Pete Hodgson

Pete Hodgson

@thepete.net

Outside expert helping engineering teams adopt AI

212 followers 146 following142 posts

1000% It's actually very sad how much "scratching my own itch" creativity energy is being sunk into products that only 1 person will ever use, because scratching our own itches has become so cheap. Quite a counter-intuitive result.

Matt Wynne@mattwynne.net

I can't shake this thought that we're "coding past each other" - everyone building stuff with their sirens, but not spending enough time integrating other people's ideas.

This new benchmark is an apt metaphor for the AIs ability to gaslight humans with legit-looking-but-wrong output. Oh, you can't find the raccoon? Huh. Sorry you're having trouble with that, meatsack. I tried to not make it TOO hard for you, lil bud, but I'll make it even easier next time.

Simon Willison@simonwillison.net

I came up with a somewhat foolish new benchmark for testing image generation models, to exercise the new ChatGPT Images 2.0: "Do a where's Waldo style image but it's where is the raccoon holding a ham radio" simonwillison.net/2026/Apr/21/...

simonwillison.netWhere’s the raccoon with the ham radio? (ChatGPT Images 2.0)OpenAI released ChatGPT Images 2.0 today, their latest image generation model. On the livestream Sam Altman said that the leap from gpt-image-1 to gpt-image-2 was equivalent to jumping from GPT-3 …

Pete Hodgson reposted

@gergely.pragmaticengineer.com

Tomorrow on The Pragmatic Engineer Podcast: Designing Data Intensive Applications with @martin.kleppmann.com

Pete Hodgson reposted

STOP SCROLLING. Mandatory watch. It's < 3m. Share to everyone. 🫡

Pete Hodgson reposted

Camille Fournier

@skamille.themanagerswrath.com

💯💯💯 bcantrill.dtrace.org/2026/04/12/t...

bcantrill.dtrace.orgThe peril of laziness lost | The Observation Deck

Bryan Cantrill@bcantrill.bsky.social

The peril of laziness lost bcantrill.dtrace.org/2026/04/12/t...

bcantrill.dtrace.orgThe peril of laziness lost | The Observation Deck

NORTH VIRGINIA OPOSSUM CRUSING THE COMMONWEALTH SINCE DUSK 😂 Sometimes the weird goof energy of an LLM hits just right

Simon Willison@simonwillison.net

From the comments: /* Earring sparkle */, , ,  ... static.simonwillison.net/static/2026/...

Reminder for whoever needs to hear it, today's cutting edge AI models are still very jagged in their "PhD-level intelligence". Put more bluntly, Opus 4.6 Extended sometimes can't count to 2. 🫠

>As we’re entering the phase of single-use plastic software, we might be moving the whole layer of responsibility elsewhere. "single-use plastic software" is a really a striking way of framing it. It feels very accurate, and I'm not sure how I feel about it.

Armin Ronacher@mitsuhiko.at

Today Thorsten Ball said he's the bottleneck. I too am the bottleneck and it made me think. lucumr.pocoo.org/2026/2/13/th...

lucumr.pocoo.orgThe Final BottleneckAI speeds up writing code, but accountability and review capacity still impose hard limits.

It's hard to find balanced opinions on AI coding. Lots of the people pushing the envelope on AI-Assisted Engineering also have somewhat bonkers AI-maximalist perspectives. Here's some people who're in the thick of it, advancing the field but who I also trust to bring a thoughtful, balanced take 🧵

View full thread

@birgitta410.bsky.social ThoughtWorks Distinguished Engineer, focused on AI-assisted Engineering Dexter Horthy: x.com/dexhorthy Coined the term Context Engineering, was into Ralph before it was cool 🧵...

x.comdex (@dexhorthy) / Xdex (@dexhorthy) / X

Who'd I miss? I also posted this list on LinkedIn to, and got some nice additions from folks: www.linkedin.com/posts/beinga...

It's hard to find balanced opinions on AI coding. Lots of the people pushing the envelope on AI-Assisted Engineering also have somewhat bonkers AI-maximalist perspectives. Here's some people who're in the thick of it, advancing the field but who I also trust to bring a thoughtful, balanced take 🧵

@simonwillison.net the OG LLM user, very plugged in to what's happening, prolific writer, never hyperbolic Mitchell Hashimoto: mitchellh.com/writing/my-a... Built Hashicorp (Vagrant, Terraform, etc), now building ghostty, and using AI a lot to do it. 🧵...

simonwillison.netSimon Willison’s Weblog

@mitsuhiko.at lucumr.pocoo.org/tags/ai/ Grizzled veteran engineer, using AI a lot but very pragmatic about it @addyosmani.bsky.social addyosmani.com/blog/ai-codi... He writes REALLY LONG posts, but they're full of great info 🧵...

lucumr.pocoo.orgEntries tagged

Hot-ish take: devs should be spending time now learning out how to be *efficient* with the tokens in their agent's context windows 1) it makes the agent perform better 2) tokens are gonna start costing more over the next couple of years These Ralph shenanigans might look real silly in retrospect

“Have you tried out Ralph Wiggum loop?!” “Are you still using beads?” “Is Gas Town actually legit?” “Know anyone who’s into BMAD?” These days I’m glad my family doesn’t know what my Serious Work Conversations sound like. It’s embarrassing. Let’s talk about how to deal with all this noise. 🧵...

You’re staring at a breathless social media post (probably with a really cringe AI-generated image of a robot “writing software”) or Slack share talking about the latest AI-coding flavor of the week. You’re wondering “is this actually worth me checking out?” Here’s the heuristic I use... 🧵...

focus on whether it helps address the LLM’s fundamental limitations by asking: does this new approach systematically improve how the agent manages context? does this new approach systematically improve how the agent gets human feedback? Read on and I’ll elaborate: blog.thepete.net/blog/2026/01...

blog.thepete.netAssessing the Latest AI Coding HotnessCut through the breathless hype and figure out whether the new framework with a weird name actually addresses coding agent's fundamental limitations - bad context, and a mediocre sense of software des...

Claude Code is an amazing product. It does amazing things. I love it. Amazing. It's also most typically used as a CLI tool. This does not mean that Claude Code is an amazing product BECAUSE it's a CLI. Tell me why I'm wrong!

Devs should be regularly sharing coding agent sessions with colleagues. I made a thing to make this super duper easy: custardseed.com 🧵...

custardseed.comCustardSeed - Publish Claude Code SessionsInstantly share your Claude Code sessions online. Turn your sessions into shareable web pages with a simple slash command. Store sessions as GitHub Gists and view them beautifully rendered.

Peer review of AI coding sessions brings a lot of the same value as good code review: - share and learn good prompting practices - providing feedback on peer's techniques - understand how the code was built My hope is that CustardSeed will help teams realize that value.

CustardSeed is brand new! I’d absolutely love any and all feedback on: - the concept - how you'd like to use it - what's missing - what would stop you using it today? - the documentation - the installation experience - the UX of the session viewer - the UX overall

Really nice measured, balanced and thoughtful writeup of a decidedly unbalanced and unhinged project 🙃

Maggie Appleton@maggieappleton.com

I have Gas Town derangement syndrome and spent the last few weeks writing thousands of words on agent orchestration patterns; how they shift our bottlenecks and force us to ask whether and when we should stop looking at code maggieappleton.com/gastown

maggieappleton.comGas Town’s Agent Patterns, Design Bottlenecks, and Vibecoding at ScaleOn agent orchestration patterns, why design and critical thinking are the new bottlenecks, and whether we should let go of looking at code

who hurt you?

Gonna start saying "Clod" when referring to Claude Code. I've felt a need to use its full name to disambiguate between the Claude Code - the coding agent - vs Claude - the general purpose chat bot, but it's so tedious to type it out in full. It's either CC or Clod, and Clod is more fun.

"Why is Claude Code better/worse than [other coding agent] - they’re both using the same models under the hood?" A reasonable question, so I spent some time reverse-engineering Claude Code's interactions with the model to illustrate why coding agent matters blog.thepete.net/blog/2025/12... 🧵...

blog.thepete.netSame Model, Different Results: Why Coding Agents Aren't InterchangeableReverse-engineering Claude Code reveals why it performs differently from other agents that use the same Anthropic models. The answer lies in sophisticated context engineering and tool orchestration hi...

View full thread

each coding agent can bring: - different system prompt - different tools - (hidden!) context management tricks to steer the model’s behavior - task management tools to nudge the LLM into better planning - sub-agents, which can have a *huge* impact on context management 🧵...

sub-agents in particular are a big deal, because each agent uses it's own temporary context window, keeping the main context window clean. Claude Code also uses this trick for it's built in tools (e.g. WebFetch) lots more detail in the full post: blog.thepete.net/blog/2025/12...

blog.thepete.netSame Model, Different Results: Why Coding Agents Aren't InterchangeableReverse-engineering Claude Code reveals why it performs differently from other agents that use the same Anthropic models. The answer lies in sophisticated context engineering and tool orchestration hi...

"Why is Claude Code better/worse than [other coding agent] - they’re both using the same models under the hood?" A reasonable question, so I spent some time reverse-engineering Claude Code's interactions with the model to illustrate why coding agent matters blog.thepete.net/blog/2025/12... 🧵...

blog.thepete.netSame Model, Different Results: Why Coding Agents Aren't InterchangeableReverse-engineering Claude Code reveals why it performs differently from other agents that use the same Anthropic models. The answer lies in sophisticated context engineering and tool orchestration hi...

turns out that coding agents are much more than a “wrapper” around an LLM. They do a surprising amount of work under the surface to (a) maximize the useful information available to the LLM, and (b) steer its behavior. 🧵...

I continue to believe that context management is the #1 skill to develop if you want to get the most out of coding agents. But why? Modern models have a context limit of 200,000 tokens! That seems like a lot! There's a couple of reasons why... 🧵

turns out that: - your agent can burn through it fast - your LLM's performance starts taking a hit way before that "limit" is reached further reading: - blog.thepete.net/blog/2025/10... - research.trychroma.com/context-rot - github.com/adobe-resear... - arxiv.org/abs/2404.06654

blog.thepete.netAI Coding: Managing ContextManaging your coding agent's context is super important - a bloated context window will erode the quality of your agent's work over time. Learn some new techniques for trimming irrelevant details from...

This is a great summary of what code review is REALLY for. It also demonstrates why AI-powered code review does not help AT ALL with the most important aspects of code review: aligning the team around software design (#1 and #3 in the hierarchy). blakesmith.me/2015/02/09/c...

blakesmith.me Code Review Essentials for Software Teams » Blake Smith

View full thread

The biggest risk with AI code review is that it looks so legit! Each review is so thorough and detailed. No surprise: The LLM does a *great* impression of the huge pile of human code review examples in it's training set!

I worry a lot that this will induce more `LGTM` reviews from human reviewers when they see the "comprehensive" feedback from an LLM. Particularly when coupled with the increased overall volume of PRs that AI-assisted coding is producing... 😬

This is a great summary of what code review is REALLY for. It also demonstrates why AI-powered code review does not help AT ALL with the most important aspects of code review: aligning the team around software design (#1 and #3 in the hierarchy). blakesmith.me/2015/02/09/c...

blakesmith.me Code Review Essentials for Software Teams » Blake Smith

Also, given that - the LLM won't have the full context of the codebase, and - an LLM only has mediocre software design skills (at best) you're not going to get any sort of full critique of your proposed solution from an AI-powered tool.

This doesn't mean that you shouldn't use AI code review. But you should treat it as an additional piece of feedback on a PR, rather than any sort of replacement for human review. Think of it as a super-powerful static analysis step in your CD pipeline.

AI WON'T turn you into a 10x engineer but it CAN accelerate your day-to-day work. I'm hosting a live workshop early next month to help experienced devs level up from copilot to coding agents. No breathless hype, no "learn this one magical prompt" BS, I promise! blog.thepete.net/ai-assisted-...

blog.thepete.netAI-assisted Engineering Workshop SeriesLevel up from copilot to coding agents in this practical online workshop teaching AI-assisted Engineering practices for experienced developers. No breathless claims about 10x productivity. No suggesti...

Pete Hodgson reposted

@jaderubick.bsky.social

One thing you can do with skip levels is ask questions about things you believe you've communicated. You'll be surprised how often people haven't really internalized it.

rubick.comhttps://www.rubick.com/skip-level-1-on-1s/

Slides from my #kubecon talk today Feature Flags Suck! - The Problems With Feature Flagging and How To Avoid Them speakerdeck.com/phodgson/fea...

speakerdeck.comFeature Flags Suck! - KubeCon Atlanta 2025<strong>Feature Flags Suck! - The Problems With Feature Flagging and How To Avoid Them</strong> OK, feature flags are actually pretty awesome, but if…

Blurghed about managing you AI coding agent's context. Proactively trimming and refining the context your agent is working with is probably THE best way to improve the quality of it's work. But there's more to it than firing off a `/compact` every now and then. blog.thepete.net/blog/2025/10...

blog.thepete.netAI Coding: Managing ContextManaging your coding agent's context is super important - a bloated context window will erode the quality of your agent's work over time. Learn some new techniques for trimming irrelevant details from...

Pete Hodgson reposted

Okay, for the folks who asked: here's the majority AI view, writing up the reasonable, thoughtful view on AI that the vast majority of people in tech hold, that gets overshadowed by the bluster and hype of the tycoons trying to shill their nonsense. anildash.com/2025/10/17/t... Please share!

anildash.comThe Majority AI View - Anil DashA blog about making culture. Since 1999.

1000000%

Chris Coyier@chriscoyier.net

I feel like how we feel about smoking on airplanes is how the next generation will feel about us just mainlining social media.

Super interesting proof-of-concept exploit against an AI agent via GH pull request. Very sneaky prompt injection which convinces an AI coding agent to subtly inject a remote backdoor into a codebase. A great example of @simonwillison.net's Lethal Trifecta! blog.trailofbits.com/2025/08/06/p...

blog.trailofbits.comPrompt injection engineering for attackers: Exploiting GitHub CopilotPrompt injection pervades discussions about security for LLMs and AI agents. But there is little public information on how to write powerful, discreet, and reliable prompt injection exploits. In this ...

See if you can spot the prompt injection: github.com/trailofbits/... Or if you can spot the backdoor being added: github.com/trailofbits/... More on the lethal trifecta: simonwillison.net/2025/Jun/16/...

Good example of why it's important to monitor your LLM-powered features in prod. Apparently some bug in Amazon's AI assistant where the product description isn't in it's prompt. Pre-production evals are super important, but wouldn't catch this. Monitoring some key metrics in prod would.

The point being: a) In many situations LLMs are WAY WORSE that "classic" deterministic tools. b) don't waste a ton of ultra-valuable LLM context on guidance that could be enforced a different way 🧵...

thoughtworks.comIn Praise of the ./go Script - Part IMy step-dad is a cabinet-maker by trade. We chat about his work from time to time. I've often been struck by the similarities between building furniture and building software. Take tooling for example...

c) invest in making your codebases agent-friendly by wrapping all your conventions and tribal knowledge into a small set of scripts point C is exactly the same guidance that I was giving 11 years ago(!!!): www.thoughtworks.com/insights/blo... I was ahead of the curve I guess :D

thoughtworks.comIn Praise of the ./go Script - Part IMy step-dad is a cabinet-maker by trade. We chat about his work from time to time. I've often been struck by the similarities between building furniture and building software. Take tooling for example...

Such a nice way to illustrate the point

Jessica Kerr@jessitron.bsky.social

Putting all your controllers in one directory and your views in another Is like putting the cortisone and hemorrhoid cream in the drawer with toothpaste because they're all tubes.

Pete Hodgson reposted

they should be taken to The Hague for this graph crime