assertchris

The demo for Rapid Escalation is OUT NOW. Fight through short, high-intensity space battles, earn massive upgrades, and grow stronger every run. Win Fast. Grow Faster. Play the demo on Steam: store.steampowered.com/app/4411170/Ra…

New harness building going great so far.

"I haven't been yet. I don't really have any updates." is definitely a choice to make during standup.

On the up-side; the little book I'm writing about building Friday has a definitive end. I guess I will start another little book about building Helen.

@brayniverse Anthropic killing off Agent SDK usage under subscription billing.

@assertchris Oh dear! What happened?

Looks like I get to rebuild my agent harness from scratch.

My agent's take on "We value your feedback"

@AmanVirk1 Anthropic models, Sonnet mostly.

@AmanVirk1 I spend about $3 worth of tokens for the average 5 phase plan; with deep sub-agent research per phase.

How many tokens are too many tokens for planning implementation for a feature?

@zackkitzmiller I appreciate you taking the time to talk about it with me.

@assertchris Do what you will, obviously, I'm not selling anything. It's just drastically improved cost and workflow speed over here.

Seriously though, don't sleep on whetstone. Yes it's just a few tools packaged together, but it works well with zero overheard, and a one line install github.com/z19r/whetstone

@zackkitzmiller After my tweet I dug into headroom and saw the telemetry is happening in that actually. I think it's interesting, but looks like they're working through some prompt cache issues.

@assertchris Re: phone home. A future whetstone update could do that regardless.

@aschmelyun Fun fact: I am currently running at least 111 (not a typo) MCP tools in every session.

@aschmelyun My best advice for when to make MCP tools is that you should figure out the things you want to be deterministic, with validated inputs and outputs. Added benefit is that you can blanket allow MCP tools you know are safe to always run.

Feel like I'm missing out because I don't use skills, or a lot of MCP, or multi-agent orchestrations when using AI dev tools. I'm just like "implement this feature" or "how do this work" or "no not like that, do this instead". Idk, I feel fast and accurate so why change?

@aschmelyun My skills are more like large repeated guidelines/prompts; where I expect interactivity and rarely for the agent to abort if conditions are sub-optimal.

@michaeldyrynda @aschmelyun They're a little coded to my harness (custom thing running on top of Claude Code SDK) but the basics could be the same whether I was using Claude Code CLI or something else.

@michaeldyrynda @aschmelyun Here are my workflow skills. Both the planning and implementation follow this orchestrator/sub-agent approach. I use Sonnet exclusively (at least until I hit the weekly limit, and then Opus 4.6 low effort. God results all-round. gist.github.com/fridaytherobot…