Lesson 09 | GSD Q&A (Part 3): Practical Tips & Troubleshooting
Chapter 3: Practical Tips & Troubleshooting
What to do when things go wrong, how to onboard existing projects, and daily usage tips.
Q15: How do I roll back execution errors? Can GSD undo?
GSD has no dedicated "undo" command. But atomic commits make git rollbacks easy:
# View recent commits
git log --oneline -10
# Revert to a specific commit (keeping workspace changes)
git reset --soft <commit-hash>
# Revert to a specific commit (discarding all changes)
git reset --hard <commit-hash>
# Revert just one specific commit
git revert <commit-hash>
Each Task is an independent commit, so you can precisely roll back to just before the failing Task.
Recommendation: Manually git tag before major Phases begin, e.g., git tag before-phase-3. To roll back, simply git reset --hard before-phase-3.
Q16: What if AI code quality is poor? Can I ask it to rewrite?
Yes. Three ways:
/gsd-code-review+/gsd-code-review-fix: Auto review + auto fix. The review report categorizes by severity, and the fix agent only modifies problematic areas.- Modify directly: Use
/gsd-quickto tell AI what is wrong and how to fix it. Or fix it manually. - Re-execute:
git revertthe problematic commit, then re-run/gsd-execute-phase --wave Nfor that specific wave.
Pro-tip: Don't wait until the entire Phase finishes to check the code. Use --interactive mode to check after every Task.
Q17: What if the research agent fails? Can I fill it in manually?
Research agents can fail due to network issues or API limits (which happened in this tutorial project).
How to respond:
- Retry: Simply re-run
/gsd-new-project, and the research phase will try again. - Fill manually: Do your own research and write the results into the
.planning/research/directory. Formatting doesn't matter much; AI can read Markdown. - Skip: Use
--skip-researchto skip research and jump straight to the requirements phase. Your own domain knowledge will suffice.
Failed research doesn't block the rest of the flow. Requirements can be written or supplemented manually.
Q18: How do I introduce GSD to an existing project? Do I need to refactor?
No refactoring needed. Two-step integration:
- Create a
CLAUDE.mdin the project root outlining constraints. - Run
/gsd-new-project, and GSD will read existing code to generate plans.
GSD won't modify existing code. It only generates planning files in .planning/. Subsequent Phases can incrementally add features.
Note: Existing projects might require you to supplement the requirements document. GSD tries to infer requirements from code analysis, but it may not be 100% accurate. Manual checking of REQUIREMENTS.md is recommended.
Q19: GSD generates too many files, can I slim down .planning/?
Yes. Files in .planning/ fall into two categories:
Core Files (Do not delete):
STATE.md— Project state, essential for restoring progress in new sessionsROADMAP.md— Roadmap, essential for executing PhasesREQUIREMENTS.md— Requirements, essential for Verificationphases/*/CONTEXT.md— Frozen decisions, downstream Phases rely on this
Deletable Files (Deleting won't affect functionality):
research/— Research outputs, delete after decisions are madephases/*/DISCUSSION-LOG.md— Discussion logs, purely for referencephases/*/VERIFICATION.md— Verification reports, can be deleted after the Phase passesphases/*/REVIEW.md— Review reports, can be deleted after fixes
Pro-tip: Once a Phase passes verification, keep only PLAN.md and CONTEXT.md. Delete research, discussion, and verification files. Always keep STATE.md.
Q20: What is the difference between solo use and team collaboration with GSD?
Solo Use (This tutorial scenario):
- GSD replaces the discipline of "Plan → Execute → Check"
- AI acts out all roles (Product/Design/Dev/QA)
.planning/is your personal memo
Team Collaboration:
- The
.planning/directory can be committed to git so team members share context CONTEXT.mdacts as a design document for team reviewPLAN.mdacts as task allocation; multiple people can execute different Phases in parallelVERIFICATION.mdserves as acceptance criteriaREVIEW.mdserves as the Code Review log
Core difference: Solo, GSD means "AI helps you manage the project." In a team, GSD means "AI helps the team align on context."
Tutorial Complete! Full series 14 episodes + 4 appendices, 6 acts, and 20 Q&A.