Wall Street PromptBrand
All briefs
Wall Street Prompt · Brand

The Finance Agent Guardrail Checklist

Carousel — cover + one idea per slide + caption. This is a post, not a video script.

Carousel
Instagram carousel + LinkedIn documentInstagram carousel8 slides
References & validationReference missing

No specific creator reference resolved. Add an observed post before production.

Reference creator: Deep Insight Labs / Simon Willison

No reference links resolved.

Slides

One idea per slide

Cover
THE FINANCE AGENT
GUARDRAIL CHECKLIST

Before I trust an AI finance agent,
I check 5 things.

The agent builds. You approve.
(swipe)
Slide 1
THE FINANCE AGENT
GUARDRAIL CHECKLIST

Before I trust an AI finance agent,
I check 5 things.

Source. Timestamp. Math.
Conflict. Approval.

The agent builds. You approve.
(swipe)
Slide 2
First, the trap.

The agent hands you a clean,
confident answer.

Confident is not the same
as verified.

The failure mode is quiet:
no wrong tone, no error message.
Just a number you never checked.

These five checks make it loud.
Slide 3
1. SOURCE

The common mistake: trusting a figure
because it sounds specific.

-> Every number traces to a real document
-> A filing, a transcript, a dataset
-> The agent shows you where it came from
-> No source, no use

Unsourced is not data.
It is a guess in a suit.
Slide 4
2. TIMESTAMP

The failure mode: right number,
wrong period.

-> Confirm the as-of date on every figure
-> This quarter, not last, unless you said so
-> Stale data reads exactly like fresh data
-> Make the agent state the date, not assume it

Old numbers do not announce
that they are old.
Slide 5
3. MATH

Agents narrate arithmetic
they did not actually run.

-> Re-add the sum yourself
-> Re-run the ratio, the growth rate, the margin
-> Check that the parts tie to the total
-> A fluent explanation is not a correct one

The story can be perfect
while the math is wrong.
Slide 6
4. CONFLICT

The disciplined move: ask what
disagrees with the answer.

-> Make the agent surface the counter-evidence
-> Two sources, two numbers? Name the gap
-> A clean answer with no tension is suspicious
-> Disagreement is the work, not an error to hide

If nothing contradicts it,
you have not looked hard enough.
Slide 7
5. HUMAN APPROVAL

The last guardrail, and the
non-negotiable one.

-> A person signs off before anything ships
-> The agent drafts, it does not decide
-> No output goes out unreviewed
-> The agent never gets the last word

Automate the building.
Never automate the approval.
Slide 8
RUN THE CHECKLIST

-> Source: where did this come from?
-> Timestamp: is it the right period?
-> Math: does the arithmetic tie?
-> Conflict: what disagrees?
-> Approval: who signed off?

The agent builds the work.
You approve the answer.

Save this. Run it before you trust the output.

Educational only. Not advice.

Caption

Paste under the carousel

An AI finance agent gives you speed and a clean, confident answer. Neither of those is verification. Before I trust the output, I run five guardrails.

-> Source: every figure traces to a real document, a filing, transcript, or dataset. No source, no use. -> Timestamp: confirm the as-of date. Right number, wrong period is one of the most common quiet failures. -> Math: re-run the arithmetic yourself. Agents narrate sums and ratios they never actually computed. -> Conflict: ask what disagrees. A clean answer with zero tension usually means you have not looked hard enough. -> Human approval: a person signs off before anything ships. The agent drafts and structures. It does not decide.

Notice the split. The agent is doing the building, fast. It is not the one allowed to declare the answer correct or send it out the door. That stays with you.

Automate the building. Never automate the approval.

Save this and run it before you trust the next agent output.

Educational content only. Not investment advice, and not a recommendation to buy, sell, or hold any security. Wall Street Prompt. Always verify against the primary source.

Design notes

WSP system across all slides. Near-black background, bone/off-white type, single green accent used only for arrow glyphs (->), slide numbers, and one keyword per slide. Cover: heavy condensed all-caps title (two lines), subhook in regular weight, then the five one-word checks as a tight inline row, footer (swipe) cue. Body slides: large green numbered label at top (1. SOURCE), a short concede-then-sharpen line in muted bone, then the arrow list (one idea per arrow, generous spacing), and a punchy two-line closer at the bottom set slightly larger or bold for the 'mic-drop' beat. Keep every slide visually identical in grid: same top margin, same label position, same list block, same closer position, so it reads as a lecture deck. Final slide repeats the five checks as a compact recap, the two-line principle, the save line, and a small 'Educational only. Not advice.' footer. Persistent @wallstreetprompt handle bottom-corner on every slide. One accent color only, no gradients, no shadows, tight tracking on headlines.

CTA

Save this and run it before you trust the next agent output. Comment 'AUDIT' and I'll send the full checklist.

Wall Street Prompt
Wall Street Prompt — internal