Ask for source-backed facts, distinguish facts from hypotheses, and require an explicit “not found” when evidence is unavailable. Human review remains required before research affects scoring, CRM data, or messaging.
[Image in Confluence: Bildschirmfoto 2026-07-30 um 10.34.43.png]
Before you start
- Define the decision the research must support.
- Choose the company, person, signal, region, and time period in scope.
- List the public or approved sources that may be used.
- Decide the output language and maximum useful length.
- Identify sensitive information that must not be collected or used.
Build the prompt from six components
| Component | Specify |
|---|
| Decision | Qualify, prioritise, verify, or select a conversation angle. |
| Scope | Company, person, event, geography, and time window. |
| Questions | A short list of verifiable questions. |
| Evidence | A source or public proof for every factual claim. |
| Output | Language, headings, concise format, and not-found behaviour. |
| Safety | No private, sensitive, credential, or unsupported personal information. |
Separate facts from interpretation
Ask the result to label direct evidence, derived interpretation, and unresolved questions separately. A plausible sentence without a source is still unverified.
- Fact: a source directly supports the statement.
- Hypothesis: an interpretation that must be labelled and reviewed.
- Not found: available sources do not support an answer.
- Outdated: evidence may no longer reflect the current company or role.
- Sensitive: information that must not be used for outreach.
Steps
- **Define the decision. **Write what will change if the research confirms or rejects the hypothesis.
- **Add scope and exclusions. **Name the signal, time window, geography, company or person, and what should be ignored.
- **Ask focused questions. **Use a small set of questions that can be answered with evidence.
- **Require sources. **Request proof for every factual claim and “not found” when none exists.
- **Specify the output. **Choose language, headings, length, and separate facts, hypotheses, and gaps.
- **Test known accounts. **Run the prompt on at least five positive, negative, and edge cases.
- **Review before use. **Check sources, identity, dates, and relevance before copying results into scoring, a message, or the CRM.
Review and safety checks
- Every question supports a real decision.
- The scope and time window are clear.
- Factual claims include usable evidence.
- Hypotheses and gaps are labelled.
- Sensitive or private information is absent.
- The output is concise enough for review.
- No statement enters outreach or scoring without human validation.
Troubleshooting
| Problem | Solution |
|---|
| The result is long but not useful | Reduce the questions and connect each to qualify, prioritise, verify, or select an action. |
| The answer contains unsupported claims | Require sources and “not found”; remove unsupported text from downstream use. |
| Sources are old | Add a time window and require a publication or observation date. |
| The wrong person or company was researched | Add stable identifiers and current company context, then discard the mismatched result. |
| Different languages appear | Specify one output language and test records from multiple regions. |
Frequently asked questions
What should a Deep Research prompt start with?
Start with the decision, then define scope, questions, evidence, output, and safety rules.
How many questions should I ask?
Use the smallest set that changes the decision. Fewer focused questions are easier to verify.
Can results be used directly in a message?
No. Check the source, date, identity, relevance, and wording first.
What happens when no source is found?
The output should say “not found” or identify the gap, not generate a plausible substitute.
Can I ask for private personal information?
No. Use only necessary, approved, and appropriate sources and exclude sensitive information from outreach.
Where can the prompt be reused?
Use the Deep Research prompt or Templates workflow available in your workspace, then test each downstream use.
Keep reading
Set up copywriting, prompts, and templates
Use Copywriting settings for shared style and delivery rules, and use Templates for reusable prompts, replies, and Deep Research instructions. Each campaign still needs the correct list, signal, persona, language, and sender context.
Review and approve the Knowledge DB
Review every Knowledge DB section before approving it for research or copywriting. Approval means the content is trusted enough to influence customer facing output, so persuasive wording is not a substitute for accurate evidence.
Understand P1, P2, P3, and DQ
Use P1, P2, P3, and DQ to make consistent review decisions—not as automatic truth. A priority is useful only when the team shares a definition and checks company fit, persona fit, source evidence, and the next action together.
