Joe Muoio

yes! This makes sense to help with token spend and just general efficiency. I have been using a similar technique to ask questions in rounds (just naive prompt during plan). I am excited to try the new version from someone who has put a lot more thought and testing into this

@bradenleague whats your favorite TTS software?

@JoeMoCode I can’t read

I read most of the code I write, most of the time. Do you? Why or why not?

@plainionist As the great philosopher group, Blink-182, wrote: "work sucks, I know."

I thought loop engineering was about the agent team. But I feel like I am the one stuck in an endless loop: writing tickets, reviewing specs, analyzing what to improve next, constantly context-switching between different tickets. Exhausting. 🤷‍♂️

@mattpocockuk Yeah I am stuck on this too. The way I think of it... Check in and maintain only what needs to be evergreen. Then for feature branches, I check in semi-ephemeral artifacts which includes research and plans. But need to archive or remove these when finished. External KB good, too

I use GitHub issues for PRD's and Implementation plans, but currently I check research files into git because I don't know what else to do with them. Considering storing them in an external knowledge base instead

There are a class of AI Coding assets that IMO don't really belong checked into git: - PRD's - Research files - Decision maps - Implementation plans Folks who agree with me, what are you using instead?

@plainionist It has always been a context problem. And dare I say, always will (with LLMs). You can never rely on the latent info in an LLM and it must always be grounded. Just from the basics of how it works and realities of changing context and training pipelines.

Serious question: Do you still see recent models producing “AI slop”? Or is it mostly a context problem now? 🤔

Got an ASUS Ascent GX10. What are your top 3 local models to run and why? So far, playing with qwen 30B instruct coder model and pretty amazed with how fast it is and generally quite good with tool calling.

Maybe 4.8 is "better" at coding, but I will just start trying it for planning for now. I'd much rather use a cheap model for implementation than an expensive one. Big context window and better natural language processing is useful when researching and planning, not coding.

Opus 4.8 is out, but I am WAY more hyped about Composer 2.5 for coding. Opus class has been good for planning... but it's somehow worse at coding than composer. By the time I code with AI, the task is well thought through and I just need it done.

@plainionist yeah totally there is a place for rules. I see them primarily as a way to direct agent to the place it needs to find that info rather than capturing it in a likely duplicate way. I don't give agents roles like that so I don't have opinions on that, specifically.

@JoeMoCode some rules are important but not that easy to catch as using linter, e.g. that agent stays in its role 🤷‍♂️

Sometimes my AI agents seem to have the memory of a goldfish. I have to regularly remind them to follow the rules 🤷‍♂️

@sickdotdev 100% Cursor Vertically integrated harness has most potential, but also most risk when the company is built to sell compute VSCode (with devcontainers) + Cursor CLI for in the loop dev. Cursor Cloud agents for parallelized sessions. Cursor IDE is not my cup of tea

Choose only one: - Cursor - Claude Code - Codex

github.com/mattpocock/ski… Matt's skills are great. At least the few I have used, so far. Lots of good stuff in here I have not even discovered yet, I am sure.

@plainionist not everything is so simple, but there are some things that are :) Rules are always gonna be "okay-ish" and never going to SOLVE a problem. Only mitigate how often it happens.

@plainionist That's insulting to the goldfish who do actually have memory! Shift away from rules to determinism. Example, don't have rules about code style, instead have a hook that runs your linter and hook that up to run when a file is changed by your AI harness. No need for style rules

@DanielMiessler @MattTrifilo And this is a core assumption of the argument. That we have "intelligence". I like that you defined it here and agree models give us this to an extent. Employees also give accountability. How okay is it if the workflow is wrong? who has accountability? Bigger problem, IMO

@MattTrifilo Not true. Several of the components in these algorithms require something that automation doesn’t have: intelligence. The ability to make informed decisions given the context. We have that now, so it’s game on.

Claude Code is about to release a feature called /workflows that I think will be extremely significant. Especially for Enterprise AI. I talked about this in 2024 in a post called Companies Are Just Graphs of Algorithms. Basically the idea is that all work is just an algorithm, i.e., a series of steps to accomplish a goal. Skills and Cowork have been heading in this direction already, and we've seen what that's done to company valuations in various spaces. Well this is closer to the final form. It's turning the regular, expected work that's done in companies into pseudo-deterministic workflows that follow defined SOPs. The human role will be determining what problems to solve (taste, expeirence, etc), building new products from that, and then optimizing these workflows from above. But the work itself will be these workflows executed according to SOPs.

@thdxr I been thinking this as well... but, don't forget about your own skill growth in that time! Turns out things are hard. Software is hard, and we are here and like it because it is hard. Always something new to learn and apply.

think back to projects you've worked on in the past it's hard not to imagine they'd have been completed way faster now that we have ai but everything still feels as slow and as difficult as ever

Only if the practitioner wants better software and knows what that means and what it takes. Lots of examples to the contrary -- more bad software getting made than ever before. More great software being made as well. ...just more software getting made. :)

@plainionist huh, that's a very interesting tool... 🤔🤔🤔

@JoeMoCode DSL doesn't need to be complex - can be something like "extended gherkin" 😉

Serious question: How long until Markdown is replaced by a DSL optimized for instructing AI agents? 😉

@plainionist DSL needs to REALLY show value for it to be adopted and used to the point that we then fine tune models on it at scale. Without that, you are locking yourself in and/or the model will have minimal knowledge of it, or will have to bloat context to share knowledge of DSL.

@plainionist For runtime use cases with claude for instance, I use XML in order to better separate parts. Recommended in claude docs, too. I don't use it for general plans/notes/research and still opt for md. Kinda since its the default format and I haven't felt the need yet to change.