Skip to main content

Killer Use Cases

"Making it simpler" helps onboarding, but "making it necessary" drives adoption. Below are use cases where Cognitive's contracts and auditability are hard to replace with ad-hoc prompting.

1) High-Risk Decisions With Routing (Human-in-the-Loop)​

When an AI output can cause production impact, you want:

  • A strict output envelope (always the same shape).
  • Explicit risk and confidence for routing.
  • Schema validation to block malformed payloads.

Example pattern:

  1. tier: exec or tier: decision
  2. Enforce meta.explain (short) + data.rationale (long)
  3. Route results by meta.risk:
    • low: auto-apply / auto-merge
    • medium: require review
    • high: block + escalate

2) IDE-Native Workflows (MCP) With Streaming​

For tools like Cursor / Claude Code, the ideal experience is:

  • The tool gets streaming events (progress) and a final envelope (result).
  • The runtime enforces the same schema/policy rules as CLI and HTTP.

Recommended transport split:

  • MCP/HTTP: SSE streaming
  • CLI: NDJSON streaming

The important invariant is parity: the same module, policies, and final envelope regardless of transport.

3) Composable Multi-Step Workflows (Composition)​

Composition becomes a "protocol feature" when:

  • Each step emits a validated envelope.
  • The router/aggregator consumes typed outputs (not free-form text).
  • You can audit intermediate states and failures.

This is the boundary where Cognitive is no longer "a CLI tool", but "a workflow contract system".

4) PR Review Gate (CI)​

Turn AI code review into a CI gate:

  • Run pr-risk-gate on the PR diff
  • Block merges when meta.risk === "high" or data.blocking === true

Start here:

  • getting-started/pr-review-gate
  • getting-started/benchmark-evidence