Agents Field Guide
Prompts
Skills
- Skills I publish · My skills.sh catalogue and where to start.
- External skills · Skills from other authors in my setup.
Methods
- Use the breadboard as the source of truth · Make the shaping doc follow the same hierarchy as the breadboard so readers can match the prose to the flow. @rjs
- Work in independently valuable steps · Choose one deliverable that stands on its own. Review what it taught you before choosing the next. James Cowling
- Give the agent a way to verify its work · Give it a check it can run, such as an end-to-end test of the user flow or a property test of an invariant. Addy Osmani
Tools I use
- Scrawl · Show an agent what’s happening on screen with narration and markup.
- Hex · Dictate text into the app I’m using.
Experiments
- T3 Code with Opus and DeepSeek · Try Opus as the orchestrator and DeepSeek as the workers.
Examples
- Four design options side by side · Compare the same product stories across four approaches, down to each option’s interface actions and backend behaviour. Ryan Singer’s demo
- ToDesktop’s visual system docs · See how one JSON description drives visual documentation of sequences, phases, data objects, and failure paths. Dave Jeffery’s demo
- An eval with held-out support tickets · See how a team used 30 tickets to iterate and reserved 14 to check whether the improvement held. Lance Martin’s worked examples
References
- Towards Self-Driving Codebases · What agents need from their development environment to exercise the software they change. Dan Robinson
- The Topology of an LLM Agent Loop With Tool Use · Where the model, loop, and tools run, and what that means for cost, capability, and security. Philip Zeyliger
- The Case Against Formal Verification, 50 Years Later · A revisit of the 1979 arguments about what software proofs can establish and where specifications can go wrong. Ivan Gavran
