Skills & Prompts
The second time you explain a workflow to a model, you should be writing it down instead.
What This Covers
Section titled “What This Covers”Turning one-off prompting wins into reusable assets: skills (packaged workflows the harness can invoke), prompt templates, and the discipline of maintaining them like code.
Key Principles
Section titled “Key Principles”- Encode the pattern once, stop re-paying for discovery
- Version and test prompts like code
- Periodically re-justify each skill against plain prompts on current models
What Makes a Good Skill
Section titled “What Makes a Good Skill”- One job: a skill that does three things gets invoked wrong for all of them
- Trigger clarity: the description determines when it fires; write it for the router, not the reader
- Enforcement over exhortation: a script that refuses bad output beats a paragraph asking for good output
- Progressive disclosure: load detail on demand rather than front-loading everything into context
Prompt Craft That Transfers
Section titled “Prompt Craft That Transfers”- Show, don’t describe: Few-Shot Examples outperform abstract instructions
- Constrain the output shape: Structured Output makes results machine-checkable
- Make reasoning visible when the task is judgment-heavy (Chain of Thought)
- Build in a second pass: Self-Critique catches what the first draft missed
Key Considerations
Section titled “Key Considerations”- Is this worth encoding, or is it a one-off? (Frequency times cost of rediscovery)
- Who owns the skill when the underlying tool or API changes?
- How do you know it still works? A skill with no test rots invisibly.
- Does a newer model make the skill unnecessary? Retire aggressively.
Anti-patterns
Section titled “Anti-patterns”- Skills that duplicate what the model already does well on its own
- Prompt archaeology: nobody knows why a clause is there, so nobody removes it
- Untested skills trusted because they worked once
- A skill library that grows but never shrinks