AI Tooling, skills.md & Workflows
I treat AI as part of my engineering environment rather than a chat box. That means writing repeatable skills.md instructions, keeping a simple workflow that catches bad assumptions early, and building MCPs when models need real tool access.
Personal skills.md
As I write custom skills for my own stack and recurring tasks, I publish them here. Keeping instructions in versioned skills.md files keeps agent behavior consistent across repos without re-explaining constraints every session.
software-idea-feasibility-validator
SkillAdversarial, non-sycophantic desk validation of a software idea, tool, or system before writing code. Evaluates technical feasibility, architectural soundness, workflow friction, and whether the design actually accomplishes the expected goal, ending with falsifiable kill or re-architect criteria.
- Dialectic Stress-Test
- Compresses the goal, steelmans three structural variants, then attacks the load-bearing assumptions in an isolated pass.
- Pipeline & Effort Math
- Models compound pipeline reliability, maintenance-to-value ratios, and whether a bare prompt or CLI already solves 80% of it.
- Falsifiable Kill Criteria
- Outputs concrete, dated thresholds and a bounded 1 to 3 day technical spike instead of vague "test it later" advice.
MCP Servers
Soon enough.
Workflow
Start with /grill-me
My current workflow is intentionally simple. Before letting an agent scaffold or refactor a system, I run a short structured interview ( /grill-me ) to lock down audience, architecture, and trade-offs one question at a time. It takes two minutes and prevents throwing away half a build later.