The Cost of Accuracy
You don't buy answers from a frontier model, you buy passes. Most real tasks take more than one, and the verification passes are where the budget actually goes.
dev ai claude cost agents verification evals
Filtering by
Friction Is the Feature
The four-practice discipline I built after eighteen months of heavy AI use. Without deliberate friction, AI compounds capability and dependence.
ai claude-code discipline skill-atrophy overreliance
The Harness Is the Bottleneck: Five Prompts That Broke on Opus 4.7
Adaptive thinking and a literalness shift turned the Opus 4.7 prompt into a multi-dimensional contract. Half your prompt library signs old terms.
ai claude prompt-engineering evals
LLM Fundamentals Part 9: The Agentic Loop
The agentic loop is a while loop that checks stop_reason, executes tools, and returns results. It is 10 lines of code that depend on everything from posts 0-8.
ai llm-fundamentals
LLM Fundamentals Part 10: From Loop to Agent
Scaling an agentic loop to production means managing cost, context, and complexity. A roadmap: prompt caching, model routing, compaction, subagents, MCP.
ai llm-fundamentals
LLM Fundamentals Part 8: Tool Use
Tool use lets Claude call functions you define. You ship schemas, the model picks when to call, your code runs the operations. Prerequisite for agents.
ai llm-fundamentals
LLM Fundamentals Part 7: Structured Output
Structured outputs guarantee JSON schema conformance at the API level, which removes an entire category of parsing and validation code from your pipeline.
ai llm-fundamentals
Karaoke Audio by Voice, 200 Miles Out
Built a word-by-word audio sync feature for a blog post without touching my laptop. Voice on the phone, agents on the laptop, every change shipped to prod.
ai claude-code remote agents
LLM Fundamentals Part 6: Extended Thinking
Extended thinking gives Claude a reasoning scratchpad before responding. It costs output tokens but measurably improves accuracy on hard problems.
ai llm-fundamentals
When 'Working' Isn't Repeatable
Frontier LLMs are not deterministic, even at temperature=0. Same prompt, five different answers. Why it compounds when you stack LLMs, and how to recover.
dev ai claude evals stochasticity
LLM Fundamentals Part 5: Prompt Engineering vs Harness Engineering
As models get smarter, the value shifts from crafting prompts to designing the system around them. Where the prompt ends and the harness begins is the question.
ai llm-fundamentals
The exponential power of scaffolding
A 16-hour Claude Code session shipped a long-form post, three SVG diagrams, a CI script, and four legacy fixes from a seven-word instruction.
ai productivity claude-code case-study
RAG as a Tool, Not a Pipeline: An On-Call Investigator on Mastra and Bedrock
Long context did not kill RAG for incident response. Per-section indexing in pgvector, hybrid .docx extraction, an iterative Mastra agent loop on Bedrock.
ai rag mastra aws bedrock incident-response
Chat, Cowork, and Code, Oh My!
Three Claude surfaces, three context windows, three different ways of knowing about your files. Confirmed only from Anthropic's own documentation.
ai claude claude-code
LLM Fundamentals Part 4: Messages API
Every Claude API call is stateless. You send roles, messages, and the full conversation history each time. Here is how that structure works.
ai llm-fundamentals
A Three-Bar Statusline for Claude Code
A one-command Claude Code plugin renders three progress bars: context fill, 5-hour session burn, 7-day weekly quota. No ccusage required on CC v2.1.x+.
dev ai claude claude-code plugins shell
Your Claude Skill Is a Pile of Markdown. Ship It Like Software.
Six phases to ship a Claude skill like software: behavioral eval, packaging, review, CI, install verification, governance. Each catches a different bug.
dev ai claude workflow eval plugins
Part 3: Four Fix Patterns for AI Code Review (and the AI-Auditing-AI Problem)
Part 3 of 3. Four structural fix patterns that move an AI reviewer from ignorable to trustworthy, plus the meta problem of using AI to audit AI.
dev ai claude workflow leadership
Part 1: Designing Automated PR Reviews with Claude Skills
Part 1 of 3. Building a PR reviewer as structured Claude skills: anatomy of a review skill, why markdown beats prompt blobs, what makes iteration possible.
dev ai claude workflow leadership
Part 2: Diagnosing False Positives in AI Code Review
Part 2 of 3. The cry wolf failure mode, the trust threshold step function, and a four-step cluster analysis that surfaces root causes, not symptoms.
dev ai claude workflow leadership
Context Switching Is Metabolic Interval Training
Your brain burns 20% of your body's energy. Creatine refills brain ATP 40x faster than any other pathway. What that means for devs at their keyboards.
dev ai health workflow
LLM Fundamentals Part 3: Context Windows
Context windows are working memory, not storage. More tokens means more cost, more latency, and accuracy that degrades as context grows.
ai llm-fundamentals
LLM Fundamentals Part 0: What Are LLMs?
Large language models are prediction machines built on transformers. What they are, how developers use them, and where agent workflows fit in.
ai llm-fundamentals
LLM Fundamentals Part 2: How LLMs Generate Text
LLMs predict one token at a time from a probability distribution. Temperature, top-k, and top-p control that distribution, not creativity.
ai llm-fundamentals
LLM Fundamentals Part 1: Tokens Are Not Words
LLMs don't process words or characters. They process tokens, subword units that determine cost, speed, and context limits for everything you build.
ai llm-fundamentals
When Slower Is Actually Faster in Claude Code
A structured Claude Code workflow that replaced plan-and-pray with repeatable, first-pass-clean features.
dev ai claude workflow
Telegram, Claude Code, and a Proposals Portal by Midnight
Magic link auth, progress tracking, and one-command publishing. Built with Claude Code, Astro, Cloudflare KV, and Resend from my phone.
dev ai claude cloudflare astro
My Mastra Agent Found a Production Bug in Five Minutes
I stood up a Mastra workflow from my phone via Telegram to monitor Cloudflare Workers. It caught bots crashing my media proxy on the very first run.
dev ai cloudflare workflow
Pushing to Prod From the Mountains
AI agents and channels integration make it possible to ship production code from anywhere: a summit, a grocery line, a dog walk. Why it works.
dev ai claude workflow mobile
Self-Improving Claude Code Skills, Part 2: Full Automation
The self-improving Claude Code skills loop from Part 1 now runs itself. A SessionEnd hook and daily GitHub Action generate improvement PRs automatically.
dev ai claude workflow
How Smart Are Your Claude Code Skills?
A self-improving loop for Claude Code skills that observes failures, finds patterns, and proposes fixes. Skills that improve over time instead of degrading.
dev ai claude workflow
Is Your Content Discoverable by Humans Using AI?
How I implemented LLM SEO across 3 sites with FAQ schema, entity linking, and 14 AI crawlers, plus an honest look at what actually works and what doesn't.
dev ai seo astro
Claude Code Remote Control: Ship Code From Your Phone
How I use Claude Code remote control with the iOS app to ship code from anywhere. No SSH tunneling, no exposed ports, just a Max subscription.
dev ai claude workflow mobile
Turning Claude Code Agents Into an SEO Content Machine
How I used Claude Code agents, context files, and slash commands to automate SEO for my Astro portfolio site: keywords, linking, and content audits.
dev ai claude seo astro
Replacing Pixieset From My Phone on a Sunday Morning
Pixieset can't sell individual video clips. So I built my own storefront with Claude Code, Cloudflare R2, and Stripe, from my phone, on a Sunday morning.
dev astro claude cloudflare stripe video
Building a Filterable Resume Page with PDF Export and CI/CD for Astro
How I built a filterable tech resume with multiple export formats and set up a proper CI/CD pipeline with GitHub Actions and Dependabot.
dev astro ci-cd github-actions linkedin
Two Hours to a Christmas Gift Tracker (My Kids Can't Know)
How I went from 'I need to track Christmas gifts' to a fully deployed, PIN-protected wishlist app before my morning coffee got cold.
dev astro claude parenting
Dumber Models on Purpose
Why the most capable AI isn't always the right tool, and how friction in your workflow can lead to better architecture decisions.
dev ai claude workflow linkedin