushes local knowledge back into the harness.
What to Borrow
| Borrow | How to apply |
|---|---|
| Opinionated stages | Name the stages your team actually uses |
| Review/test/ship separation | Do not collapse all validation into implementation |
| Browser and device QA | Verify real behavior, not only code |
| Team distribution | Manage commands and skills as shared assets |
| Checkpoint and learning | Store progress and convert repeated patterns into skills |
Caveats
- Opinionated workflows are fast when they fit the team.
- They create friction when they do not match the local domain.
- Do not deploy the full catalog at once; start with the core loop.
References
- gstack README, read baseline 2026-05-23 https://github.com/garrytan/gstack