29 POSTS
The claim that the harness matters more than the model is repeated everywhere. I went through what my sources actually show, what they only assert, and why a same-prompt head-to-head cannot carry the conclusion people draw from it.
When agents write the code, the scarce thing is knowing whether it is right. Practitioners handle that with deterministic gates, review depth scaled to blast radius, and a reviewer that did not write the code. Most of it is judgment, not measurement.
A coding harness can gate on tests. A harness for slides or design cannot, so it has to turn taste into reference sets, rubrics, constrained components and a human pick. Gamma's own account is the worked example, with the founder's claims marked as claims.
After reading a lot of one person's work, the useful output is a verdict on each idea: adopt, modify, reject, or conflicts with something I already do. A worked example from an audit of Lee Robinson's writing.
Agents make code cheap to write, so the scarce thing is how much complexity a codebase can carry. Lee Robinson's argument for spending abstractions like a budget, and where Ousterhout-style design fundamentals agree with it.
Generated interfaces share the same defaults, so their failures form a short list. That makes the fix a checklist run after generation, most of it checkable against source code, instead of better taste in the prompt.
Before picking a model or a framework, sort each decision in the workflow: a person, a plain rule, classical ML, or a generative model. Every wrong assignment has its own cost, and the costs run in both directions.
AI-assisted CAN bus reverse engineering works when the model sits inside a pipeline of deterministic decoding and validation. It fails when the model is asked to guess signals from raw frames.
RAG is a ladder of levels, and each level was added to fix a specific failure in the one below. A solo builder should climb only as far as their own failures require.
There are three ways to give an agent a tool. The choice mostly comes down to who else needs the tool and what it costs in context, not which one is best.
A skill that holds up is built in a fixed order: do the task by hand, write down what you did, turn it into a skill, then test it against cases. Most broken skills skipped a step.
What makes a coding agent reliable is mostly the layer around the model: the context it is handed, its tools, the checks on its output and the code that orchestrates sessions. That layer is ordinary engineering.
For a single person's knowledge base, markdown articles compiled by an LLM plus a generated index answer questions without retrieval infrastructure. That holds up to a size limit, and the limit is worth naming.
An agent loop is a trigger, an action and a stop condition. The stop condition is the only hard part, because done has to be something a machine can check.
When I cannot judge a plan myself, the model that wrote it should not be its only reviewer. A model from a different vendor can critique it in a bounded loop, and any factual claim the reviewer makes still has to be checked.
Running a coding agent fully offline is practical now. The binding constraint is GPU memory, and the honest question is what quality you give up at each memory tier.
The CoALA taxonomy names four kinds of agent memory: working, semantic, procedural and episodic. In practice each one is a plain file, and knowing which kind a fact is tells me which file it belongs in.
A skill library gets worse past a certain size, because everything in it competes for the model's attention. Deleting is a real optimization, and the test for keeping a skill is whether the agent does better with it than without it.
Compaction keeps a general summary that the tool chose. A handoff document keeps what I decided the next session needs. They solve different problems, and the handoff is the one worth writing by hand.
Devsplate argues vibe coding fails horribly in C++ because the language has no runtime safety net. The directional claim is right. The percentages are unsourced. And the question the video never asks is the most important one for working C++ engineers: is the AI even checking for these issues?
The CEO of Obsidian runs a flat vault with one folder. I run ten numbered folders. Both are right. Here is what changes when an LLM is your primary reader.
Deep transformers dilute early information as layer outputs accumulate. Notes on attention residuals, a proposed fix for that signal dilution.
Nate B Jones argues the AI job market has split: generalist roles are flat while roles that design, run and manage AI systems go unfilled. Notes on the seven skills he says employers cannot find.
Google's Stitch puts a whole design system in one plain-text DESIGN.md file that agents read as context. Notes on why that makes generated UI consistent, and on creative tools moving toward CLI and MCP interfaces.
Traditional RAG falls apart when your codebase spans hundreds of repos. Graph RAG fixes retrieval -- ICM fixes the workflow around it. Here's how to combine them.
A breakdown of when to use coding agent SDKs (Claude Agent SDK, Codex SDK) vs traditional frameworks (Pydantic AI, LangGraph) for building AI agents -- plus the current state of RAG and why it's not dead.
Jake Van Clief walks through building a content creation workspace from scratch using Claude's desktop co-work mode -- without VS Code. He builds a four-pillar content system with modular brand voice markdown files and makes the case for subscription-based AI workflows over API-heavy custom frameworks.
How structured folders and markdown files can replace complex frameworks for building powerful, persistent AI workspaces.
Programming feels hard not because you're bad at it, but because you're actually trying to understand it. A breakdown of why the struggle is the real skill-builder.