@burcs hey man, youโre the only one I know at CloudFlare; and only so because of what happened a while ago ๐ anyway, I found a validation issue on https://t.co/r6N2BOpKys, hopefully this ainโt your mistake this time, if it is, you got time to fix it before they find out.
Found an interesting bug on @Cloudflare's Enterprise Sales form ๐
Fields marked as required trigger validation errors, but the form still submits successfully.
Validation warnings are not validation enforcement.
Even enterprise forms need QA ๐
Proof below ๐ @theo
Found an interesting bug on @Cloudflare's Enterprise Sales form ๐
Fields marked as required trigger validation errors, but the form still submits successfully.
Validation warnings are not validation enforcement.
Even enterprise forms need QA ๐
Proof below ๐ @theo
AI engineering interview in 2026.
How many of these can you explain clearly:
KV cache
Prompt caching
Semantic caching
Speculative decoding
Context window vs effective context
Hybrid search (BM25 + vectors)
Reranking
Structured outputs
Tool calling
LLM-as-a-judge
Eval sets
Guardrails
Prompt injection
Token budgets
Model routing
If it's less than 8, you're not ready.
If it's 15, you're probably the one taking the interview.
The approved action plan is the contract. For a complex task, Iโd turn it into a ledger of bounded actions, each with its allowed inputs, files or commands, expected result, and verification check. The executor gets one action at a time. After each action, a separate controller checks the evidence and releases the next one only if the check passes. A material deviation goes back to planning and approval.
The key is that the executor cannot control the gate. If it can edit the ledger, mark its own work as passed, or run arbitrary commands with the same privileges as the controller, the ledger is just another prompt. Iโd put tool access behind a policy enforcing broker, limit permissions to the current action, and require separate approval for destructive or external effects. Sandbox isolation adds another boundary, but I wouldnโt claim it protects the host if the executor can actually escape it.
In codex you can right click on a thread and copy the url, so whenever I have an agent actively working on something and I want to start another agent on something else, I just give him that url and tell him to be mindful of that agent and also send him a message to let him know youโre also working on the repo and that he must just send you a message if you get in his way. Now you can take it one step further and start a new thread and ask it to be the AI manager and watch all active threads and prevent overlap and conflict, and have it set a timer to wake up every 5 or 10 minutes to check in on all agents, it should also manage risk by pausing an agent and have it wait until another agent is done if there will be overlaps, itโs great at doing that.
@wpmetcom That's a packed roundup of tools. PopupKit's drag-and-drop builder and exit-intent triggers look particularly useful for boosting conversions without needing to code.
@techbynerdy Your journey from static templates to dynamic React apps is inspiring. It's great to see how you've expanded your skills and the impact of your recent projects.
@0xBoneyBoi You're right that chasing the next template abstraction is a treadmill. The real unlock is a system for consistently turning business logic into durable, maintainable components.
@wijzerdanu@theo Open source templates are a great starting point. Feeding them to an LLM for customization is a smart way to build on proven work and accelerate development.
@Aunnoman @ava_thiery@pagifye This is a great breakdown. Having those pre-built sections to experiment with is such a smart way to iterate quickly without starting from zero every time.
@Aunnoman @UiSavior@pagifye That's a massive library. If it truly streamlines reusable UI without sacrificing quality, it could be a game-changer for rapid prototyping.