Topic
18 posts
An eight-minute vibe-coding challenge at a Mouse AI League mentor mixer, where I volunteer as a mentor and judge: an AI agent built a French flashcard app in eight minutes, and the speed of the machine is what makes product judgment the…
About ten on one subscription, and the model is never what stops you. Four shared resources under the agents set the ceiling: the link to the vendor and its quota window, the git checkout, the port your tests bind, and the file every bra…
Ballast declines 24 of its 89 golden questions on purpose and answers the other 65 with citations. As of the August 16 ledger the answer-or-decline call was right on all 89, and one of the 65 answers was unfaithful to its sources. Where…
Thirty days of running an agent fleet from chat: 8.7 billion tokens, 70,227 calls, 2,861 sessions, a 98 percent cache hit rate. At each vendor's list price that traffic bills $7,566. Three subscriptions covered it for $599. Where the tok…
I split a day of development across two AI subscriptions: judgment on the Anthropic meter, typing on Moonshot's. Five merged PRs later, here is where the tokens went, how K3's pricing actually compares, and how I'm tuning the mix.
I built a self-healing RAG pipeline, a guardrails gateway, and an eval gate as one system, then threw a 44-question battery at it. Zero hallucinations, because the behavior that matters most is refusal. Here is how trust got built into t…
A year of building and operating a small fleet of finance and content products almost entirely through an AI coding agent. What worked, what was hard, the honest failures (including a flagship signal that measured nothing and an edge tha…
Prompt caching looks like a flag you flip for a cheaper bill. It is really the reuse of a stored prompt prefix, governed by three rules, and applying it across four parts of my own system showed where it pays, where it quietly does nothi…
Four days after I said goodbye to Opus, an export-control directive pulled Fable 5 offline and the fallback became the workhorse again. What I shipped in the window, what it cost, and the model-tiering plan for when Fable comes back.
I handed a backlog to Claude Fable, told it once it could merge, and let it run. It shipped seventeen items across five repos. The line that mattered was not in the work it finished. It was in the work it refused to touch.
Anthropic shipped Claude Fable 5 and Mythos 5: same model, two names, one safeguard layer apart. What the new frontier model means for running agents in production.
Dumping the whole corpus into an AI agent makes it worse, not better. The fix is architectural: each task loads a curated slice, not everything you have. Here is the method, and the same move at three different layers: specs, sensor data…
A small ARM box that started as a local LLM experiment and ended up a self-governing node: private retrieval, a resident agent under a written constitution, a code-enforced safety fence, and a nightly job where it audits itself and files…
Anthropic published a guide on building a session-level orchestration mode. I built it two ways, on the CLI and on the API, and then hit the part the guide does not cover: an orchestrator that fans out is useless without a backlog of rea…
How I replaced manual CSV exports with a live Garmin data feed for my AI marathon coach: a scheduled unofficial-API poller, resilient session handling, and the design calls that keep training and recovery data fresh and trustworthy.
Spec-driven development reads like a methodology for controlling AI agents. It isn't. It's a methodology for managing context across stateless sessions. The spec is the persistent memory.
Two production sites, a blog, and two personal AI projects, shipped this week from a phone. The chain is voice dictation into Perplexity Computer, a spec, then Claude Code on the web. The interaction model is the story.
How I built a personal AI coaching system for marathon training, layering deterministic guardrails over an LLM narrative engine, ingesting Garmin FIT files, and designing for my own injury history.