01 · Mental model
Add continuity before autonomy
A chat is useful for exploration and one-off assistance. A project adds stable instructions, approved references, examples, and a repeatable workspace. A skill packages the procedure. An agent adds controlled action only when the task genuinely needs a loop and tools.
This capability ladder prevents teams from building an autonomous system when a documented workspace or reviewed skill would deliver the same value with less cost and attack surface.
02 · Visual explanation
03 · Compare and decide
Choose the smallest sufficient capability
| Decision lens | Use this | When |
|---|---|---|
| Chat | The human supplies context and drives every turn | Exploration, explanation, drafting |
| Project | Context and references should persist | Repeated analysis in one domain |
| Skill | The method and output must be versioned | Reviewable repeated expert work |
| Agent | The system must choose and execute several bounded steps | Multi-step evidence collection and synthesis |
04 · Cybersecurity example
Architecture review adoption path
A team repeatedly reviews cloud architecture diagrams and produces similar findings.
Start with chat to discover the useful questions.
Create a project with standards and examples.
Package the review method and schema as a skill.
Add read-only tools only when evidence collection is repetitive.
Outcome: The team earns automation through measured repetition rather than starting with autonomy.
05 · What to remember
The 60-second recall
Persistence and autonomy are different capabilities.
A repeatable task is often a skill before it is an agent.
Upgrade only when the current rung has a measurable limitation.
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