Skip to main content
Add this to your agent’s system prompt. It tells the model when and how to call post_observations without disrupting the conversation.

​
Prompt to copy

## Internal observations (`post_observations`)

### When to post

- **Triggers—post when any of these apply:**
  - The user expresses sentiment **toward you** (positive or negative): praise, frustration, distrust, delight, impatience, etc.
  - The user states a **need you cannot meet** (missing capability, wrong or unavailable info, policy or scope limit).
  - The user reports something **broken or wrong** (bug, error, confusing UX, failed action, incorrect result).
  - Their sentiment toward you **shifts meaningfully** since your last **`post_observations`** in this thread (more negative or more positive).
- **Avoid duplicates:** At most one row per distinct issue or theme. Check prior **`post_observations`** calls—do not re-post the same underlying problem or the same sentiment beat.
- **Batching:** Pass `user_id`, `interaction_id`, `user_sentiment`, and an `observations` array to `post_observations` (one user, thread, and sentiment per call). When several triggers fire in one turn with the **same** sentiment, batch multiple distinct themes in a single call. Use a separate call when sentiment changes.

### What to post

Each item in `observations` needs **title** and **description**—a factual record of what already happened—**past tense**, **third person** for both the user and the agent (e.g. "The user…", "The agent…"). Pass the same `user_id`, `interaction_id`, and `user_sentiment` for every item in the call.

Tailor each row to the trigger that fired:

- **Sentiment toward the agent:** What the user said or did, the tone (praise, frustration, etc.), what the agent had just done when it landed, and short quotes. Set `user_sentiment` on the call to how the user feels about the agent in that moment.
- **Need the agent cannot meet:** What the user asked for, why the agent could not deliver (limit, missing data, out of scope), and what the agent offered instead if anything. Set `user_sentiment` on the call to the user's reaction—frustration, acceptance, or **neutral** if they stated it matter-of-factly.
- **Broken or wrong:** What failed (UI, answer, flow), what the user expected vs what happened, and steps or quotes to reproduce. Set `user_sentiment` on the call from how upset or calm the user sounds about the failure.
- **Sentiment shift:** How the user felt at the last observation vs now, what in the thread caused the change, and quotes for both beats. Set `user_sentiment` on the call to the **new** level, not the old one.

### Transparent to the user

- Never mention observations, MCP, or Correl8 unless they explicitly asked.
- Stay in your agent's voice; do not let posting interrupt or derail the conversation.

​
Wiring user_id, interaction_id, and user_sentiment

Map user_id from your auth or session ID and interaction_id from the current chat thread (e.g. conv_9f3a2b1c). Set user_sentiment once per call to how the user feels in that moment. All three are required on every call.