01 · Mental model
The deliverable is a controlled system, not a clever prompt
Start with a bounded, repeated task and an expert reviewer. Write non-goals before the system instruction. Threat-model the workflow before selecting tools. Define structured output before model orchestration. Package the expert method as a skill. Expose narrow read-only verbs. Add budgets and stop states before more intelligence.
This order makes the system testable at each boundary and prevents the prompt from becoming an informal substitute for architecture, authorization, or evidence.
02 · Visual explanation
03 · Compare and decide
A demo versus a releasable capability
| Decision lens | Demo proves | Release evidence proves |
|---|---|---|
| Possibility | One compelling run can work | Representative cases pass repeatedly |
| Control | The happy path looks bounded | Attacks and failures stop safely |
| Quality | Output looks useful | Rubrics, schemas, and reviewers agree |
| Operations | The UI responds | Cost, latency, logs, rollback, and ownership are ready |
04 · Cybersecurity example
First agent: architecture review assistant
The agent reviews a supplied design and produces findings without changing any environment.
Define required input and non-goals.
Package the review method and findings schema.
Use retrieval for official guidance and read-only metadata tools.
Attack the inputs, score results, and require expert acceptance.
Outcome: The first release is useful, bounded, observable, and intentionally unable to deploy changes.
05 · What to remember
The 60-second recall
Choose a task with a clear expert reviewer and definition of done.
Write non-goals and output contracts before prompts and tools.
Ship only what can be bounded, tested, observed, and stopped.
Teach-back prompt: Explain this concept to a teammate using the diagram, then name one failure mode and the control that stops it.
06 · Questions people ask
FAQ
07 · Primary sources