video - no vibes allowed - Solving hard problems with agents - 2025-12
https://youtu.be/rmvDxxNubIg?si=50LQcPbs8tvxLGl-
Solving hard problems with coding agents
At 2:40 maintaining mental alignment
...
I think this is what I'm doing with app specification
At 3:40 he talks about intentional compaction
...
essentially compacting the session memory into a summary or context compression
At 4:30 he makes the point that LLMs are like pure functions
At 6:00 he talks about context window zone -- the dumb zone which is about 40% full. And he makes the point that if you have a bunch of model context protocol - MCPs, you pretty much always end up in the dumb zone.
At 7:30 he says that even better than subagents for managing context is frequent intentional compaction.
...
This makes me think about how Claude Code clears the context window after planning
At 10:25 spec-driven development is broken. Not the idea, but the phrase.
The problem is semantic diffusion. People mean different things by it.
...
Related to app specification
At 13:00 he talks about internal documentation that is explicitly geared towards bootstrapping the context needed for an agent.
...
like in film - memento
At 13:30 the problem is that the documentation skews from code, so you have to spend time updating it.
At 14:20 maybe a slash command building up a research document for the coding agent
...
Could be an agent skill specifically geared towards research in a part of the codebase. Building the research document feels like a plan document for a coding agent
At 15:00 you need mental alignment when working with coding agents and how PRs are for mental alignment
...
mental alignment
Mentions how tweet - Mitchell Hashimoto put his Amp threads in his pull requests
At 16:15 your plan document for a coding agent should express your intent, but also lead to reliable execution, which is why they are starting to have the actual code snippets in them.
At 17:00 he talks about human leverage
...
feels related to bottleneck - human reviewers
At 18:00 the research, plan, implement -- RPI. He also talks about harness engineering