Wall Street PromptBrand
All briefs
Wall Street Prompt · LinkedIn one-page

The Finance Agent Guardrail Checklist

LinkedIn one-page
Before I trust an AI finance agent, I check 5 things: source, timestamp, math, conflict, human approval.
LinkedIn1080 × 1350
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.

Post content

Dense one-page content — sections, tables and frameworks

The Setup

An AI finance agent can read a 10-K, pull comps, and draft an IC memo in seconds. That speed is the trap. A fast answer feels like a checked answer. It is not. Before any agent output touches a model, a memo, or a position size, it passes 5 guardrails. Skip one and you are not researching. You are guessing with extra steps.

  • The agent builds. You judge. That order never reverses.
  • Every guardrail below maps a common mistake to its failure mode to a disciplined workflow.
  • None of this is about distrusting AI. It is about knowing exactly where it breaks.
Why it matters: An unguarded agent does not fail loudly. It fails confidently. Confident wrong is more expensive than visibly broken.

The 5 Guardrails

These run in order. Each one is a gate. Output that fails any gate goes back to the agent before it reaches your judgment.

#GuardrailThe question you askWhat it catches
1SourceWhere did this number come from, and can I open the document?Fabricated figures, blended sources, training-data memory dressed as research
2TimestampAs of when? Which filing, which quarter, which call?Stale data, mixed periods, last year's guidance read as current
3MathDoes the arithmetic reproduce by hand or in a sheet?Wrong margins, miscompounded growth, ratio errors, unit slips
4ConflictDoes this contradict the filing, the call, or another source?Cherry-picked lines, ignored footnotes, one-sided framing
5Human approvalDid a person sign off before this drove a decision?Silent automation, unreviewed outputs feeding into sizing or memos
Tip: Run them in sequence. A timestamp check is meaningless if the source is fabricated. Source first, always.

Mistake, Failure Mode, Workflow

This is the discipline layer. For each guardrail, name what people actually do wrong, what it costs, and the fix.

GuardrailCommon mistakeFailure modeDisciplined workflow
SourceAccept the number because it sounds rightWrong revenue base anchors the whole modelRequire a citation with a page or line reference. Open it. No link, no number.
TimestampAssume the latest data was usedQ3 metric compared against Q1 guidanceForce every figure to carry a period and filing date. Reject undated claims.
MathTrust the agent's arithmeticMargin off by 400bps flows into comps and sizingRecompute one ratio by hand. If it breaks, recheck the whole output.
ConflictRead only the summary the agent surfacedFootnote risk or one-off charge gets droppedAsk the agent to argue the opposite case from the same filing. Compare.
Human approvalLet the agent output flow straight into a decisionPosition sized on an unreviewed draftInsert a named human gate before any output drives sizing or a memo.
Why it matters: The pattern repeats across all five. The mistake is always a shortcut. The failure mode is always downstream. The workflow always reinserts judgment.

Where Each Guardrail Lives In The Workflow

Guardrails are not a final review step. They sit at specific points in the research pipeline. Mapping them keeps you from checking too late.

StageAgent taskGuardrail appliedGate owner
IngestPull 10-K, 10-Q, earnings call textSource + TimestampAgent flags, you verify
ExtractLift revenue, margins, segment dataSource + MathYou recompute one line
SynthesizeBuild comps, draft narrativeConflictYou request the counter-case
DecideFeed into IC memo or sizingHuman approvalNamed human signs
Heads up: Most people only check at the Decide stage. By then a bad source has already shaped the comps. Catch it at Ingest.

The Operating Principle

The checklist is one rule applied five times: never let the agent be the last set of eyes on anything that moves money or shapes a thesis.

  • Speed is a tool, not a verdict. A 3-second answer still earns a 3-minute check.
  • Citations are not decoration. If you cannot open the source, the number does not exist yet.
  • Recompute one figure every time. The agent that is right 95 times trains you to skip check 96.
  • Make the counter-case mandatory. An agent that only confirms your view is an echo, not research.
  • Name the human. Approval with no name attached is automation pretending to be oversight.
Key: Human judges. AI builds. The guardrails are how you keep that line from blurring under deadline pressure.

Caption

LinkedIn post copy

An AI finance agent will read a 10-K and draft your IC memo in seconds.

That speed is the trap. A fast answer feels like a checked answer. It is not.

Before I trust any agent output, I run 5 guardrails:

1. Source. Can I open the document the number came from? 2. Timestamp. As of when? Which filing, which quarter? 3. Math. Does one ratio reproduce by hand? 4. Conflict. Does this contradict the filing or the call? 5. Human approval. Did a named person sign off before it drove a decision?

Each guardrail catches a specific failure. Fabricated figures. Stale data. Wrong margins. Cherry-picked lines. Silent automation.

Most people only check at the end. By then a bad source has already shaped the comps. Catch it at ingest.

The principle is one line: never let the agent be the last set of eyes on anything that moves money or shapes a thesis.

Human judges. AI builds.

Which guardrail do you skip most? Comment it and I will show you the failure mode it hides.

Educational only. Not investment advice.

Visual design notes

  • Near-black forest-green background (#0B1410 base). ONE green accent only (#2FBF8F / WSP teal-green) for the headline keyword, table header rules, and section number badges. Everything else off-white (#E8EDE9) and muted gray (#8A968F).
  • Headline in heavy condensed uppercase (Anton or Archivo Black), left-aligned, top-left. Emphasize the word GUARDRAIL in the green accent; rest in off-white. Title spans 2 lines max.
  • Teal subtitle bar directly under the headline: full-width thin banner, dark-green fill, off-white text, single line. This frames the whole post.
  • Dense tables are the visual spine. Use thin 1px green top-border on each table header row, hairline gray row dividers, generous left padding, no vertical gridlines. Zebra-stripe rows at 4% white opacity for scan-ability.
  • Section 2 (The 5 Guardrails): render the # column as small circular green-outline badges (1-5) to make the checklist feel countable and scannable.
  • Section 4 (Where Each Guardrail Lives): add a slim left-to-right pipeline diagram strip above the table. 4 nodes Ingest, Extract, Synthesize, Decide connected by thin green arrows, each node tagged with its guardrail. Keeps it operational, not just textual.
  • Left-align all body text and headings. Section headings in semibold uppercase tracking, smaller than headline, with a short green tick-mark to the left. Notes (Why it matters, Tip, Heads up, Key) in italic muted gray with the label in green.
  • Footer locked to bottom edge: thin green hairline above a single-line footer in small-caps gray. Keep 64px outer margins on a 1080x1350 canvas. No em dashes anywhere in rendered copy.

Production checklist

  • ☐Set up the 1080x1350 WSP one-pager: near-black forest-green background, 64px margins, single green accent token, condensed headline type scale, left-aligned grid.
  • ☐Build the header: heavy condensed title with GUARDRAIL in green accent, then the full-width teal subtitle bar one-liner underneath.
  • ☐Build the 4 tables (5 Guardrails, Mistake/Failure/Workflow, Workflow Stages, plus the bulleted Operating Principle block): green header rules, hairline row dividers, zebra stripes, circular number badges on the checklist table.
  • ☐Add the 4-node pipeline diagram strip (Ingest, Extract, Synthesize, Decide) above the workflow-stage table with thin green arrows and guardrail tags.
  • ☐Add WSP branding: section tick-marks, italic green-labeled notes, bottom green hairline and footer line.
  • ☐Proofread for voice and compliance: short declarative sentences, no em dashes, no buy/sell/hold, no targets, no promised returns. Confirm Educational only. Not investment advice. sits in the caption.
  • ☐Export PNG at 1080x1350 for the LinkedIn feed and a matching PDF version for DM delivery.
Follow Dave Wang · Wall Street Prompt · For more AI investing workflows · COMMENT & REPOST

CTA

Save this checklist. Before your next agent output drives a decision, run all 5 gates. Comment the one guardrail you skip most and I will show you the failure mode it hides.

Wall Street Prompt
Wall Street Prompt — internal