Comment by apatheticonion | Hacker News Reader

Comment by apatheticonion | Hacker News Reader

I see a large emphasis placed on headless agentic/vibe coding, what I don't see people talking about is how great guided coding is.

I have over a 15 YoE writing software and guided coding sessions - that is, using an editor like Zed or VSCode with an LLM integrated, writing code how you normally would but using a flash model to prompt away the annoying parts and/or plan - is as productive as vibe coding, produces significantly higher quality, is actually enjoyable, and you actually stay sharp.

Flash models (DeepSeek v4 flash) tend to be so fast that you don't have time for parallel agents, you lock in and rapid fire prompts, building high quality software while incrementally reviewing it as you go. VETO bad edits and try again or rewrite them manually.

By contrast, I have noticed headless agentic coding tends to be an unreviewable black box. The major issues I've found is that, even with a human-in-the-loop, you accumulate defects which compound and eventually you're spending millions of tokens to make trivial changes in a ridgid codebase.

Ultimately, tiny, highly cached, fast models like Qwen's 27b/a3b range or DeepSeek flash are highly capable and relatively inexpensive to run. Hoping people realise we don't need 14 trillion parameter models and I'll be able to buy some ram for my workstation

This is EXACTLY how I am working with these tools, you summarized everything about why I use this approach perfectly. Zed + OpenCode + Mimo/GLM/Kimi models are incredible. I haven't touched a "frontier model" for months.

I really treat them as "delegation layer". They're systems that sit right alongside your stack, and at this point, at every level in it, like a connective tissue that intersects at all points. I can delegate a spec, a function, a file, a whole feature set, an app, a debugging loop, etc.. My day-to-day isn't all that different, it's just like having the "Ship's Computer" that I can ping anytime I need something. I'm enjoying the benefits that a smart data processing pattern interpolator provides, but I'm not giving up any personal control or agency. My goal isn't to go faster, nor is it to avoid coding, because coding isn't drudgery or below us just because we have new code generation tools.

I had a job interview at Canva where they had an "AI coding round". They wanted to see how I used AI in my workflow.

Because I was using DeepSeek flash, the interviewer was confused at how fast I was getting edits - he'd only used Claude and was used to it thinking for 40 minutes to update a function.

We were building a frontend application and I was able to do that whole guided prompt/review loop. I finished the task with no defects (high quality code) and way ahead of time.

I could tell the questions he asked were prepped for vibe coding with slow frontier models, like "how do you review the code to ensure it is safe to commit" - "make small/reviewable edits and read the code the LLM emits?".

Despite meeting the criteria of a principal engineer in the other rounds, the feedback on the AI round was that "I didn't delegate enough to the AI and was too hands on" lol.

> the feedback on the AI round was that "I didn't delegate enough to the AI and was too hands on" lol.

the real feedback they couldn't say is that you were a threat to his position.

So frustrating. Too many hiring teams seeking the "perfect" candidate instead of hiring for completely valid and amazing potential.

I'd wager you're far better than whatever candidate they landed on, because the person who apparently qualified for that role probably isn't thinking while they prompt away.

Why are engineering teams looking for brainless monkeys? It's to their detriment all around if they aren't hiring capable people.

Did that AI round feedback scuttle a potential offer?

Unfortunately they decided not to proceed despite glowing feedback in the other rounds

I've gotten immense benefit from "guided coding". I've learned a lot, and it's saved me a lot of time digging through documentation. I also don't have anyone dictating to me how many tokens I should be burning, and I'm not forced to review my co-worker's 3000 line PR.

I agree with your last statement. Frontier models are nonsense. We should be concentrating on focused models that are cheap to run, and don't require destroying the SBC industry, and making my hobbies financially out of reach.

Headless agentic coding only works when:

* The scope is constrained and clearly defined

* You articulate an exhaustive and objective set of success criteria

* You provide a deterministic tool for the AI to evaluate its output against those criteria and receive feedback (compiler errors, unit tests, regression tests, etc.)

You're essentially writing a comprehensive spec for the AI to follow. The time suck was always from making the spec and the QA bulletproof.

Even then, it tends to code itself into a corner. Code tends to be ridged and difficult to expand.

One shot benchmarks that have exhaustive prompts are impressive, but benchmarks that emulate real software development (iterative changes over time with changing requirements) have close-to-zero pass rates.

Well yes, if the thing you're working on is subject to changes in requirements, then you fundamentally can't one-shot it, AI or no AI. That's why the waterfall process fell out of fashion long ago.

But there are cases where there are no changes, or only backwards-compatible changes. Take the core WebAssembly specification, for example. Extremely detailed, formally verified, comes with a large test suite. Someone else has done the hard work of specifying the behavior of every possible edge case. A WASM runtime implementer can point their agent to that spec, plus some prior art. The remaining manual work is in constructing a test harness and collecting a corpus of test WASM binaries that exercises decent coverage.

Agreed. I think VS Code + Codex or similar is great, personally. I never (1) hit any kind of limits with the $20 plan or (2) lose control of the code base whatsoever. I'm also not shipping at 300x or whatever some other folks claim.

In my comment yesterday i mentioned about "inline autocomplete" is the best way to use LLM in coding, you still lay the foundation, meanwhile the LLM follows you and finish the detail. You are step-locking with the LLM.

Reply on news.ycombinator.com