Matt Galligan pfp

Matt Galligan

@mg

257 Following
25484 Followers


Matt Galligan pfp
Matt Galligan
@mg
"Context engineering" is a great step-up from "prompt engineering." Coincidentally, it's also great to describe the coping mechanism my brain persistently operates under as someone with ADHD.
0 reply
0 recast
6 reactions

Matt Galligan pfp
Matt Galligan
@mg
Anyone regularly using Descript for coding/agent related videos? I'm gearing up to drop a bunch of stuff over the next few weeks and would really prefer to be leaning in on AI to help produce videos for them. Any feedback is appreciated!
0 reply
0 recast
4 reactions

Matt Galligan pfp
Matt Galligan
@mg
That's what open source is all about! Feel free to review and tweak as you like.
1 reply
0 recast
1 reaction

Matt Galligan pfp
Matt Galligan
@mg
Sharing my main "engineer" prompt that I use with Claude Code, Cursor, Codex, etc. Still a work in progress, but it's been pretty great for improving consistency and code quality. 🤖 https://gist.github.com/galligan/d19ab747766e2a3dcf61a8627ccbb274 I just @ mention the file in CLAUDE.md, etc.
3 replies
3 recasts
27 reactions

Matt Galligan pfp
Matt Galligan
@mg
When it's an instance-by-instance rename, yes. It still didn't catch 100% from what I can tell but it got close!
0 reply
0 recast
0 reaction

Matt Galligan pfp
Matt Galligan
@mg
My coding agents are over here lathering each other with compliments in the comments and I'm just impressed AF on the 12k line refactor in sub 2hrs lol
2 replies
0 recast
4 reactions

Matt Galligan pfp
Matt Galligan
@mg
Ok, Claude and Code Rabbit really need to get a room 😆
0 reply
0 recast
1 reaction

Matt Galligan pfp
Matt Galligan
@mg
I tried to make my coding agents less sycophantic. Well…uh…that seems to have worked! 🤨
0 reply
0 recast
7 reactions

Matt Galligan pfp
Matt Galligan
@mg
Dude. That means a ton to hear. Same. So appreciate you sharing that screenshot. 🫠
1 reply
0 recast
1 reaction

Matt Galligan pfp
Matt Galligan
@mg
Gave Claude Code a meaty task to chew through. Ran 30 minutes without intervention or stopping, got to only 3% of context left. Asking to pull changes from main, branch, commit, push, and pr got it to 0% 😅 Wow
0 reply
1 recast
10 reactions

Matt Galligan pfp
Matt Galligan
@mg
Watching Claude Code work as a GitHub action is just so satisfying. Currently using it to: - Review PRs - Implement changes it suggests - Research project, writing results to the report - Fire up new PRs to address issues Seriously this thing is 🔥🚀🤯
0 reply
0 recast
4 reactions

Matt Galligan pfp
Matt Galligan
@mg
100% I'm also careful to make sure when pasting the results back in, that it should account for any conventions in the repo that would apply vs. just treating it verbatim.
1 reply
0 recast
1 reaction

Matt Galligan pfp
Matt Galligan
@mg
Here's a little tip for working with coding agents on research tasks within a repo: Instead of asking the agent to research something, ask it for a sytem prompt with all the context another LLM would need to do complete the research. Paste the prompt into o3, then paste results back. No more endless tool calls or waiting around, just keep coding along, and when o3 is done, the results have been better anyway.
1 reply
0 recast
9 reactions

Matt Galligan pfp
Matt Galligan
@mg
Yep! Care of my ChatGPT fever dreams…
0 reply
0 recast
1 reaction

Matt Galligan pfp
Matt Galligan
@mg
A guy can dream
3 replies
4 recasts
28 reactions

Matt Galligan pfp
Matt Galligan
@mg
That's the plan! Likely would need to be powered by AI, but a grepa MCP server / set of rules files could solve for that.
0 reply
0 recast
1 reaction

Matt Galligan pfp
Matt Galligan
@mg
Basically! But I'm putting together a CLI tool, and scripts that will streamline it a bunch. So you'll be able to do `grepa inventory` and quickly get all tags, sort by frequency of use, see the files that have them as well as the string that follows the tag. Should make finding stuff lightning fast.
1 reply
0 recast
1 reaction

Matt Galligan pfp
Matt Galligan
@mg
⚠️ This is very much a proof-of-concept and I’d love feedback before v0.1! Drop a reply, issue, or PR! 🙏
1 reply
0 recast
1 reaction

Matt Galligan pfp
Matt Galligan
@mg
🍇 Grepa = “grep-anchors.” - ultra-terse tags (":ga:tldr start here", ":ga:perf,todo") - one anchor pattern per repo - dead-simple to insert, grep on, and agents are fast to understand how to search/write them Conventions & docs » https://github.com/galligan/grepa
1 reply
0 recast
5 reactions

Matt Galligan pfp
Matt Galligan
@mg
Watching my AI agents go crazy just to locate a single line of code  felt like hunting for a **needle in a stack of needles**. 🔍 🔍 🔍🤬 So I whipped up a little experiment called grep-anchors or "grepa". Just drop ":ga:<text>" anchors in comments and your tools (or agents) *teleport* to the right spot. Like the TODO's and FIXME's before, but with some conventions for predictability (and composability). 🚢 Early days, but I’m already seeing way fewer fruitless tool calls!
8 replies
9 recasts
53 reactions