Prompting the Agent

Strong Agent tasks describe the business outcome, the inputs, the constraints that matter, and what a good answer looks like. They do not need to prescribe every tool call.

A practical task shape

1. State the goal.
2. Supply the seed data or identity anchors.
3. Explain the matching or decision criteria.
4. Define the evidence standard.
5. Name any exclusions or safety constraints.
6. Describe the desired presentation.

Analyze the supplied customer accounts and identify stage-matched lookalikes.

For each seed:
- resolve the canonical company and domain;
- determine current company stage using dated public evidence;
- preserve the original POC and verify their current role;
- find lookalikes with similar workflow, buyer persona, employee scale, and funding stage;
- exclude companies that are materially larger or later-stage than the seed;
- cite the official site, company LinkedIn, and an independent source where available;
- separate confirmed facts from inference and explain low-confidence fields.

Give the agent judgment, not ambiguity

Prefer: “Match stage using funding, employee scale, and commercial maturity; explain exceptions.”

Avoid: “Find similar companies.”

The first version gives the model room to research while still defining what similarity means.

Tell users when research is deep

Deep research can take longer and use more ACU than a direct lookup. In user-facing products, set that expectation before or during the run and stream progress when possible.

Ask for evidence at the claim level

Useful instructions include:

• “Include a source URL for every stage or fundraising claim.”
• “Never infer payment status.”
• “Use null when no reliable source supports the field.”
• “Label a claim as inference when the evidence is indirect.”

Use schemas only when the caller needs them

Do not add output_schema to every turn. Conversational synthesis is usually better for exploration and follow-ups. Add a schema when downstream code, a table, or an API consumer requires exact fields.