30 of 30Clear filters ×
A game about AI maturity, built partly for fun and partly out of frustration. Agents of Mild Panic is a skill tree you play, because maturity comes from more than just doing a skill.
I tested whether a model rebuilds a design better from the brand rules or from a senior designer's corrections. The corrections won: 26% fewer fixes on new pages, and ahead in 8 of 9 blind reviews.
We keep saying the bottleneck has shifted. But faster was never the success. The value is in doing what you couldn't do before, and a lot of the people doing it never had a pipeline.
Auto mode is now the default in Claude Code. It's a classifier, not a wall, and it's easy to leave on when your cloud, endpoint protection, MDM, connector permissions and the tool layer sit around it.
I asked Claude Code to find one setting in a 50, 500 and 5,000 service repo, with and without grep and a CLAUDE.md. Grep never cost more than 4 cents. Without grep, the CLAUDE.md didn't stop the scan.
For the first time, our work itself costs money. Stop watching the meter: give a problem $150 in tokens, aim it at what makes the work slow, and scale up when it pays.
I built a coding job full of traps to see if a skill could lift a cheaper model to Opus. Every model got every check right, and that's the finding.
I gave Claude Code the same coding job twenty times, half with a skill and half without. Telemetry showed what the skill changed, but not whether it was right.
I priced four kinds of design job in Claude, token by token. The build is the cheap part. Checking it is what closes the outcome.
Claude Design is a fantastic product, and without serious effort you'll still end up looking like everyone else. That isn't the AI looking like AI. It's the AI predicting from what you gave it, so creativity, skills and other mediums are what make the difference.
Most AI adoption metrics count activity: tokens, pull requests opened, lines suggested. There are four engineering metrics you already have the data for (testing, harness changes, understanding the business and review time), and outcome metrics on top of them that take real work to get.
Anthropic publishes a whole shelf of design skills and plugins. I took one, frontend-design, built the same climbing club site three times (no skill, the stock skill, and my own version of it) and screenshotted what changed.
Managed AI platforms are getting good at the how: agents, connectors, identity, guardrails. What they can't tell you is what any of it was for. A hypothesis: the missing link is the definition of an outcome, and a Work Pack is what that looks like written down.
Your teams are already using harnesses they understand, and mostly that's fine. When they're stuck and nobody can agree how the work should flow, two approaches keep coming up: start from a kit with GitHub's Spec Kit, or start from a path with Anthropic's AI-native SDLC playbook and build it up yourself in Claude Code.
A hypothesis. Platform engineering's old trick was making the hard parts invisible to app teams. Do the same for the whole business, so someone in talent can build an idea without knowing what identity is, and you've got a way to transform it.
Clear the skills hurdle and you'll start seeing a return on AI. Stop there and you'll see very little of it. The return is in what skills add up to across teams, domains and the business.
How I built a puzzle game in five weeks: an artifact prototype, Claude Code for the build, Gemini for the art and a second opinion, and a small team of agents that each earned their place.
I've now taken five legacy estates onto Kubernetes. I'm not sure I'd do it again. Not because it's the wrong technology, but because the landscape has shifted.
Redesigning your AWS account structure? Get your ingress and egress right on day one and forget everything else. I learned this the hard way twice.
A hundred agents building a product that one agent could have built. Are we overthinking ourselves into a burst?
The software factory is just the latest incarnation of spec-driven development. The spec keeps getting smarter. The people who own the problem keep getting left out.
Engineering hasn't gone away. It's just different: more complex than it's ever been, and simpler too.
To keep up with our agentic workflows. How long until pull requests get a meter?
Everyone talks about AI making us produce more. Nobody talks about what we consume, or where it came from.
People keep asking what I use. The tools that build trust so companies can move faster with AI.
Before you implement an AI-SDLC, understand how your people actually work. The technology gets easier with the right lens.
Everything is a signal, at least to some extent. So what if your SDLC telemetry told you who to hire next? Emphasis on the maybe.
How I actually use AI day to day: which models, which tools, and the three phases I work in. Because I wish someone had shown me this a year ago.
Give someone AI and they solve the problem that bugs them. That's how you end up with 18 versions of the same agent.
Committees, prompt engineering, agents of agents, another committee. Then you realise you're solving the same problem you've had for 10 years.
Nothing in here yet. It's thinking.