The Practical Guide to XML Prompting

Why XML for prompts?

XML is particularly well suited for prompts for several reasons:

• Clear structure: you cleanly separate instructions, context, inputs and expected outputs via tags.
• Parser-friendly: the output can be read with any standard XML library and is far more robust than homemade delimiters.
• Extensible: you can add new tags without breaking older ones.
• Human-readable: developers and business teams alike can read and edit prompts.

Use XML to enforce outputs, extract information, route tasks, enforce formats, mark up few-shot examples and run self-checks. And all of that without requiring hidden step-by-step reasoning (chain-of-thought).

Core rules that save you trouble

1. Always produce well-formed XML: one root tag, every tag closed, no stray special characters in the text (except escaped or in CDATA).
2. Name tags after the domain: prefer <invoice>, <clause> or <persona> over generic <section1>.
3. Separate cleanly: instructions, examples, inputs and outputs belong in their own blocks (<instructions>, <examples>, <input>, <output_format>).
4. Define the behavior in the error case: specify an <error> schema and define when it is used.
5. Skip chain-of-thought: do not require step-by-step reasoning, but short justifications or checklists (such as <key_points> with an upper limit).
6. Use CDATA for code, JSON, regex and anything with curly or angle brackets.
7. State the output schema clearly: an <output_format> with all required tags and constraints.
8. Version your prompts: a <meta version="1.2" id="email-reply-v12"/> makes changes traceable.

<code><![CDATA[
def add(a, b): return a + b
]]></code>

A minimal, reusable skeleton

Start every new prompt with this skeleton. It contains metadata, task, instructions, input, output format and an error handler.

<prompt>
  <meta version="1.0" id="GEN-001"/>
  <task>Fasse einen langen Artikel für eine vielbeschäftigte Führungskraft zusammen.</task>

  <instructions>
    - Schreibe in einfacher Sprache.
    - Halte dich an 120 bis 150 Woerter.
    - Ergaenze 3 stichpunktartige Highlights.
  </instructions>

  <input>
    <article><![CDATA[
      {{ARTIKEL HIER EINFUEGEN}}
    ]]></article>
  </input>

  <output_format>
    <summary>120 bis 150 Woerter.</summary>
    <highlights>
      <bullet>Praegnante Aussage</bullet>
      <bullet>Praegnante Aussage</bullet>
      <bullet>Praegnante Aussage</bullet>
    </highlights>
  </output_format>

  <error_handler>
    Wenn die Eingabe fehlt oder nicht auswertbar ist, gib zurueck:
    <error code="INPUT_INVALID" message="Grund hier"/>
  </error_handler>
</prompt>

Prompt library: ready-to-use templates

All templates deliberately avoid hidden chain-of-thought. Adopt them as-is or adapt the tag names to your domain.

1) Structured extraction

<prompt>
  <task>Extrahiere Firma, Produkt, Preis und Stimmung aus einer Rezension.</task>

  <input>
    <review><![CDATA[
      Ich liebe den Acme Air 2000, nur 249 Dollar und leiser als mein alter Ventilator.
    ]]></review>
  </input>

  <output_format>
    <extraction>
      <company>string</company>
      <product>string</product>
      <price currency="USD">number</price>
      <sentiment one_of="positive|neutral|negative">string</sentiment>
      <key_points max="3">stichpunktartige Phrasen</key_points>
    </extraction>
  </output_format>

  <instructions>
    - Ist ein Feld unbekannt, setze ein leeres Tag (z. B. <price/>).
    - Halte <key_points> auf hoechstens 3 kurze Stichpunkte.
  </instructions>
</prompt>

2) Classification with fixed labels

<prompt>
  <task>Klassifiziere die Absicht eines Support-Tickets.</task>
  <labels>
    <label id="bug">Fehlermeldung</label>
    <label id="billing">Abrechnung</label>
    <label id="feature">Feature-Wunsch</label>
    <label id="other">Sonstiges</label>
  </labels>
  <input>
    <ticket>Meine Karte wurde diesen Monat zweimal belastet.</ticket>
  </input>
  <output_format>
    <classification>
      <label_id one_of="bug|billing|feature|other">string</label_id>
      <justification max_chars="180">Ein bis zwei Saetze.</justification>
    </classification>
  </output_format>
</prompt>

3) Email reply (professional and safe)

<prompt>
  <persona>
    <role>Customer Success Manager</role>
    <tone>Professionell, hilfsbereit, knapp</tone>
    <signature>Viele Gruesse, Alex (Customer Success)</signature>
  </persona>

  <task>Entwirf eine Antwort, die bestaetigt, beantwortet und naechste Schritte vorschlaegt.</task>

  <input>
    <email_from_customer><![CDATA[
      Hallo, ich komme nach dem Zuruecksetzen meines Passworts nicht mehr ins Dashboard.
    ]]></email_from_customer>
  </input>

  <output_format>
    <email>
      <subject>string</subject>
      <body>Klartext, kurze Absaetze</body>
      <next_steps>
        <step>Stichpunkt</step>
      </next_steps>
    </email>
  </output_format>
</prompt>

4) SQL generation with schema context

<prompt>
  <task>Erzeuge eine SQL-Abfrage aus einer natuerlichsprachlichen Frage und einem Schema.</task>
  <schema><![CDATA[
  tables:
    orders(order_id, customer_id, total, created_at)
    customers(customer_id, region)
  ]]></schema>
  <input>
    <question>Gesamtumsatz pro Region fuer das laufende Jahr 2024.</question>
  </input>
  <output_format>
    <sql><![CDATA[
      SELECT ...
    ]]></sql>
    <assumptions max="3">
      <item>kurzer Satz</item>
    </assumptions>
    <safety_checks>
      <check>Sind die Datumsfilter korrekt?</check>
      <check>Verbinden die Joins die richtigen Schluessel?</check>
    </safety_checks>
  </output_format>
</prompt>

5) Meeting notes to action items

<prompt>
  <task>Wandle ein Meeting-Transkript in strukturierte Notizen um.</task>
  <input>
    <transcript><![CDATA[(Transkript einfuegen)]]></transcript>
  </input>
  <output_format>
    <notes>
      <topics>
        <topic>...</topic>
      </topics>
      <decisions>
        <decision>...</decision>
      </decisions>
      <action_items>
        <item owner="Name" due="YYYY-MM-DD">...</item>
      </action_items>
    </notes>
  </output_format>
</prompt>

6) Intent routing (switch)

<prompt>
  <task>Leite die Nutzereingabe an den richtigen Handler weiter.</task>
  <handlers>
    <handler id="KB_SEARCH">Wissensdatenbank-Suche</handler>
    <handler id="HUMAN_HANDOFF">Uebergabe an Mensch</handler>
    <handler id="ORDER_STATUS">Bestellstatus</handler>
  </handlers>
  <input>
    <utterance>Wo ist meine Bestellung Nr. 1234?</utterance>
  </input>
  <output_format>
    <routing>
      <handler_id one_of="KB_SEARCH|HUMAN_HANDOFF|ORDER_STATUS">ORDER_STATUS</handler_id>
      <entities>
        <order_id>1234</order_id>
      </entities>
      <confidence>0-1</confidence>
    </routing>
  </output_format>
</prompt>

7) Critique and improve (self-check)

<prompt>
  <task>Kritisiere und verbessere den Entwurf, begrenze die Begruendung.</task>
  <constraints>
    <rationale_limit>120 Zeichen</rationale_limit>
  </constraints>
  <input>
    <draft><![CDATA[(Entwurf einfuegen)]]></draft>
  </input>
  <output_format>
    <improvement>
      <revised><![CDATA[(ueberarbeiteter Text)]]></revised>
      <rationale>max. 120 Zeichen</rationale>
      <checks>
        <check id="clarity">pass|fail</check>
        <check id="tone">pass|fail</check>
        <check id="length">pass|fail</check>
      </checks>
    </improvement>
  </output_format>
</prompt>

Parsing and validation

With the Python standard library you can read the XML output in just a few lines:

import xml.etree.ElementTree as ET

def parse_extraction(xml_text: str):
    root = ET.fromstring(xml_text)
    ext = root.find('extraction')
    return {
        "company": (ext.findtext('company') or '').strip(),
        "product": (ext.findtext('product') or '').strip(),
        "price": (ext.findtext('price') or '').strip(),
        "sentiment": (ext.findtext('sentiment') or '').strip(),
        "key_points": [e.text.strip() for e in ext.findall('key_points/bullet') if e.text]
    }

Validation checklist:

• The root tag matches the expectation.
• All required children exist (not empty, unless empty is permitted).
• Allowed enumerations are respected (one_of attributes).
• No unexpected tags (optional, but useful).

Robustness and anti-patterns

A few proven techniques make your prompts considerably more stable:

FAQ

Why XML instead of Markdown headings? XML is machine-parseable and consistent. Markdown is good for humans, but ambiguous for machines.

Can I mix XML with Markdown? Yes. Keep the machine-readable parts in XML, human notes can sit as bullet points inside <instructions>.

What about security and compliance? Use <checks> to enforce rules such as no personal data or no medical or legal advice. On a violation, output an <error>.

How do I avoid rambling justifications? Limit them via attributes such as max_chars or max and prefer checklists over prose.

Three principles to take away

• Short, specific tag names beat clever ones.
• Show one complete example per prompt so the model imitates the format.
• Prefer constraints over prose (150 words maximum, 3 bullet points, fixed enumerations).