Why this matters

T Today reads brokerage screenshots and returns a 做T plan — buy zones, sell targets, cash notes. A confident wrong ticker is worse than “I can’t read this.” Users will act on PLTR advice when they meant TSLL.

Vision models want to be helpful. Without explicit rules they fill gaps: blur → guess, old context → reuse HODU from yesterday’s lab save. Prompt design here is product safety, not polish.

Related: Prompt vs Context vs Harness engineering · Two-layer decision engine.

The three behavior rules

We state these plainly in the system prompt and enforce them in code around the call:

“Helpful” hallucination is a bug in a finance coach, even a paper one.

Layer 1: System prompt (behavior contract)

All screenshot honesty rules live in buildAdvisorySystemPrompt() — invest-t-advisor/src/lib/risk-lab/advisory-context.ts.

The critical block (paraphrased; see repo for exact string):

NEVER guess or infer a ticker symbol.
If you cannot read a symbol with full confidence from the image:
  → extractedPortfolio = null
  → tDecision = null
  → summary = honest message asking for clearer image or text (e.g. "TSLL 200 shares, $5000 cash")

If user sends a screenshot, read symbols into extractedPortfolio only when clearly visible.
Brokerage screenshots: ticker, "N shares", day % — screenshot is source of truth.
tDecision.symbol must come from screenshot or typed holdings, not old lab data.

Single-stock chart: read ticker from chart title only — never guess from price or prior context.
If title unreadable, admit it and ask for a clearer image.

Why system, not user message? User presets change by flow (morning_t, t_analysis, overnight). Honesty rules must survive every path. System prompt = invariant behavior.

Why JSON fields, not prose? The UI renders structured cards. “Sorry, can’t read” must map to extractedPortfolio: null so we don’t show a fake holdings table. The schema is part of the prompt:

"extractedPortfolio": {"cash": number|null, "holdings": [...]} | null
"tDecision": { "symbol": string|null, ... } | null

Null is the machine-readable “I refused to guess.”

Layer 2: User message (task, not ethics)

buildAnalysisUserMessage() in advisory-presets.ts adds session time, screenshot type A/B (portfolio list vs single chart), and “tDecision REQUIRED when holdings are known.”

That REQUIRED line applies when input is readable — not when honesty rules fire. We rely on system prompt priority: unclear image → null outputs beat “always fill tDecision.”

When we tighten further, we’ll add one explicit line to presets:

If ticker unreadable per system honesty rules, skip tDecision — do not invent.

Layer 3: Context assembly (don’t poison vision)

Prompt text saying “screenshot is truth” fails if the same request includes HODU 200 sh · last $41 from yesterday’s lab.

Fix in three places:

flowchart TD
  IMG{Screenshot attached?}
  IMG -- yes --> NO_LAB[Do not send lab holdings to model]
  IMG -- yes --> UI[Hide / auto-off Include Lab checkbox]
  IMG -- yes --> RUN["run.ts: useLabStateForAi = false"]
  IMG -- no --> LAB[Optional lab snapshot for text-only quotes]
  NO_LAB --> PROMPT["advisory-llm.ts: if (!input.image) { ... lab blocks }"]
  PROMPT --> USER["User line: screenshot is sole source of holdings"]

1. advisory-llm.ts — lab snapshot and quote lines only when !input.image
2. run.ts — useLabStateForAi = input.image ? false : …
3. UI — picking a screenshot auto-unchecks “Include saved lab portfolio” and hides the checkbox

Prompt + context hygiene together. Either alone leaked old tickers in production.

Layer 4: Model settings (harness, not prompt)

Same file, runRiskAdvisoryChat():

Rules in the prompt; validation in the harness. See harness engineering post.

Example outputs

Clear Robinhood-style list → extractedPortfolio.holdings: [{symbol:"TSLL", shares:200}, …], full tDecision, normal plan card.

Blurry crop / unreadable title →

{
  "v": 1,
  "en": {
    "summary": "I can't read the ticker symbols clearly from this image. Please upload a sharper screenshot or type your holdings (e.g. TSLL 200 shares, $5000 cash).",
    "extractedPortfolio": null,
    "tDecision": null,
    ...
  },
  "zh": { "...": "无法清晰识别截图中的股票代码。请上传更清晰的截图，或用文字描述持仓（例如：TSLL 200股，现金 $5000）。" }
}

Typed fallback — user message $12k cash, PLTR 200 shares → parse text into extractedPortfolio; screenshot optional.

Checklist for similar products

1. Write the refusal path in the schema — null objects, not empty strings that look like success
2. Put “never guess” in system prompt — early, imperative, with concrete fallback copy
3. Strip conflicting context — stale DB rows + new image = confused model
4. Align UI with prompt — don’t show “include old data” when image is attached
5. Validate JSON after the call — prompt compliance is probabilistic; parsers are deterministic
6. Bilingual refusal copy — same facts in zh and en leaves; user locale picks display

Takeaway

Good vision prompts aren’t “describe this image.” They’re decision trees in prose: read clearly → analyze; read poorly → refuse with a path forward; never guess — and the harness (context, schema, UI) must match.

Code: advisory-context.ts, advisory-llm.ts, advisory-presets.ts.