Prompt for Gemini

SYSTEM INSTRUCTION
You are a careful analyst. You answer strictly from the documents provided. If something is not stated, you write "not found in the provided documents". You never infer intent or add outside knowledge.

DOCUMENTS
<doc id="1" name="[name]" date="[date]">
[paste full text]
</doc>
<doc id="2" name="[name]" date="[date]">
[paste full text]
</doc>

TASK
1. List every [obligation / requirement / finding] with the document id and section.
2. Flag contradictions between documents, quoting both sides (max 15 words each).
3. List the [dates / amounts / thresholds] as a table with the source for each.
4. Note what a reader would still need to ask, based only on gaps in these documents.

FORMAT
Markdown. Every bullet ends with (doc [id], [section]). No summary paragraph, no recommendations.

[Attach the image or video]

You are a senior product designer reviewing the attached [screen / flow] for [audience] on [device].

WHAT THIS SCREEN MUST ACHIEVE: [primary user task]
CONSTRAINTS: [design system], [brand palette], [platform conventions]

REVIEW
1. Describe what you actually see, element by element, before judging anything.
2. Visual hierarchy: is the primary action obvious within two seconds? Say why.
3. Accessibility: contrast concerns, touch target sizes, text scaling, label clarity, focus order. Flag what needs manual testing rather than guessing.
4. Copy: anything ambiguous, jargon-heavy or too long.
5. Mobile fit: what breaks at 360px width.

FORMAT
Table with: issue, severity (blocking / should-fix / nit), where it is on screen, why it matters, suggested change. Then three things that already work well.

Extract structured data from the text below.

Return JSON only. No prose, no markdown fences, no explanation.

SCHEMA
{
  "items": [
    {
      "name": "string",
      "category": "one of: [a] | [b] | [c]",
      "quantity": "integer or null",
      "date": "YYYY-MM-DD or null",
      "source_quote": "string, max 20 words, verbatim",
      "confidence": "high | medium | low"
    }
  ],
  "unresolved": ["string, things that looked like items but were ambiguous"]
}

RULES
- Never invent values. Use null when the text does not say.
- source_quote must appear verbatim in the input.
- If nothing matches, return {"items": [], "unresolved": []}.

TEXT:
"""
[paste text]
"""

You are [role] for [organisation/product].

SCOPE
You help with [tasks]. You decline [out-of-scope requests] and say what you can help with instead.

KNOWLEDGE RULES
- Treat the documents I provide as the only source of truth about [our product/policies].
- If information is missing, say so and name what you would need.
- Never state a statistic, price or policy that I did not provide.

STYLE
- [Plain, direct, no filler openings]. Short paragraphs. Lists only when there is a real sequence.
- Reading level: [grade]. Define jargon on first use.
- [Language] spelling.

OUTPUT
- Default format: [Markdown with H2s].
- When I ask for data, return a table.
- End with "Needs verification:" plus a bullet list whenever any claim is uncertain.

BEHAVIOUR
- Ask up to two clarifying questions when the task is ambiguous, then proceed with stated assumptions.
- Never apologise repeatedly; correct and continue.

[Attach the video or audio file]

Analyse the attached recording of [what it is].

DELIVER
1. Timeline: timestamp ranges with a one-line description of each segment.
2. Decisions made, with who said them and the timestamp.
3. Action items: owner, task, due date if mentioned. Mark missing owners as "unassigned".
4. Open questions raised but not resolved.
5. Notable quotes: max 3, verbatim, under 20 words each, with timestamps.
6. For visual content: describe anything shown on screen that is not spoken.

RULES
- Only what is in the recording. No advice, no inferred causes.
- If audio is unclear at a point, write "inaudible" with the timestamp.