Skip to main content
Guidelines tell your agent when and how to post observations and conduct research. They are markdown text stored per project — not copied into your agent’s system prompt. Your agent loads them by calling the MCP tool get_correl8_guidelines(user_id=…) once at conversation start (see Agent system prompt).

​
Two guideline types

Correl8 ships defaults for both. Customize them to match your product, agent voice, and research style. Active research studies are loaded separately via the get_active_studies tool (question and eligibility per study, plus any answer already logged for the user). Study content is configured on the Research page — not in the research guidelines text.

​
Edit guidelines

You can edit either type in three places:
  • Settings → Details — inline fields for Observation guidelines and Research guidelines
  • Observations — Observation guidelines button above the list
  • Research — Research guidelines button above the list
Click a field or button to open the editor, change the markdown, and save. Changes apply to the next conversation that loads guidelines — agents do not re-read mid-thread.

​
Reset to defaults

When a field is customized, Reset to default appears. Confirm to restore Correl8’s built-in text for that type. Resetting one type does not affect the other.

​
Tuning tips

Observation guidelines — Adjust triggers to your product: which frustrations, missing capabilities, or sentiment shifts matter. Add examples in your voice (“The user…”, “The agent…”). Tighten or relax duplicate and batching rules if your agent over- or under-posts. Research guidelines — Set how assertive the agent should be when gathering signal, how many follow-ups are OK, and how to balance user requests vs research. Keep study-specific logic in each study’s eligibility criteria, not here. Keep the system prompt short — The agent system prompt only needs to say read guidelines once. Put behavioral detail in guidelines so you can iterate in the app without redeploying your agent. Verify in the app — After changes, run a test conversation and check Observations or Research for new rows. If behavior is off, edit guidelines again rather than stacking rules in the system prompt.