Core Techniques

Prompts are code — versioned, tested, engineered. Six techniques that separate 'it sometimes works' from 'it ships.'

▶ Watch this reel

What you'll learn

Clear instructions & role
Few-shot examples
Chain-of-thought
XML / delimiters
Output format control
Negative instructions

Remember this

Role + task (verbs) + format + constraints — specificity is the #1 quality lever
Few-shot examples beat descriptions; chain-of-thought for multi-step problems
Tag data with XML; spec outputs field-by-field; phrase constraints as positive behavior

Clear instructions & role

Structure: role → task (verbs) → format → constraints.
Vague prompts leave degrees of freedom that become surprises.

Few-shot

2–5 input→output examples teach tone, format, boundaries.
Order simple → tricky; keep examples perfectly consistent.
Almost-right output? Add the almost-wrong case done correctly.

Chain-of-thought

'Think step by step' / scaffolded steps for multi-step problems.
Costs tokens + latency; reasoning can still be wrong → verify finals.

XML / delimiters

Wrap external content: <document>, <transcript>, <user_input untrusted="true">.
Semantic tag names steer attention; tagged data blurs less into instructions.

Output format control

Field-by-field specs · length in real units (words/items, not 'be concise').
Feeding a machine? Use structured output schema (GA-04) to enforce.

Negative instructions

'Don't X' mentions X. Pair prohibitions with positive, concrete replacements — placed last.

Code: A production-grade prompt template

PROMPT = """You are a senior support triage engineer.

Classify each ticket by urgency.

<rules>
- P1: money lost or service down
- P2: degraded but usable
- P3: feature requests and questions
</rules>

<examples>
Q: app crashes on login
A: {"urgency": "P1", "reason": "service unusable"}

Q: report export is slow
A: {"urgency": "P2", "reason": "degraded, workaround exists"}
</examples>

Answer with JSON only: {"urgency": "...", "reason": "one sentence"}.
Reason in at most 2 short steps, then the JSON.

<ticket untrusted="true">
{ticket}
</ticket>
"""

# Usage — THIS is a prompt template (GA-06 versions it):
ticket = "charged twice for one order"
messages = [{"role": "system", "content": PROMPT.replace("{ticket}", ticket)}]