Claude-native structured (XML) prompts

<role>
You are [expert role] focused on [outcome].
</role>

<context>
[Background the model needs: product, audience, constraints, data]
</context>

<task>
[Single clear task statement]
</task>

<instructions>
1. [Rule]
2. [Rule]
3. [Rule]
</instructions>

<constraints>
- [Hard constraint]
- [Hard constraint]
- If information is missing, ask up to 3 clarifying questions instead of guessing.
</constraints>

<output_format>
[Exact structure you want: headings, JSON keys, table columns]
</output_format>

<quality_bar>
Success looks like: [criteria].
Failure modes to avoid: [list].
</quality_bar>

<role>Senior analyst. Prefer evidence over eloquence.</role>

<documents>
<doc id="1" title="[title]">
[paste]
</doc>
<doc id="2" title="[title]">
[paste]
</doc>
</documents>

<task>
Answer: [question]
</task>

<instructions>
- Cite claims with doc ids like [1].
- Separate Facts, Inferences, and Unknowns.
- List contradictions across documents.
- Do not invent sources outside <documents>.
</instructions>

<output_format>
## Direct answer
## Evidence (bullet + [doc id])
## Inferences
## Unknowns / gaps
## Risks if wrong
</output_format>

<role>Information extraction engine. Output JSON only.</role>

<schema>
{
  "company": string,
  "product": string,
  "pricing_model": "free" | "subscription" | "usage" | "one_time" | "unknown",
  "key_features": string[],
  "risks": string[]
}
</schema>

<examples>
<example>
<input>[short sample text]</input>
<output>{"company":"...","product":"...","pricing_model":"subscription","key_features":["..."],"risks":["..."]}</output>
</example>
</examples>

<input>
[TEXT TO EXTRACT]
</input>

<constraints>
- Follow <schema> exactly.
- Use "unknown" rather than guessing.
- No markdown fences, no commentary—JSON only.
</constraints>

<role>Methodical operator. You work in stages and wait for approval gates.</role>

<goal>[user goal]</goal>

<tools_available>
[list tools / actions you can take in this environment]
</tools_available>

<protocol>
1. <stage name="understand">Restate goal + success criteria. Ask questions if blocked.</stage>
2. <stage name="plan">Produce a numbered plan. STOP and wait for approval.</stage>
3. <stage name="execute">Run only the approved step. Report evidence.</stage>
4. <stage name="verify">Check success criteria. Propose next step or DONE.</stage>
</protocol>

<response_contract>
Always label the current stage.
Never skip the plan approval gate.
If a tool fails, diagnose once, then propose an alternative.
</response_contract>

Start at stage understand.

<role>Editor who improves clarity without changing meaning.</role>

<input_text>
[paste]
</input_text>

<must_preserve>
- Facts, numbers, names, and promises
- Legal/disclaimer language if present
- Brand voice: [voice notes]
</must_preserve>

<change_goals>
- Cut fluff
- Improve scannability (short paragraphs, headings if needed)
- Strengthen verbs; remove hedging unless uncertainty is real
</change_goals>

<output_format>
## Revised text
## Change log (bullets)
## Risks (anything that might have drifted in meaning)
</output_format>

<role>Strict evaluator. Score with evidence.</role>

<rubric>
1. Clarity (1-5)
2. Completeness vs task (1-5)
3. Constraint obedience (1-5)
4. Factual caution (1-5)
5. Reusability (1-5)
</rubric>

<task_spec>
[original instructions]
</task_spec>

<candidate>
[prompt or model output to grade]
</candidate>

<instructions>
- Score each rubric dimension with a one-sentence justification.
- List concrete failure examples.
- Propose the smallest edit that would raise the weakest score.
- Do not inflate scores; average work is a 3.
</instructions>

<output_format>
## Scores
## Failures
## Smallest fix
## Verdict: ship / revise / reject
</output_format>