Shrutam CCA-F Crash › Domain 4 › d4-l08-negative-constraints Hinglish English →

Negative Constraints — Why 'Don't' Fails

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 liye ek Claude-powered loan-summary microservice design kar rahe hain — output must be strict JSON, no extra fields, no prose leakage. Aapne prompt mein likha: "Do not include any explanation, do not add commentary, do not output plain text." Production mein gaya toh Claude ne fir bhi ek polite preamble inject kar diya — "Certainly! Here is your JSON:" — aur downstream schema-validation pipeline crash ho gayi. Aaj samajhte hain ki negative constraints kyun fail hote hain, aur unhe positive output-schema enforcement mein kaise convert karte hain.

Key Takeaways

LLMs process negations weakly — 'Do not output X' activates the concept of X in the generation path, making suppression unreliable without hard structural enforcement.
Replace every negative constraint with a positive output specification: define the exact schema, field list, and data types using JSON Schema or a response_format parameter so the model has no ambiguity to resolve.
In production pipelines, treat prompt-level negative constraints the same way you treat application-layer input validation — necessary but insufficient; always back them with a schema-validation guardrail at the API boundary.
Memory anchor — schema-validation-guardrails: if your constraint lives only in natural language and not in a machine-readable schema, it is a lint warning, not a compile-time error.

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