A prompt is not a workflow. It is one step, and one step has no place to catch a wrong answer.
All briefs
Wall Street Prompt · LinkedIn one-page
A PROMPT IS NOT A WORKFLOW
LinkedIn one-page
Reliable AI finance work needs a mini-system around the prompt: Source, Prompt, Checklist, Human Review.
LinkedIn1080 × 1350
References & validationExact post missing
Reference creator resolved, but the exact source post was not captured. Capture the specific post (not just the profile) before treating as validated.
Reference creator: Nicolas Boucher / Notion template creators
Post content
Dense one-page content — sections, tables and frameworks
THE PROBLEM: A PROMPT IS A SINGLE STEP
Most analysts treat one clever prompt as the whole job. It is not. A prompt is one step inside a workflow. When the prompt is the only control, three things break.
| Common mistake | Failure mode | What it costs you |
|---|---|---|
| Prompt with no source attached | Model fills gaps with plausible numbers | Fabricated revenue, made-up margins in your memo |
| Prompt, then copy-paste the answer | No step to catch the error before it spreads | Wrong comp set flows into the IC deck |
| One prompt expected to do five jobs | Output is shallow on each, deep on none | Re-do the work, twice the time |
| No human read before sending | AI confidence reads as accuracy | You own a number you never checked |
Why it matters: In finance the cost of a confident wrong answer is higher than the cost of no answer. A single prompt has no place to catch the wrong answer.
THE FIX: A 4-PART MINI-SYSTEM
Wrap every prompt in four parts. Each part has one job. The model does three of them. You own the fourth.
| Part | Job | What goes in | Owner |
|---|---|---|---|
| 1. SOURCE | Ground the work in real data | 10-K, 10-Q, earnings call transcript, comp file, your notes | AI builds |
| 2. PROMPT | Tell the model the exact task and format | Role, task, output structure, constraints, what to refuse | AI builds |
| 3. CHECKLIST | Force a self-check before the answer | Required fields, math rules, citation rule, refusal rule | AI builds |
| 4. HUMAN REVIEW | Judge the output before it moves | Spot-check numbers vs source, sign off, or send back | Human judges |
Key: Human judges. AI builds. The model assembles the draft from the source against the checklist. You decide if it is right. The line never moves.
PART BY PART: WHAT EACH ONE PREVENTS
Each part of the system kills one specific failure. Drop a part and that failure comes back.
| Part | What it does | The failure it kills |
|---|---|---|
| Source | Pins every claim to a document the model can read | Hallucinated figures and invented quotes |
| Prompt | Locks task, format, and refusal rules so output is repeatable | Vague, drifting answers you cannot reuse |
| Checklist | Makes the model verify its own work before it answers | Missing fields, broken math, uncited numbers |
| Human Review | Adds judgment the model does not have | Confident output shipped without a human check |
Heads up: The checklist is the part most people skip. It is the cheapest to add and catches the most errors. Put the checklist inside the prompt as a do-not-answer-until list.
WORKED EXAMPLE: EARNINGS QUICK-READ
Same task, two ways. One prompt versus the mini-system. The task: summarize a quarter from the 10-Q and the call.
| Step | Just a prompt | The mini-system |
|---|---|---|
| Source | None. Model recalls from memory | Attach the 10-Q and the earnings call transcript |
| Prompt | Summarize NVDA's quarter | Extract revenue, gross margin, guidance, and 3 risks. Cite the page or quote for each. Mark anything not in the source as Not stated. |
| Checklist | None | Every number has a citation. Margin math checks. Guidance is a direct quote. Refuse to estimate. |
| Review | Read it, trust it | Spot-check 3 numbers against the filing. Sign off or send back. |
| Result | Plausible, partly invented, hard to defend | Sourced, checkable, ready for the IC memo |
Why it matters: The mini-system version takes ten extra minutes to set up once. After that you reuse it every quarter on every name. The single prompt is faster once and wrong eventually.
BUILD YOUR FIRST MINI-SYSTEM
Pick one recurring task you do by hand. Wrap it in the four parts. Reuse it. Below are the questions to answer for each part.
| Part | Question to answer | Example: comp table builder |
|---|---|---|
| Source | What document grounds the truth? | Latest 10-Ks plus your peer-set list |
| Prompt | What exact output and format do I need? | EV/EBITDA, P/E, rev growth for 6 peers in one table, cited |
| Checklist | What must be true before it answers? | All 6 names present, every multiple cited, no blanks |
| Review | What do I check before I trust it? | Re-pull 2 multiples by hand, confirm the peer set, sign off |
Tip: Save the prompt plus checklist as a reusable block. Next quarter you swap the source and run it again. That is the difference between a prompt and a workflow.
Caption
LinkedIn post copy
Visual design notes
- Canvas 1080x1350, paper bg #fbfaf6, near-black ink #191919 for text and borders, ONE green/teal accent #15d6d6 only. No coral on shadows. Text is black or white, never gray.
- Header centered: H1 'A PROMPT IS NOT A WORKFLOW' in Sergio Trendy ~56px uppercase, two lines with break after 'PROMPT IS NOT'. Emphasize the word WORKFLOW in teal, rest in ink. Subtitle in italic Montserrat below, then the teal subtitle banner bar spanning full width with the framing one-liner in ink on teal.
- Hero element below header: a horizontal 4-block flow diagram SOURCE -> PROMPT -> CHECKLIST -> HUMAN REVIEW with teal arrows between blocks. Blocks 1-3 carry a small 'AI builds' tag, block 4 carries a 'Human judges' tag in teal-pale. This anchors the whole post.
- All 5 sections use dense 2-3 column tables with 2.5-3px ink outer borders and 1.5px ink internal grid lines. First column gets soft beige fill (#f5f2ea) and bold weight. vertical-align: middle on every cell.
- Section headers as black bars with white text and a 3-4px teal bottom border, each with a teal numbered badge (1-5), 34px disc, black Montserrat 900 number, 2.5px ink border.
- 'Why it matters' / 'Heads up' / 'Key' / 'Tip' notes as teal-pale callout strips under each table with a small coral-free teal ribbon label top-left and ink body text.
- Alternating density: the 4-part system table (Section 2) and the worked-example table (Section 4) are the visual anchors, larger row padding. Section 1 and 3 tighter. Left-aligned content throughout, generous main gap 20-24px between sections.
Production checklist
- ☐Design the 1080x1350 one-pager in the WSP template: paper bg, ink text, single teal accent, Sergio Trendy H1 with WORKFLOW emphasized, teal subtitle banner.
- ☐Build the hero 4-block flow diagram (Source -> Prompt -> Checklist -> Human Review) with teal arrows and AI builds / Human judges tags.
- ☐Build all 5 section tables: mistake/failure/cost (3-col), the 4-part system (4-col), part-by-part (3-col), worked example (3-col), build-your-own (3-col). First-column beige fill, 1.5px ink grid, vertical-align middle.
- ☐Add teal numbered badges 1-5 and black section header bars with teal bottom border; add the Why it matters / Heads up / Key / Tip callout strips in teal-pale.
- ☐Add header and footer branding: Dave headshot 48px circle, 'Dave Wang' 21px, WSP footer line on teal bar with COMMENT & REPOST icons.
- ☐QA pass: verify nothing overflows 1350px, gaps follow the WSP spacing scale, no gray text, shadows ink or teal only, every table fully visible above the footer.
- ☐Export PNG at native 1080x1350 (pixelRatio 2) and PDF for DM delivery; name files 2026.06.02_A Prompt Is Not a Workflow.
Follow Dave Wang · Wall Street Prompt · For more AI investing workflows · COMMENT & REPOST
CTA
Comment SYSTEM and I will send you the 4-part mini-system template (Source, Prompt, Checklist, Review) as a one-page PDF you can reuse on any name.