AI Discussion

[matthewkeil]

There has been SOOO much activity in the AI space the last few weeks that is important to have another AI sync to surface important activity.

What sparked this meeting was surrounding token efficiency with regards to logging. I dont want to spoil the details but its fundamentally around creating AI-first development policies. Using this as a frame of reference we can create coherent workflows that are both human and more importantly AI friendly. This will be the heart of the discussion.

We also think its important to share some wins and losses that each of us have had.

If there is anything you want to add to the meeting topics feel free to add a note below and you will be included in the meeting flow and some time will be reserved for you.

Meeting Agenda

AI-First Development Policies and Procedures

Things are moving super fast with AI these days and it has wholesale changed the job description for all developers (along with many other careers). We have been adapting quickly to the changes but its still coming at us like a freight train. Having a coherent plan for how all the specifics come together such that both we humans and our bot companions are working both effectively and efficiently is critical for success. While the short list of ideas below is a taste there is lots of work to do so that we can learn from each other and refine the process as the tech itself matures (on a daily basis).

Log Context Reduction

There is a lot of excess context that we send to the frontier models when doing node debugging. It is possible to handle this in several different way, but lets invite discussion on which might lead to better outcomes and higher token efficiency. Some possible ideas are:

• throw more tokens at the problem and just let money fix it
• figure out a "grep" skill that can be used to filter data that gets sent to the model
• create a "bot" log level that is between "info" and "verbose". let the bot instrument the code where it needs logs to debug what is happening. Send only error, warn, info and bot logs to the models

The above could be augmented by given the skill a resource.md file that has one full slot worth of information (or possibly two so it sees an epoch transition as well) that it can use to decide which lines are pertinent to grep/instrument for analysis and to gain context on full slot timelines.

Spec/Plan Driven Development

There are a few ways to slice the pie when building out features and often the complexity of the task necessitates different approaches to how an AI driven development workflow should play out. In particular spec/plan driven development has been a key area of success for some, and for others it feels like a waste of time. This also seems to echo what some of the youtube "prophets" tend to espouse. We would like to invited discussion on how some of us are approaching the process of prompting and context management these days (with 1M token context windows).

Round-Table Discussion

The Power of OpenClaw and Introducing Lodekeeper

Wow... just wow.... what a game changer

[joshdougall]

I've been iterating on AI-assisted development patterns across infrastructure-general and my personal IaC repos. A few things I'd like to share and get feedback on:

• CLAUDE.md layering (global/project/repo) for scaling context across repos without duplication
• Persistent memory + session continuity (STATE.md) as a token reduction strategy, similar to what Lodekeeper is doing with daily logs + MEMORY.md
• multi-agent skills: orchestration vs. single-agent use cases
• my love for devil's advocate agents
• MCP servers as an alternative to the log context reduction problem: give the model query tools instead of filtering logs into the prompt

Also keen to discuss:

• Code review signal-to-noise tuning (shared pain point, would love to collaborate on calibrating this)
• Dare I say OpenClaw for infra work?
• ZeroClaw vs NemoClaw vs OpenClaw

[nflaig]

re memory system: memory-system.md is pretty sophisticated

[nflaig]

also have a Devil's Advocate now

[nflaig]

Dev Workflow has worked pretty well for bigger tasks, it's pretty openclaw specific but I am sure can be easily adapted for claude/codex, however it does use opus/gpt and even gemini as models

[jflo]

Ralph loops - still necessary? Claude now has a 1M token context window, and built in compression supposedly got better.

[nflaig]

I haven't used the 1M context with openclaw for now because it's not really a game changer there, I think if you wanna stay in the loop of development it doesn't really matter but if it's a task that the agent should work on for hours then it's still needed, so far the longest time I had lodekeeper work on a task without interruption or input was almost 48h