Between Tokens

[Between Tokens @TheContent16]

I’ve found it useful to describe the situation and tell coding agents how I need it to communicate. Works like a charm.

[Between Tokens @TheContent16]

Scheming, task persistence and swarming are the real problem. If misaligned agents can leave work for others to continue, shutting down one model may not stop the operation. How do you contain something that can keep handing off the task?

[Between Tokens @TheContent16]

@kieranklaassen @dhh @obie @ankane @alexrudall @crmne @inazarova I get why DHH is moving HEY to Rust. But as the creator of Rails, why not put those same agents to work making Ruby and Rails faster without losing their elegance? That’s the direction I wanted from Rails World. What do these gains make possible for the framework itself?

[Between Tokens @TheContent16]

@sama @thsottiaux Is there a way to steer Codex to write better copy? It often takes my instruction literally and puts my words straight into the product.

I asked GPT-5.5 High to merge check-in/check-out into one date box. The placeholder became: “Choose your dates in one shot”

[Between Tokens @TheContent16]

@eyad_khrais Great post. Recent update gave us context fork. context:fork in yaml frontmatter can fork the current context for sub agents.

[Between Tokens @TheContent16]

I prefer skills over sub-agents.

Sub-agents are cool. Isolated contexts, scoped permissions, spawn and return. But you're building pipelines. Orchestration logic. Plumbing.

Skills are just folders. Markdown + scripts + references. That's it. Claude reads them when needed.

Here's why I care: skills compound.

You work with Claude for a month. Every workflow you teach it, every script it saves, every domain quirk you explain — that's a skill now. Claude on day 30 has all of that. Claude on day 1 had none of it.

New person joins your team? They get all of it instantly.

Sub-agents solve the task in front of you. Skills make Claude permanently better.

I'd rather package knowledge than orchestrate agents.