I'm going to test it out, thanks Elio!

Huntley's article: https://ghuntley.com/replaced/

Show Your Work! by Austin Kleon, a short book I reread regularly: https://austinkleon.com/show-your-work/

What I share as open source: https://github.com/IgorWnek

Then show your work. Publish the repo, write a short post, share what broke.

One piece of advice: this week, build a small proof of concept you've always wanted, with a coding agent. Then publish it, together with what you learned.

Pay for AI yourself and use it on your own code. If your company doesn't allow AI, keep your subscription away from its code and your clients' code.

A $20 plan is enough to start. Treat it as an investment, not a cost.

Using AI at work isn't enough. Build the thing you've always wanted to build. With coding agents, that takes evenings, not months, and that experience is what makes you valuable at work.

Doing more has consequences too: far more code to review and a very different way of working. You'll discover that along the way, and it's normal. More on that in future posts.

I still know engineers who say they won't use AI. AI won't replace them. Engineers who use it every day will.

Without AI, the work is too slow compared to theirs. They try more ideas, ship more, and still have time to check what they ship.

GPT-6 Luna costs the same for shorter prompts. Anthropic's table puts Haiku 5.5 ahead, 39.2% vs 16.4% on Terminal-Bench 4.0.

My guess: GPT-6.1 Luna lands tomorrow. GPT-6.1 Sol came the day after Sonnet 5.5.

https://www.anthropic.com/claude-haiku-5-5

Same update: Sonnet 5.5 cache reads drop from $0.20 to $0.10 per million tokens. Anthropic puts the saving at around 20% on most agentic work, in line with my own estimate.

With Haiku taking the right subagent work, I expect my Claude Max plans to go much further.

I expect it to earn a place as a subagent on well-specified work: implementing a plan, triage in code review, codebase research. Sonnet and Opus keep complex agentic coding.

Prompting guide tip: at low and medium effort it sometimes reports a change as done without running a check. Name the check.

I leave room for another provider or tool to earn a role. Argon leads the broader Vals Index; task mix matters. I'd compare accepted results, review time, failures, and cost on my own work.

Terminal comparison:
https://www.vals.ai/benchmarks/terminal-bench-4

Vals' Mini-SWE-agent terminal runs put Argon behind Opus, Sonnet, and Astra. It scores above Sol 6.1, but costs about 10x as much per test at the published API rates. Some Claude tasks used fallbacks.

That doesn't establish value under my subscriptions.

And that's a clever solution. 👍

3/3 The proactive move is counterintuitive: shrink the Concern circle, expand the Influence one. AI engineer: stop chasing releases, ship the eval. Product engineer: stop reading the roadmap, talk to five users. Principal: pick the two places where you actually move the system. Sick! 💪🏻

Covey's distinction: Circle of Concern (everything you care about) vs Circle of Influence (what you can actually affect). Most engineers I know - drown in the Concern circle. Model news, org politics, every team's stack choices.

Awesome news! Thanks! 🚀