Best practices

Getting better Reports starts with better prompts. This guide distils what works.

Principle 1: Specify scope tightly

The system delivers what you ask for. Vague prompts produce vague reports.

Bad: "Tell me about the semiconductor industry."

Better: "Summarise the 2025 financial performance of the top 5 fabless semiconductor companies by revenue, with focus on AI-accelerator product lines."

Best: "Top 5 fabless semiconductor companies by 2025 revenue: revenue growth YoY, AI-accelerator product line revenue share, gross margin trend, R&D spend as % of revenue, and one notable strategic risk per company. Use SEC filings as primary source where applicable."

Principle 2: Name your output shape

The system can produce tables, bullet lists, scoring matrices, executive summaries, legal memos, and narrative reports. Say which.

Principle 3: Specify time windows

Models without specific dates can mix recent and dated information. Always say "as of [date/quarter/year]".

Bad: "Latest trends in vector databases."

Better: "Vector database landscape as of Q1 2026: which products are production-grade, with citations to recent benchmarks (2025 or later only)."

Principle 4: Tell it what NOT to do

Counter-instructions are as powerful as instructions.

• "Do not include marketing claims; cite primary sources only."
• "Do not list companies I haven't named; restrict to the 5 listed."
• "Do not speculate on future events beyond 12 months."
• "Skip preamble and disclaimers; output direct findings."

Principle 5: Use Refine before Submit

The Refine tool is free (200/day) and uses no Report quota. It tightens vague prompts. Always Refine first if you're under 1,500 chars and your prompt isn't already specific.

Workflow:

1. Type a rough question
2. Click Refine — review the rewrite
3. Edit further if needed
4. Submit

Principle 6: Attach instead of paste

For long context (>1,000 chars of background), upload as a file (.md, .txt, .pdf, .docx) rather than pasting into the prompt. This:

• Keeps your prompt readable
• Treats context as evidence, not as the question
• Doesn't bloat your local draft history

Principle 7: Attach large context as files

Every Report uses the same full-strength synthesis, so there's no tier to upgrade — but you can still help the engine by feeding it the right evidence. For high-stakes or large-context work:

• Upload primary documents (filings, contracts, datasets) instead of pasting them
• Name the specific sources you trust ("use the 10-K and the earnings-call transcript")
• Keep the prompt to the question; let attachments and your Vault carry the context

The engine handles inputs up to roughly 250,000 tokens; beyond that, trim to the most relevant material.

Principle 8: Iterate, don't perfect

You won't write the perfect prompt on the first try. Strategy:

1. Run once to gauge how the system interprets your question
2. Read the Report — note where it went wide or shallow
3. Refine the prompt to tighten scope or demand primary sources
4. Re-run

Each run counts as one Report. Two or three focused iterations almost always beat one over-engineered prompt.

Principle 9: Avoid forbidden patterns

The system will decline or degrade outputs for:

• Personal advice (medical/legal/financial without professional review)
• Detailed operational instructions in CBRN / weapons / malware domains
• Identifiable-person attack vectors
• Election interference content
• Content sexualising minors (zero tolerance)

If you're a qualified professional needing technical detail in a regulated domain, the system may still produce material with appropriate caveats. Frame the request honestly.

Principle 10: Verify before you cite

Reports are research synthesis, not court-admissible evidence. Best workflow when stakes are high:

1. Read the Report's claims
2. Click each inline citation to open its source
3. Confirm the source supports the claim as stated
4. Spot-check 2–3 specific numerical claims against the source

If a claim doesn't check out, your prompt may have triggered a hallucination — refine and re-run.

Anti-patterns to avoid

Examples from real workflows

Legal research

"Compare California, New York, and Texas state law as of 2025 on [specific issue]: statute citations, key case law citations, practical differences for [specific party type]. Bluebook citation style. Flag any pending legislation."

Investment due diligence

"Conduct due diligence on [Company]'s [Year] [period] performance: revenue and operating margin trend, debt structure changes, named executive compensation, top 3 disclosed risks from 10-K, and any whistleblower or material litigation disclosed. Cite SEC filings only."

Competitive intelligence

"[Competitor] product line as of Q1 2026: feature parity vs [our product] on [list 5 features], pricing observed in public materials, recent customer wins disclosed via case studies or press releases. Skip rumors; primary sources only."

Academic literature

"Survey the 2024–2026 literature on [research topic]: 10–15 most-cited papers, each with one-sentence finding, methodology category, and link. Group by sub-topic."

Policy analysis

"Compare proposed [regulation] across [Jurisdiction A] and [Jurisdiction B] as of Q1 2026: scope, enforcement mechanism, penalties, implementation timeline, public comments received. Cite official government sources."