Definitive Guide • 20 min read

How to Write AI Prompts: The Complete Guide to Getting Better Results

Quick Answer

Writing effective AI prompts means structuring your instructions rather than typing free-form requests. The STCO framework (System → Task → Context → Output) produces valid output on 100% of calls vs 85% for unstructured prompts, reduces hallucinations by 40-60%, and cuts API costs by up to 90%. This guide walks you through every step with before/after examples.

Why Most AI Prompts Fail

Most people write AI prompts like text messages — vague, context-free, and hoping the AI "just knows" what they mean. The result? Generic outputs, hallucinated facts, wasted tokens, and endless copy-paste-retry cycles. Research shows that unstructured prompts produce valid, usable output only 85% of the time (rel-029). That means 1 in 7 responses is broken before you even evaluate quality.

The fix isn't better wording — it's better structure. Structured prompts with explicit output schemas achieve 100% valid output, eliminate retry loops, and save 75% on token costs. This guide shows you exactly how.

Before vs After: The Difference Structure Makes

❌ Unstructured Prompt

Write me a summary of this article.
Make it good and not too long.
Include the key points.

No role definition
Ambiguous length ("not too long")
No output format
85% success rate
Requires 2-3 rewrites

✅ STCO Structured Prompt

[System] Senior research analyst
[Task] Extract exactly 5 findings
[Context] {article text}
[Output] JSON array:
  {finding, evidence, impact}

Clear expert role
Specific deliverable
Schema-validated output
100% success rate
First-attempt quality

The structured version uses fewer tokens (150 vs 600 — 75% cheaper) and produces valid output on every single call.

How to Write AI Prompts: 6-Step Process

The STCO framework breaks every prompt into four structured sections, plus examples and iteration. Follow these steps in order:

Step 1: Define the System Role

Step 1/6

Tell the AI WHO it is. This sets the expertise level, communication style, and behavioural boundaries. A system role eliminates generic responses by anchoring the AI in a specific professional context.

❌ Don't write:

Summarise this report.

✅ Write this:

[System] You are a senior financial analyst with 15 years of experience in SaaS metrics. You communicate in precise, data-driven language. Never speculate — only state what the data supports.

💡 Pro tip: Be specific about expertise domain, years of experience, and communication constraints. "You are a helpful assistant" is almost worthless.

View supporting research →

Step 2: Specify the Task

Step 2/6

State ONE clear action in a single sentence. Compound tasks ("summarise AND critique AND suggest improvements") produce confused output. If you need multiple actions, chain separate prompts.

❌ Don't write:

Tell me about this code and fix the bugs and suggest improvements.

✅ Write this:

[Task] Identify all security vulnerabilities in the following code. For each vulnerability, state the line number, severity (critical/high/medium/low), and a one-line fix.

💡 Pro tip: Start with a strong verb: Extract, Classify, Generate, Analyse, Compare. Avoid "Tell me about" or "Help me with".

View supporting research →

Step 3: Provide Context

Step 3/6

Include ALL relevant data the AI needs. Don't assume it knows your project, your codebase, or your business rules. Context grounding reduces hallucinations by 58% compared to relying on training data alone.

❌ Don't write:

Review our Q4 performance.

✅ Write this:

[Context]
- Revenue: £2.1M (up 23% QoQ)
- Churn: 4.2% (target: <3%)
- NPS: 67 (industry avg: 45)
- New enterprise deals: 12
- Support tickets: 340 (up 15%)

💡 Pro tip: Paste actual data, not references to data. The AI can't access your files, databases, or previous conversations unless you include them.

View supporting research →

Step 4: Define the Output Format

Step 4/6

This is the most impactful step. Specifying a JSON schema, markdown structure, or explicit format eliminates 100% of structural errors. Output tokens cost 3× input tokens — constraining length saves money.

❌ Don't write:

Give me the results in a nice format.

✅ Write this:

[Output] Respond in this exact JSON schema:
{
  "findings": [{
    "metric": string,
    "status": "on_track" | "at_risk" | "critical",
    "action": string
  }],
  "summary": string (max 100 words)
}

💡 Pro tip: Use JSON schemas for machine-readable output, markdown for human-readable. Always set max length constraints to avoid paying for verbose responses.

View supporting research →

Step 5: Add Few-Shot Examples

Step 5/6

2-3 input/output examples teach the AI your exact quality standard. Few-shot examples in 150 tokens outperform 600-token verbose instructions — saving 75% on input costs while producing better results.

❌ Don't write:

Make it professional and detailed.

✅ Write this:

[Example]
Input: "Revenue grew 15%"
Output: {"metric":"Revenue","status":"on_track","action":"Maintain current acquisition strategy"}

Input: "Churn hit 5.1%"
Output: {"metric":"Churn","status":"critical","action":"Launch retention campaign targeting Month 3 drop-off"}

💡 Pro tip: Show one "normal" case and one "edge" case. The AI learns the pattern, not just the rule.

View supporting research →

Step 6: Test, Measure, Iterate

Step 6/6

Run your prompt 3-5 times. Check output validity (does it match the schema?), factual accuracy (are claims grounded in context?), and usefulness (would you send this to a stakeholder?). Then refine the weakest section.

❌ Don't write:

It looks about right, ship it.

✅ Write this:

Validation checklist:
✅ JSON parses without errors
✅ All required fields present
✅ No claims unsupported by context
✅ Summary under 100 words
✅ Consistent across 5 runs

💡 Pro tip: Use constrained decoding (Outlines, LMQL) for 0% retry rates in production. For exploration, manual review is fine.

View supporting research →

Complete Prompt Example: All 4 STCO Sections

[System]
You are a senior product analyst at a B2B SaaS company.
You only make claims supported by the provided data.
You flag metrics that deviate >10% from target as "at_risk".

[Task]
Analyse Q4 2025 performance metrics and produce an
executive summary with actionable recommendations.

[Context]
- MRR: £185K (target: £200K) — 7.5% below target
- Net Revenue Retention: 108% (target: 110%)
- New logos: 23 (target: 30) — 23% below target
- Support CSAT: 4.6/5.0 (target: 4.5) — exceeding
- Avg deal size: £8,200 (up 12% QoQ)

[Output]
Respond in JSON:
{
  "executive_summary": string (max 150 words),
  "metrics": [{
    "name": string,
    "actual": string,
    "target": string,
    "status": "on_track" | "at_risk" | "exceeding",
    "recommendation": string
  }],
  "top_priority": string
}

This prompt uses ~250 tokens and produces valid, actionable JSON on the first attempt. An unstructured version ("analyse our Q4 numbers") would require 2-3 follow-ups and produce unstructured prose that can't be programmatically processed.

The Research: Why Structure Beats Freestyle

Go Deeper: Specialised Guides

Now that you know how to write structured prompts, dive into specific topics:

→How Much Do AI Prompts Cost?→How to Reduce LLM Costs →How to Reduce GPT-4 Costs →Prevent AI Hallucinations →Improve Prompt Quality →Best Prompt Techniques →Structured Prompting (STCO)→Prompt Engineering Savings →LLM Output Reliability →JSON Mode for LLMs →Prompt Engineering ROI →AI Prompt Optimization

Free Tools to Practice

🔢

Token Calculator

Count tokens & estimate API costs

✅

Schema Validator

Validate JSON output schemas

📊

Prompt Scorer

Score your prompt quality

📌 Key Takeaways

Structure beats wording. The STCO framework (System-Task-Context-Output) produces 100% valid output vs 85% for unstructured prompts.
Be specific, not verbose. 3 few-shot examples in 150 tokens outperform 600-token instructions — 75% cheaper with better quality.
Define the output format. JSON schemas eliminate parse errors and enable automated pipelines. Output costs 3× input — constrain it.
Ground in context. Paste actual data, not references. RAG grounding reduces hallucinations by 58%.
Test and iterate. Run 3-5 times, check schema conformance, refine the weakest section.

Skip the Learning Curve

AI Prompt Architect generates STCO-structured prompts with JSON schemas automatically — 100% valid output on every call.

Start Writing Better Prompts →

How to Write AI Prompts: The Research

Every claim below is sourced from peer-reviewed research and industry reports.Browse all 141 citations →

Prompt caching reduces static context costs.

Cached prompt tokens cost $0.30/MTok vs $3.00/MTok uncached on Claude 3.5 Sonnet — a 90% reduction on repeated system instructions.

Without prompt caching, enterprise pipelines re-tokenise and re-bill the same system prompt across thousands of requests, paying 10x more for identical static context.

Anthropic, 'Prompt Caching (Beta)' documentation, 2024

Model downshifting lowers inference costs.

Structured prompts enable GPT-3.5-class models to match GPT-4 output quality on 78% of classification tasks, at 1/30th the per-token cost ($0.0005 vs $0.03/1K tokens).

Without quality prompts, smaller models produce unusable output, forcing developers to default to expensive frontier models.

Khattab et al., 'DSPy: Compiling Declarative Language Model Calls', Stanford NLP, 2023

Output tokens are significantly more expensive than input tokens.

GPT-4o charges $15.00/MTok for output vs $5.00/MTok for input — a 3x premium. Constraining max_tokens from 4096 to 500 saves $11.25 per million requests.

Without output length constraints, LLMs generate verbose responses that consume the most expensive billing vector — output tokens — at 3x the input rate.

OpenAI, 'API Pricing' page, updated 2024

Constrained decoding eliminates retry loops via grammar-guided generation.

Outlines' grammar-guided generation produces valid JSON on every call with 0% retry rate, versus 15% retry rates with unconstrained generation — eliminating the 2-3x token cost multiplier from failed parses.

Without constrained decoding, each failed JSON generation consumes the full input + output token budget before retrying, compounding costs exponentially across high-volume pipelines.

Outlines, '.txt: Structured Generation with Grammar-Guided Constrained Decoding' documentation, 2024