Draft — personal notes on working with AI coding agents. Not yet ready for wide circulation.
Note: A lot of this being captured in srid/agency
Make it research a lot (to counteract hallucination during planning)
Split into multiple phases (easier to review & refactor)
Then, execute one phase at a time.
See my /plan command
For AI autonomy we need strong feedback loop from e2e tests and local CI
for bug fixes, create a failing test first, then fix it.
parallelize for faster loop: juspay/kolu#69
avoid or minimize flaky tests (timeouts, etc.)
Use unique ports (to avoid port conflict with multiple AI sessions)1
See my /do command
Ralph loop for refactoring — iterative refinement passes that simplify and beautify code. Good for structural improvements.
Review fully, leave no stone unturned — read every line the AI generates. Understand it all.
Including architecture level review.
Use AI to learn the code it generates — the code AI writes is code you otherwise wouldn't have written. Treat it as a learning opportunity, not a black box.
Create a "skill" out of it - to help future projects
To Organize
Plan -> Probe (grok )
Parallel AI sessions
Review overhead
Early prototype before planned implementation
Deterministic orchestration
End-to-end autonomy challenges
/srid-do doesn't always run /elegance and /hickey
/code-police (after every change)
CI
Stops when CI fails sometimes ("flaky")
Needs fine-grained CI faiure feeback for parallel fix.
Log all the above to a .ai/session/pr-NNN.md file for later audit
Need to re-run only failed CI steps (not whole)