Check fails
Implement→Check
↩ return
Back to Implement, then Check again.
Workflow
How I build with AI, from idea to merge.
AI helps across the development lifecycle. Scope, decisions, implementation, verification, review, and completion remain visible and controlled.
1. The workflow
1. Specify
Define problem, goals, constraints, acceptance criteria, and likely risks.
2. Implement
Make source changes in small reviewable slices.
3. Check
Gather behavioral evidence against acceptance criteria.
4. Review
Inspect quality, drift, maintainability, risk, and consistency.
5. Complete
Handle approved commit, PR, merge, archive, and cleanup actions.
2. Repair loops
Review requests changes
Implement→Check→Review
Return to Implement, then re-check, and get a fresh Review.
State or environment is unclear
Route to status or diagnostic check. Resolve the issue before proceeding.
3. Core principles
Files outlive conversations
Durable artifacts carry the work forward.
One phase owns each concern
Boundaries keep focus and state clear.
Evidence beats confidence
Show results. Assume nothing. Verify.
Verification should match risk
Right depth, right cost, right time.
Humans keep consequential control
Approval gates before actions that matter.
Trivial work stays trivial
Use simpler flows when risk is low.
4. Where AI helps, where I stay responsible
AI helps with
- framing assumptions and questions
- surfacing edge cases and alternatives
- exploring codebase context
- drafting specs and plans
- scaffolding implementation slices
- review assistance and summaries
I stay responsible for
- scope and success criteria
- architecture and trade-offs
- correctness, security, and performance judgment
- user and product decisions
- approvals for push, PR, merge, and cleanup
- final accountability
5. How it scales with risk
Trivial work
edit → suitable verification → summarize
Small tracked change
specify → implement → check → complete
Non-trivial change
specify → approval → implement → check → review → complete
Not every change deserves the same ceremony.
How this shows up in practice
Loreo
I mapped ingestion stages, failure states, and recovery paths before implementation, then verified the importer against a real 4,405-link archive.
PlanningIngestionVerification
DevCache
I used the workflow to turn a broad knowledge-management idea into scoped slices, test architecture options, and keep implementation decisions documented as the product evolved.
ArchitectureDecisionsIteration
A note on tooling
The exact prompts, commands, and internal setup I use are private. What matters publicly is the methodology: clear phases, durable context, verification, and explicit control.
