Without format instructions, the AI guesses how to structure output — and its guess is almost never optimized for your workflow.
Format control means specifying: length, structure, schema, delimiters, and what to exclude. One of the easiest quality improvements at zero extra cost.
02 Weak vs. Strong
EX 01Parsing applicant emails for a database
Parse this email into structured data for our ATS. Return ONLY valid JSON — no markdown backticks, no explanation, no text before or after.
Schema (null for missing fields):
{
"name": string,
"email": string,
"role_applied": string,
"years_experience": number,
"current_company": string | null,
"salary_expectation_usd": number | null,
"earliest_start_date": string | null
}
Parse "$165-185k" as the midpoint (175000). Parse "available in 4 weeks" as null — can't compute absolute dates.
Email:
---
Hi, I'm Priya Mehta, reaching out about the Senior Product Designer role. I have 8 years of product design experience, currently at Figma. Looking for $165-185k. Could start in about 4 weeks. LinkedIn: linkedin.com/in/priyamehta
---
→ Why it works
Schema with types and null handling. Parsing rules for salary and relative dates.
EX 02Vendor comparison table
Compare AWS, GCP, and DigitalOcean for a production Go API (~50k requests/day, managed PostgreSQL, Redis cache).
Return a markdown table ONLY. No intro, no conclusion, no explanation.
Columns: Criterion | AWS | GCP | DigitalOcean
Rows (include all, in this order):
1. Managed Kubernetes product name
2. Managed PostgreSQL price at 8GB RAM
3. Managed Redis price at 2GB
4. Min viable setup cost/month for our use case
5. Time to first deploy (team new to the platform)
6. Best feature for Go backends (be specific)
7. Biggest operational pain point (be honest)
8. Best suited for (one sentence)
→ Why it works
Table schema with exact rows defined. 'No intro, no conclusion' prevents filler that pushes the table below the fold.
03 Key Points
01Specify length in measurable units: sentences, words, bullets, characters
02'Return only valid JSON, no markdown wrapping' — critical for code integration
03'No preamble. Start with the output.' — eliminates filler
04For comparisons: always request a table — scannable in 10 seconds
05Specify what NOT to include: 'no markdown', 'no bullets', 'no intro paragraph'
04 Model-Specific Notes
Claude follows XML output format tags reliably: <output_format>Return only JSON matching schema</output_format>. Prefilling the assistant turn with '{' forces JSON mode.
05 For Your Role
Tell it the shape: 'Just the answer in one sentence', 'A numbered list of 5', '3 short paragraphs, no headings'. The AI will comply.