how I used to write prompts for a coding session:
- describe the whole feature
- describe the edge cases
- describe what not to do
- add examples
now:
- write a spec file before I open Claude
- one task per session
the difference: actually finishing things
@ksohit i think blast radius scoping is the right call... the hard part is the dependency graph. if your test runner doesn't know what a change actually touches, it ends up running everything anyway
@kamilkwapiszpl yeah the "add AI to the existing mess" pitch is the easy sell. the redesign question is harder to get a yes on but it's the only one that actually changes anything
@PatrickOjo_ the audit trail problem is the hard one. git history made sense when each commit was a human decision. when an agent commits 500 times, what's the unit of review?
Looking to connect with founders doing AEO or AI SEO right now, not just reading about it.
Specifically:
- tracking GSC impression/click gaps weekly
- writing for specific queries, not broad topics
- have one thing that worked that surprised you
@Dillon_Carter true, the sop just forces you to write down what you were doing in your head. when the agent gets it wrong it usually means you hadn't written it down clearly enough
@samrexford github codespaces + claude code is underrated.. remote context without the VPS headache. did you wire it with the https://t.co/tkVJDwDMkn remote or via SSH?
@stretchcloud most orchestration problems are actually retrieval problems in disguise. the agent looks broken but the knowledge layer just can't answer the question it's being asked
genuine question for solo founders: at what point did your first 10 users stop feeling like an accident?
not when you hit 10. after. when you believed someone was choosing you, not stumbling across you.
@pauliusztin_ the split makes sense. AGENTS.md is policy, MEMORY.md is experience. the failure mode is when the agent starts writing to MEMORY.md things that should have been policy from the start
@heyrimsha the catch is real, raw chain of thought isn't the ground truth anyway. but "trust us, we checked" has never been a sustainable safety argument in any other industry either