Kiro Intent Classifier

Overview (confidence: high)

Kiro’s Mode Classifier is the smallest prompt in the corpus at approximately 500 tokens. It uses a compact JSON envelope that maps intent categories to mode handlers. It is more of a routing table than a traditional system prompt.

Envelope Format (confidence: high)

The entire instruction set is a JSON object with no markdown narratives. This JSON-only format is unique in the corpus — only Kiro uses it as the sole format. It is extremely efficient for routing and classifier tasks but is probably not suitable for general-purpose agents that need behavioral nuance. See template-variable-prompts for another minimal prompt approach.

Intent Classification Pattern (confidence: high)

The prompt follows a classify → gather → respond pipeline:

1. Classify: Determine user intent from a set of exhaustive categories
2. Gather: Collect preliminary context needed for the matched category
3. Respond: Route to the appropriate mode handler

Exhaustive Enumeration (confidence: high)

The prompt demands that all possible intents be listed with no truncation — never use “etc.” or “and more.” This is critical for classifier prompts where missing an option means misrouting a request.

Communication Style (confidence: high)

Follows the enumerate exhaustively style — systematic, comprehensive, and robotic. This style is one end of the communication-style-spectrum — appropriate for its role as a classifier rather than a general-purpose assistant.

Implications (confidence: high)

Kiro’s tiny JSON classifier proves that some tasks need minimal prompts. At ~500 tokens, it’s the most token-efficient prompt in the corpus. This is evidence that prompt size should match task complexity — not every tool needs a 30K-token epic.

Sources

• [raw/prompts/articles/taxonomy-synthesis-2026-07-16.md]