Most AI research fails for one reason: it mixes three things that need different burdens of proof. Fact, inference, opinion. Tag every line before the memo misleads.
All briefs
Wall Street Prompt · LinkedIn one-page
The Prompt That Separates Fact, Inference And Opinion
LinkedIn one-page
Most AI research fails because it mixes three things. Tag every line before the memo misleads.
LinkedIn1080 × 1350
References & validationReference missing
No specific creator reference resolved. Add an observed post before production.
Reference creator: Research methodology creators
No reference links resolved.
Post content
Dense one-page content — sections, tables and frameworks
THE MISTAKE: ONE BLENDED PARAGRAPH
Ask AI to research a company and it hands you smooth prose. The problem is what is hiding inside the prose. A verifiable number, a chain of reasoning, and a vibe all wear the same confident tone. You cannot size a position on a paragraph you cannot audit.
- Fact: pulled from a 10-K, 10-Q, or earnings call transcript. Checkable against the source.
- Inference: a conclusion the model built by connecting facts. Only as strong as its links.
- Opinion: a judgment, a tone read, a 'looks attractive.' Not evidence. Often the model guessing.
Why it matters: when these three blend, you inherit the model's confidence without its sources. The memo reads clean and reasons dirty.
THE THREE LAYERS, DEFINED
Before you trust a line, you have to know which layer it lives on. Same sentence, very different burden of proof.
| Layer | What it is | How to verify | Failure if blended |
|---|---|---|---|
| FACT | Reported figure or stated event | Match to filing line item or transcript quote | Hallucinated number reads as truth |
| INFERENCE | Conclusion built from facts | Re-trace the logic chain step by step | Weak link passes as proven |
| OPINION | Judgment, tone, or estimate | Label as the model's view, not data | Vibe gets sized like evidence |
Key: a fact has a source. An inference has a chain. An opinion has neither. Make the model say which.
THE PROMPT STRUCTURE
You do not ask for a cleaner summary. You force the output into three tagged columns so the layers cannot hide in each other. Paste this skeleton, then drop in your filing or transcript.
| Prompt block | Instruction to the model | What it forces |
|---|---|---|
| ROLE | Act as a research analyst building an IC memo input | Sets the standard of proof |
| TAG RULE | Label every claim as [FACT], [INFERENCE], or [OPINION] | No untagged sentences allowed |
| SOURCE RULE | For each [FACT], cite the exact filing line or transcript quote | Kills unsourced numbers |
| CHAIN RULE | For each [INFERENCE], list the facts it rests on | Exposes weak logic |
| FLAG RULE | Mark any claim you cannot source as [UNVERIFIED] | Surfaces the guesses |
Tip: the [UNVERIFIED] flag is the most valuable line in the output. It tells you exactly where to do your own work.
BEFORE VS AFTER: THE SAME RESEARCH
Same question, same model. The only change is the structure you demand. One output you skim. The other you can challenge.
| Dimension | Blended output (default) | Tagged output (this prompt) |
|---|---|---|
| Reads as | One confident paragraph | Labeled claims by layer |
| Auditable | No, sources are buried | Yes, every fact cites a line |
| Risk surface | Hidden inside prose | Flagged as [UNVERIFIED] |
| Your job | Trust or re-do it all | Verify the facts, judge the rest |
| Memo input | Misleading and smooth | Honest and structured |
Why it matters: the tagged version is longer and uglier. That is the point. Ugly research you can verify beats clean research you cannot.
THE VERIFY-BEFORE-YOU-SIZE WORKFLOW
The prompt produces the draft. You do the discipline. Run every tagged output through this loop before a single number reaches a memo or a position.
| Step | Action | Decision gate |
|---|---|---|
| 1. Pull | Feed the filing or transcript, run the tagged prompt | Output uses all three labels |
| 2. Spot-check facts | Verify 3 to 5 [FACT] lines against the source | Any mismatch means re-run |
| 3. Trace inferences | Follow each [INFERENCE] to its facts | Broken chain gets demoted to opinion |
| 4. Quarantine opinion | Move [OPINION] lines to a separate note | Never sized as evidence |
| 5. Resolve flags | Do your own work on every [UNVERIFIED] | No flags left before sizing |
Heads up: the model can tag a fabricated number as [FACT]. The tag tells you where to look. It does not replace the look. Human judges. AI builds.
Caption
LinkedIn post copy
Visual design notes
- Near-black forest-green background (#0B1410 to #0E1A14), one teal-green accent (#1FB58F) used only for the emphasized word, the subtitle banner, and table header rules. Everything else off-white on dark.
- Heavy condensed sans headline (Anton or Bebas Neue) for the title, 'Separates' set in the teal accent. Subtitle sits in a thin full-width teal-rule banner directly under the header, left-aligned.
- Five dense 2-to-4 column tables stacked vertically, left-aligned, with hairline row dividers in muted green. Header row of each table gets a solid accent underline, not a fill, to keep it clean and printable.
- Use a small color-coded tag chip system for FACT (solid accent), INFERENCE (outlined), OPINION (muted gray) so the three-layer concept is reinforced visually wherever the labels appear.
- Section headings in small-caps condensed with a left tick mark in accent. 'Why it matters / Key / Tip / Heads up' notes set in italic with a thin left border bar so they read as teacher asides.
- Tight vertical rhythm to fit 5 sections in 1080x1350: limit intro lines to two rows, keep table cells terse. Footer pinned to bottom with WSP wordmark lockup left, accent divider above it.
Production checklist
- ☐Design the 1080x1350 one-pager on the WSP dark-forest template: header with title (emphasize 'Separates' in accent), teal subtitle banner, footer lockup.
- ☐Build all five tables with consistent column widths, hairline dividers, and accent-underlined header rows. Verify the 4-column tables (Sections 2 and 3) stay legible at export size.
- ☐Add the FACT / INFERENCE / OPINION tag-chip color system and apply it consistently across the layer table, prompt table, and before/after table.
- ☐Set the four teacher notes (Why it matters, Key, Tip, Heads up) with the italic left-border treatment so they read as asides, not body copy.
- ☐Add header and footer WSP branding, confirm one-accent discipline (no second color creeps in), and proof for any em dashes (use periods or commas instead).
- ☐Export PNG at 1080x1350 for the LinkedIn feed and a matching PDF version for DM and lead-magnet delivery. QA both for cropped tables and legibility on mobile.
Follow Dave Wang · Wall Street Prompt · For more AI investing workflows · COMMENT & REPOST
CTA
Save this prompt skeleton. Next time AI hands you research, make it tag every line before you trust a single number. Comment FACT and I will share the full IC-memo prompt template.