Skip to content

Teach an agent your report format

Useful for: avoiding the same formatting correction in every conversation. This is reviewed guidance stored for an agent and authenticated owner—not model training, and not a promise the model will always follow it.

Before starting: use a learning-capable gateway, a saved agent with a working model, and an authorized reviewer. Start with a test agent and non-secret text.

1. Enable deliberate learning

In the agent editor, enable Learning Notebook and save. For this exercise, leave Propose useful lessons off so you choose when a lesson is drafted. Equivalent fragment:

learning:
  enabled: true
  auto_propose: false
  max_proposals: 2

If the agent has an explicit built-in allowlist and you want in-chat learning tools, include learning.search, learning.read, and learning.propose. The dedicated Teach screen can draft a lesson without a tool-enabled chat.

2. Teach one specific preference

Web: Brain Mem → Lessons, then select the agent. iPhone: Agents → your agent → Learning.

Open Teach a preference or workflow and enter:

For my project status reports, use these headings in order: Summary,
Decisions needed, Risks, Sources. Keep Summary to three bullets or fewer.
If there is no evidence for a claim, mark it as unknown. This is my preferred
report format, not permission to browse, send messages, or change any records.

Choose Create learning draft. This uses one governed model call and can cost money. It can also return no suitable lesson; nothing is auto-approved.

3. Review the draft, not just its title

In Needs review, open the lesson. Check its content, type, source quote, and limitations. It should be a preference about this format—not a “fact” that all reports have no risks, or permission to send reports automatically.

Tick I reviewed this lesson, its sources and limitations, then choose Approve lesson. If it misrepresents your request, reject it and teach a clearer, narrower version instead. A citation proves where text came from, not that a claim is universally correct.

4. Test in a fresh conversation

Start a new chat with the same agent and owner identity. Do not repeat the format instructions. Ask:

Write my project status report. Fictional notes: the draft is done;
the demo has no owner; no launch date has been approved.

Expect the four headings, a short Summary, and honest unknowns. Inspect the lesson's load count as supporting evidence that guidance was available, but judge the actual answer yourself. A loaded lesson is not proof it helped.

5. Prove you can stop using it

Open the active lesson under In use, review it, and choose Disable lesson. Start another fresh chat. The lesson is no longer supplied to future runs. An existing conversation or already-running response may still contain the old guidance, and the model may independently choose the same headings.

Draft restoration creates a new reviewable draft; it does not silently reactivate the lesson. Approve the restoration only if you want it used again.

Boundary checks

  • Another agent or a different managed key/owner should not inherit this private lesson. A shared key is a shared identity; it does not isolate people.
  • Helpful/Needs improvement is feedback, not automatic approval or rewriting.
  • Do not teach passwords, API keys, temporary facts as timeless policy, or instructions to bypass safety controls.
  • If the model keeps ignoring the lesson, inspect approved content, identity, the selected agent, and a fresh run before adding duplicate drafts.

Continue with Learning Notebook for memory types, privacy, retrieval, and limits.