Senior AQA / SDET π·πΈπͺπΊπΊπΈ Playwright, API testing, AI u testiranju. Gradim B2B app za veleprodaje na Balkanu. SR-EN. π Open to remote projects.
@DavidTranAI@NotionHQ One bot owning the page is the smart part. I run a few role-based bots and the moment two of them write to the same notes it gets messy π Do the others just read it, or do they send updates through the primary one?
@digitalfoundry Respect for re-testing in public. A ~10fps swing between 5.00c and 5.01 is exactly why the patch version should sit right next to every benchmark score.
@robwalling Building got cheap, finding people who'll actually pay didn't. Curious if the book covers B2B niches where the buyers aren't on X or Reddit at all.
@kentcdodds The reaper agent is the scary one. Writing code is cheap now, proving something is actually dead isn't. I'd want coverage data and prod usage logs in front of it before it deletes anything.
@DaveShapi The prompt is more of a suggestion than a contract. Anything that really matters I'd put in a check the agent can't talk its way past, like a failing test or a lint rule. Did repeating the rule at the end of the task help at all?
@jacobAlswin Distribution for small B2B SaaS is still mostly manual. Catalogs, invoices, and getting in front of the right niche eat more time than building the product itself.
@debs_obrien@bot Tiny Worker glue between Alexa and a chat agent is oddly more reliable than one mega-bot trying to do everything. Did the calendar query need any auth dance on the Alexa side?
@mattpocockuk "Automated checks are too lax" is the one that bites. Agents will greenlight almost anything if CI only runs unit tests. Once we failed the PR when the Playwright smoke didn't pass, agent output quality jumped overnight.
@nicky_sap The tsc-as-its-own-job trick is underrated. We've had suites where Jest was type-checking the same files across shards. Splitting that out cut our PR gate almost in half before we even touched shard count.
@r_davidjones@NotionHQ The Ready for Bots column is a nice trigger. How do your bots handle a card that's underspecified, do they ask in the card or just bounce it back to Ideas?
@DavidTranAI@NotionHQ Which of the seven did you think has the lowest setup cost? Curious how you'd keep the shared memory page tidy once several agents write to it.
@NotionHQ My setup is a few Grok bots with their own roles (QA, dev, product) checking each other's output, then I write the Claude Code prompt and test the result myself. The handoff notes are the part I'd love to keep somewhere like this.
@Nanjialonguazt@edinsoncode Agreed, retention is the real test. Building a small B2B SaaS, the boring stuff like onboarding and billing edge cases decides who stays more than any new feature.
@StephenVivash The spec-first approach is where it clicks for me too. When the acceptance criteria are written down, you can turn them straight into Playwright tests and the AI output gets checked against something real.
@levelsio Running a whole OS inside a tweet is a nice reminder that the browser is the ultimate test target. I'd love to see the Playwright trace for that thing.