Skills for agentic workflows.
See ./code for coding-specific skills. I use a very hands-on and highly structured
approach when coding with agents. Note that I use amp.
It's easy to get lost in what is going on, so my workflow self-documents the thinking behind the implementation. When I'm diligent, this is where 60% of my time goes.
Features start by turning an idea into a plan. In most cases the first draft of the plan is written by the agent. I'll then go make edits as needed, making changes myself is usually faster than asking the agent to do it. Plans consist of granular steps, each step corresponding to roughly the amount of work that would go into a commit.
Once implemented, a corresponding proof is written (see below) that provides additional documentation on exactly what was done, why, and how a human can verify it. This not only helps me verify work (or debug why the implementation is wrong) but I have found it to improve the accuracy of the result.
I maintain that agents are not intelligent in any traditional sense. They require a lot of coaxing to correctly implement what they are told in a reasonable way. One of the more effective strategies I've found is to have the agent generate a proof of its work. This is not entirely novel and is inspired by others like Simon Willison.
Note that this is not actual proof that the code works; that is ultimately the job of the human. This is only for the agent to deliver more reliably. Do your job and put the effort in to understand the implementation and own the final result.
- https://simonwillison.net/2026/Feb/10/showboat-and-rodney/
- https://simonwillison.net/2025/Dec/18/code-proven-to-work/
One gripe I have with agentic coding is that as agents improve and I change my mind about how I want things done, the codebase starts to become disjointed in philosophy and style. The code audit skill helps address this by carefully stepping through and making things more consistent with the current ruleset.