Every framework in the library has:
  • a slug (stable identifier — pin this in your code),
  • bilingual prompts (en and zh, kept structurally parallel),
  • an ELO score updated from real user feedback,
  • versioned content (FrameworkVersion) so you can pin a specific revision for reproducibility.
The library is at 5 frameworks today and is growing toward 30. New frameworks ship behind the same routing engine — no client changes required.

​
Live frameworks (V0)

First Principles

Strip the problem to its irreducible facts and reason up from there. Best for: ambiguous specs, “we’ve always done it this way” inertia, novel design decisions.
slug: first-principles

Inversion

Solve the inverse — what would guarantee failure? — then avoid those moves. Best for: risk audits, pre-mortems, spotting hidden constraints.
slug: inversion

Second-order Thinking

Don’t stop at the first consequence — chase the consequences of the consequences. Best for: policy decisions, refactor vs. rewrite, anything with feedback loops.
slug: second-order-thinking

Probabilistic Thinking

Replace yes/no with distributions. Best for: forecasting, incident triage, calibrating confidence on partial info.
slug: probabilistic-thinking

Sunk-cost Fallacy

Recognize when past investment is biasing a forward decision and correct for it. Best for: project kills, deprecation calls, vendor switches.
slug: sunk-cost-fallacy

​
Pinning a specific version

Every routing response includes selected.frameworkVersionId. Store it next to your agent’s output if you need full reproducibility — the same frameworkVersionId will always return the same prompt text, even if we later edit the canonical framework.

​
Excluding frameworks

If you’ve already chained one framework and want the next-best for the same intent, pass excludeSlugs: 
{
  "intent": "should we deprecate this product line?",
  "excludeSlugs": ["sunk-cost-fallacy"]
}
The router will still rank candidates the same way; it just removes the listed slugs from the pool.

​
Roadmap

The next batch (30 total) covers decision-making, debugging, planning, communication, and learning patterns. We ship them in small waves as ELO data accumulates so the rerank step has signal to work with from day one.