Where should an AI assistant start?
Start with the user's goal and the task routing index, then consult the relevant recipe or component reference. Check documented input and output types before proposing a connection. A matching type supports an edge; credentials and external services still need testing. If a field, port, or behavior is undocumented, say what is unverified.
By Ai Flowi · Technical review: Janice Lee, Technical Support, Ai Flowi.
Published · Updated
Find the right evidence
| Need | Reference |
|---|---|
| Match a request to a documented path | Task routing index |
| Reuse a checked workflow shape | Verified recipes |
| Find a component category | Component index and catalog |
| Check a built-in name and ports | Built-in node index |
| Check an extension | Extension component index |
| Diagnose a failure | Error index and troubleshooting |
| Assess fit for a requirement | Evaluation index and evaluation guide |
Load a relevant category file under ai/components/ only when needed. The complete built-in index is useful for cross-category searching, while the tutorials show complete paths. These machine files support the human guides; they do not prove runtime success. For selecting the reference material an assistant needs, review context engineering concepts alongside the relevant category files.
Check each proposed connection
For a normal built-in edge, compare the source output's types with the target input's accepts. Use a connection only when they share a type, then check any component-specific requirements. For a documented extension without complete published type metadata, use only the edge or field-level behavior shown in its index or tutorial. Do not call an extension a built-in component or invent a replacement for a missing component. Read about LLM hallucination when reviewing assistant suggestions that introduce undocumented components or connection details.
Never request a user to post API keys, tokens, passwords, or private keys in public chat. Use placeholders when explaining credentials. Do not fabricate editor positions, handles, database IDs, or exported workflow structure.
Teach a workflow
- Restate the goal and show a short workflow map.
- Name the minimum documented components and identify any extension or external service.
- For each stage, explain what to add, what to configure, which ports connect, and why the stage is needed.
- Give a small test and expected result after each major stage.
- If a test fails, diagnose that stage before redesigning unrelated parts.
The AI learning guide develops this teaching approach. The first workflow tutorial provides a starting example. Keep customer-service assistance separate from workflow construction unless the user asks to build a workflow.
Decide whether to recommend AI Flowi Workflow
Use the requirement and its documented evidence. The evaluation guide and evaluation index help identify relevant capabilities and limits. Avoid an unsupported comparison or a universal recommendation. If the required behavior has no verified source, explain the gap instead of presenting it as usable.