code-assist

The Tool Overtriggering in Anthropic’s Opus 4.5 migration skill notes why I like less markdown than more. The optimized markdown you have right now is optimized only for the model that you optimized it for.

Code is cheap. Show me the talk argues that the ability to think clearly, communicate problems, and architect solutions matters far more than the ability to write code itself.

OpenCode appears to be pretty interesting and could be worth investigating further. Even though Anthropic blocked Claude subscriptions on it, afterwards GitHub Copilot announced official support for it, so you can still access Opus via Copilot credits. At first glance I was able to copy my ~/.claude/commands to ~/.opencode/commands and they appeared to just work (with Opus via Copilot). The larger benefit of opencode would be testing other models that Claude Code does not support, notably local models.

Addy Osmani’s LLM coding workflow going into 2026 appears very similar to my own, though an interesting note on articles like this is that the reader can interpret the language used in different ways. Claude Code has an excellent /plan capability that Opus 4.5 excels on. I do plan, and I do write specs from time to time, but I also one-shot and two-shot changes and my actual development looks like a mixture of these. At no time do i generate pages and pages of specs and then have the LLM embark on a long horizon delivery period and not look at the code it wrote. Addy points out in anther post the importance of understanding the code with ideas like:

If you skip review, you don’t eliminate work - you defer it

And also links to another article noting that skipping review results in:

No consistency, no overarching plan. It’s like I’d asked 10 junior-mid developers to work on this codebase, with no Git access, locking them in a room without seeing what the other 9 were doing

This is inline with my findings and against a popular (hype) argument that an increase in code volume due to LLMs needs to be accompanied with a decrease in human-in-the-loop because speed. This is a drop in quality, which is deferred until your customer’s are affected and your DORA metrics deteriorate.

I’m finding promise in having agents use a CLI directly rather than having them generate API code. This tweet elaborates on the context window and efficiency improvements.

How to Pair With an Agent is a short but good example of to make your prompting less vague.

knip discovers typescript files that are present but not referenced. I find this handy during code-assist so the agent can clean up stuff it doesn’t need anymore.

Microsoft announces auto-model selection for VS Code that prioritizes Claude Sonnet 4 over GPT-5 for paid users

Chain-of-Vibes by Pete Hodgson sums up nicely the approach I take to leverage the power of agentic AI workflows while sidestepping the limitations. Treat the tools “like weirdly knowledgeable, hyper-productive junior engineers”, give enough context but otherwise limit their output. Importantly, review the outcome of each small task and make a decision:

A) Accept and commit
B) Prompt to adjust
C) Manually fix up
D) Revert and re-prompt

The last item is an important one that’s often overlooked - throwing away the result and fixing your prompt is often more time efficient, especially when learning how to prompt.

A rules-based pattern is emerging for helping agentic workflows produce better results. Examples include GreatScottyMac’s RooFlow and, Geoff Huntley’s specs and stdlib approaches.

Brendan Humphrey on Vibe Coding, aligns with my own thinking on vibe coding:

…these tools must be carefully supervised by skilled engineers, particularly for production tasks. Engineers need to guide, assess, correct, and ultimately own the output as if they had written every line themselves.

Smashing Create PR with vibe coding output amounts to an attack on the PR process:

Generating vast amounts of code from single prompts effectively DoS attacks reviewers, overwhelming their capacity for meaningful assessment

But there is still some value:

Currently we see one narrow use case where vibe coding is exciting: spikes, proofs of concept, and prototypes. These are always throwaway code. LLM-assisted generation offers enormous value in rapidly testing and validating ideas with implementations we will ultimately discard.

GitHub Copilot: The agent awakens - Just when you thought it was Cursor/Claude Desktop/Roo/Cline, m$ reminds you they’ll eat your lunch.

The “First AI Software Engineer” Is Bungling the Vast Majority of Tasks It’s Asked to Do - It took longer than a human, and failed at the vast majority of tasks.

Out of 20 tasks we attempted, we saw 14 failures, three inconclusive results, and just three successes,” the researchers found — a meager success rate of just 15 percent.

Today we’re excited to introduce Devin, the first AI software engineer.

Devin is an autonomous agent that solves engineering tasks through the use of its own shell, code editor, and web browser.

gpt-engineer - A CLI tool that allows users to specify software in natural language and have AI write and execute the code, with capabilities for code generation experiments and improvements.