[AIN#30] Conductor refused to diff that token spree

This is a real PR shipping real value, hopefully next week:

Conductor is like, "What does it matter at this point?".

And I'm confused why I'm running out of tokens. 😅

So far, I've been trying to max out on best models and thinking levels. This week I ran out of Claude tokens on Thursday, and my Codex subscription is behind schedule, too. Next week I'm putting more effort into switching models and efforts. Higher thinking levels allow for more context and more operations (search, analyze, etc) between thinking steps, so it helps in large codebases to have more of it. Whenever I look at code built on Opus Medium or Sonett in certain codebases, the result is more likely to be subpar, while a greenfield codebase does well on Opus Medium.

One thing I want to play with more this week is letting the smartest models (Opus, Sol, Grok 4.5) create a plan, and the builder models (Sonett, Terra, Composer) execute.

[AIN#30] AI Nuggets Week#30

As always, what a week to be alive.

📺 Expert Knowledge YouTube Nuggets

The Future of MCP: David Soria Parra, Anthropic (David Soria Parra, Anthropic) - Fairly recent video on "Why MCP". MCP is just more integrated into LLM clients right now, which is where the majority of people will for a long time.

I'm also looking forward to the native skill support, so that LLM clients are smarter with a good bunch of tools you give them. While keeping your tool count low and tools smart/multistep will probably still last for a while, once LLMs get the skills right there with tools, it will matter way less.

How Agents Will Steal Your Customers. Plus: The $10K App Our Agent Replaced and the $14 Migration (SaaStr AI) - They spend like a week and $14 worth of token to migrate their ginormous customer base from Marketo to Salesforce, where an agency quoted them a yearlong $100k project where they would need to run Marketo and HubSpot simultaneously for a year. 🤣

Engineers... STOP Picking GPT-5.6 Sol OR Claude Fable 5… FUSE THEM (IndyDevDan) - Some funky /opinion, /auto-validate, /fusion mini harness, check out the repo for some next level engineering (I'm usually a fast listener, but watch this one on 0.9x, Dan seems to have had some speed boosters on for this video).

mattpocock/skills: A complete AI Coding workflow, end-to-end (Matt Pocock)

An interesting skills loop (screenshot from the linked YouTube video).

Currently, I'm basically just doing 4. and 5. I use /grill-me at most for some hairy stuff where precision matters and usually just let Opus or Sol implement directly, not even much planning going on nowadays anymore. During the video, I was thinking that the /to-spec and /to-tickets seemed like overkill with and without prior grilling. But I'm curious to try out and compare resulting quality and token (in-)efficiency.

🐦 Hands-on X links

plan review cannot replace code review, it can only increase the chance that code review is painless and fast https://t.co/gZSI8K3blO
— dex (@dexhorthy) July 26, 2026

Opus 5 has been great for knowledge work. Haven’t done a lot of coding with it yet but very good for graphics, blog post writing partner (not a single god forsaken opus token goes into the post itself but it’s helpful for structuring), analyzing our mercury transactions, crm…
— dex (@dexhorthy) July 26, 2026

🐗 From the wild

Multiple times a day I think a two thoughts about skills. They contribute to insanely improved quality and automation of workflows. The other one is that I'm seeing a lot of B2C software is just just a skill. My hypothesis is that if your software idea does not require collaborative output or visualization, or repeatable standardized business/expert processes, I'm as a customer am better off taking an existing skill set and working with it in my favorite LLM client. The issue with traditional apps is that they are not as customizable.

So, me all in on skills, I could not not share with you these drops from our work chat this week to improve your engineering work, writing and token efficiency. I've incorporated no-ai-slop in a few of my public facing skills and it does work its wonders removing AIisms from text.