What's your agent input pipeline?

[JSONFIRST]

Most agent failures happen at input — not at the model level. The model is fine; the problem is it receives ambiguous, under-specified, or contradictory instructions.

Our stack: raw user text → JSONFIRST (structured JDON) → confidence gate → agent execution.

What does your input pipeline look like? Do you do any normalization or validation before the agent acts?

[Nyrok]

Completely agree that most failures are input failures. The model does not hallucinate because it is broken. It hallucinates because the input was ambiguous about what it was supposed to do.

My input pipeline for prompt construction works like this:

raw intent -> typed semantic blocks -> compiled prompt -> model

The key step is the decomposition into typed blocks. Instead of sending one blob of text, I break the input into labeled sections: role, objective, context, constraints, examples, output format, etc. Each block type has its own validation rules. Then a compiler assembles them in the right order for the target model (Claude gets XML tags, OpenAI gets markdown headers).

This catches a lot of the "under-specified or contradictory instructions" problem at the pipeline level. If a user skips constraints, you know. If the role contradicts the objective, you can flag it before execution. You are essentially type-checking the prompt the same way you would type-check function arguments.

Your JSONFIRST approach of parsing to structured JDON before execution is the same instinct applied one level deeper. The structured intermediate representation is what makes everything downstream predictable.

I built this as an open-source tool: flompt (visual prompt builder, 13 block types, compiles to model-optimized output). The decompose step can also work in reverse: paste a freeform prompt and it breaks it into typed blocks so you can see what is missing.

Source: https://github.com/Nyrok/flompt