Mark Headd

Humans required: Why AI coding still requires human experts to work most effectively | Ad Hoc As agencies explore and adopt AI-assisted coding tools, there's a growing body of research on how to use them most effectively.

Agentic coding is changing the way software is developed, but it likely won't remove humans from the process. Human expertise is as critical as ever.

adhoc.team/2026/03/09/h...

#AI #ChatGPT #Claude #Gemini #Cursor

Modernizing Government Regulations with AI How AI tools can help identify administrative burden in government regulations, and what lessons transfer from legacy system modernization.

New on the SpecOps Blog - Modernizing Government Regulations with AI. spec-ops.ai/blog/posts/m...

Agent Slimmer Optimize your AGENTS.md to its minimum effective form. Research-backed skill for trimming AI coding agent context files.

Trim down those bloated AGENTS.md files.

mheadd.github.io/agent-slimmer/

#ai #cursor #copilot #agents #claude

If this doesn't get your body moving in some kind of way, get your pulse checked. 🏵️ ❤️ 🎤 🎧

Left side says We Love Open Source. #WeLoveOpenSource. ATO. A community education resource from All Things Open. Right side has Illuminated robotic balls, blacking color with a white stripe and blue lights.

🚀 NEW on We ❤️ Open Source 🚀

AI is making code cheaper to write & easier to understand, shifting the real scarcity to domain knowledge. Mark Headd explores why future collaboration may center on sharing specifications and institutional expertise.

allthingsopen.org/articles/ai-...

#WeLoveOpenSource

This is important.

"CLAUDE.md is not a one-time setup. It’s a living document — but not an ever-growing one."

Amplifying — AI Benchmark Research Systematic analysis of how AI systems make decisions — from product recommendations to developer tool choices.

For government projects with mandated platforms and compliance requirements, those defaults are wrong. That's what AGENTS files are for. 3/2

amplifying.ai/research/cla...

#AI #ChatGPT #Claude #ClaudeCode #LLMs

New data from amplifying.ai shows Claude Code defaults to Railway for deployment, hand-rolls auth from scratch, and never picks AWS/GCP as a primary choice. 2/2

New research has lots of people saying AGENTS.md and CLAUDE.md files aren't worth it. But the research shows that AI coding agents follow these instructions reliably — the problem is most of these files are full of things agents would figure out anyway. 1/2

What the Research Is Starting to Tell Us About Agent Skills A new benchmark study provides empirical support for the value of human-curated instruction sets in AI agent performance, and what it means for SpecOps.

New research: AI agents given human-curated Skills improved performance 16 points on average. AI Agents that generated their own skills files? Essentially zero benefit. Human expertise still matters most.

spec-ops.ai/blog/posts/s...

#AI #Claude #ChatGPT #LLMs #Codex #Copilot

Me, coming back to work after vacation:

"I'm coming back to this project after being away for about a week and there have been several non-trivial changes committed to the code. Can we do a fresh git pull and discuss the changes that have been made?"

SpecOps: Specification-Driven Legacy Modernization Specifications over code. Preserve institutional knowledge, verify behavior, and generate modern systems with AI.

You can learn more about the SpecOps Method and obtain a copy of the book I wrote about it here: spec-ops.ai

I wrote a book about exactly this challenge. The SpecOps method proposes creating that verification artifact for legacy modernization efforts by leveraging AI to generate specifications that domain experts can read and confirm. The trends called out in the Anthropic report verify this approach.

This is why the role of humans doesn't disappear or diminish as AI coding tools gets more capable. AI can handle legacy code base analysis and new code generation. Humans are best suited to handle judgment and verification. Each actor in the process is focused on what it does best.

Domain experts know whether business rules are correct. But they often can't read COBOL, Java or other legacy languages. The verification problem in legacy modernization isn't about technical capability. It's about creating an artifact that human judgment can engage with.

The report makes a prediction of significant relevance to governments: AI will handle COBOL fluently soon. But reading old code isn't the hardest part of legacy modernization. Verification is. AI can describe what code does. It can't verify that system behavior matches policy intent. Only humans can

Anthropic's new report on #AI coding trends shows that engineers use AI in 60% of their work but fully delegate less than 20% of tasks. They call this "the collaboration paradox." The report also predicts AI will handle COBOL fluently soon. Let's dig in. 🧵 resources.anthropic.com/hubfs/2026%2...

Why your coding agent keeps undoing your architecture Think of your codebase as a state machine. The code is the current state. The architectural decisions are the transitions.

"Your coding agent is not malicious. It's amnesiac."

mahdiyusuf.com/why-your-cod...

True story

Thanks, Randy!

Government Software at the Crossroads Government tech leaders are signing longterm software contracts during the biggest shift in how software works since cloud computing. The assumptions behind those decisions may already be obsolete.

Lots of chatter online about what people are seeing with the rapid evolution of #AI coding tools, but I don't see as much conversation about what these changes mean. Here's what I think it means for a sector that I have worked in on and off for my entire life: government. civic.io/2026/02/11/g...

🤖 The SaaSpocalypse - The week AI killed software How Opus 4.6, Plugins Codex 5.3 wiped $285bn from SaaS valuations

I been saying this: "The SaaSpocalypse - The week AI killed software" www.fintechbrainfood.com/p/the-saaspo...

How government platforms make AI coding assistants more effective | Ad Hoc AI coding assistants are transforming how developers work, generating everything from boilerplate code to complex business logic in seconds. But these tools are only as effective as the context they h...

How government platforms make #AI coding assistants more effective.

adhoc.team/2026/02/05/h...

Cool new feature to add agentic workflows to your GitHub repos. Another banger coming from the group at @githubnext.com and further validation that the future of software development is opening up for people that have developed the ability to communicate clearly. Words > Code github.github.io/gh-aw/

Watch Mob War: Philadelphia vs. The Mafia | Netflix Official Site A brutal mob war between John Stanfa and Joey Merlino plunges '90s Philadelphia into chaos and bloodshed in this captivating true-crime series.

There's a good series on Netflix about the Philly Mob Wars, which features Skinny Joey in his youth. www.netflix.com/title/81763602

Astute watchers of the ESPN 30 for 30 "The Philly Special" will notice a special cameo by a quintessentially Philadelphia character: Joseph Salvatore Merlino. I think he has a steak shop now in Philly or something. Wild seeing him in this. en.wikipedia.org/wiki/Joey_Me...

If there was a snow shoveling event in the Winter Olympics, I'd crush that shit.

30 For 30: The Philly Special [FULL EXCERPT] | NFL on ESPN YouTube video by ESPN

ESPN 30 for 30 episode on the iconic Philly Special play from Super Bowl 52 drops tonight at 9:30 PM EST.

INJECT THIS DIRECTLY INTO MY VEINS!!! #GoBirds #30for30

www.youtube.com/watch?v=nVYa...

Good advice here: "Fully trusting the answer an LLM gives you is foolishness, but it’s wise to use an LLM to help navigate the way to the answer."

Sometimes you just have to "git checkout main && git branch -D your-old-idea" and start again...