Shrutam CCA-F Crash › Domain 4 › d4-l04-examples-before-task Hinglish English →

Few-Shot Examples — Placement and Count

Domain 4 · 18% ~12 min Hinglish narration

Audio-only (commute / mobile data)

Same Saavi narration, smaller file. Opus 48k preferred — auto-selected by your browser.

Scenario anchor

Aap Aaranya IT BFSI ke ek senior architect hain — client hai ASX-India, aur requirement hai ki Claude ek high-volume trade-reconciliation pipeline ke liye structured JSON output generate kare, har response mein exact field names, nested objects, aur ISO-8601 timestamps ke saath. Aapne system prompt mein JSON schema diya, task instruction diya — lekin output inconsistent aa raha hai, field names kabhi camelCase hain, kabhi snake_case. Problem kahan hai? Few-shot examples ka placement aur count galat hai — aur yahi lesson us gap ko close karega.

Key Takeaways

Always place few-shot examples before the task instruction — they function as a compiled schema contract that conditions all subsequent generation in that context window.
Calibrate shot count to output complexity: 1-2 shots for flat JSON, 3-5 shots for nested or polymorphic structures; beyond 5, marginal fidelity gain rarely justifies the context-window cost.
Each example must be internally consistent — field names, data types, and nesting depth must exactly match your target schema, because Claude generalises from demonstrated patterns, not from stated rules.
Schema-validation-pipeline rule: treat your few-shot block as the reference payload in a schema-validation pipeline — if the example itself would fail your own validator, your production output will too.

← Previous lesson Same lesson in English Next lesson → Back to overview