Intent-Router

English | 简体中文

An intent compiler for AI agents.

Turns a vague request into a typed IntentSpec: it looks up what it can, asks only what it can't, calls out a request that contradicts itself, admits two readings, or rests on a premise your own repo already disproves, and refuses to emit while the intent is still underspecified. When the work gets handed off and carried out, it checks the result against the spec it wrote — one constraint at a time, not a vibe check.

What that buys you:

• Prompt equity. You get the interaction of a senior engineer without having to learn prompt engineering — the gap between the top 1% of prompt writers and everyone else stops deciding what you get back. Say it the way you'd say it to a capable teammate, "add caching to the user API", and the skill writes the briefing a senior engineer would have written first. You stop coaxing the model and start putting it to work.
• It asks one question, and the answer lands in the delivered code. Before anything reaches you it reads what your repo, ticket system or docs already answer, so the forty-six-question interview becomes the one question that genuinely needs your judgment — and in the delivery comparison, every run with the skill asks exactly that one question and 3 of 3 delivered caches state their failure policy, while none of the 5 bare deliveries do (report).
• The work carries further. Every run ends in a machine-readable IntentSpec, saved to .intent/<intent>.intent.yaml before the reply that carries it, whose probed fields carry evidence pointers and which records its own state: still waiting on a question, settled, or already carried out. A later session reads that file before probing anything — an answered question is never asked again, a settled contract is implemented as written — and the result of checking the delivered work against it is written back into the same file. So the next agent, session or teammate starts from what you actually decided instead of zero.

What "installed" changes, in the diff itself

Same weak request, same fixture, same permissions, one line apart — bare vs. with the skill (N=5 bare, N=3 skill). Excerpted verbatim from each run's own diff.patch, the part that decides what happens when the cache itself fails:

Bare — no failure path exists to read. A downed Redis takes the request down with it:

const hit = await getCache<T>(key);
if (hit !== null) {
  return hit;
}
const value = await load();
if (value !== null) {
  await setCache(key, value);
}
return value;

With the skill — asked the one question no file could answer, then wrote the answer into the code:

async function readCache<T>(key: string): Promise<{ hit: boolean; value: T | null }> {
  try {
    return { hit: true, value: await getCache<T>(key) };
  } catch {
    return { hit: false, value: null };
  }
}

Every one of the 5 bare runs shipped the first shape; every one of the 3 skill runs shipped the second. Neither line was written by hand for this README — both come straight out of the runs that produced the numbers in Evals.

          "add caching to the user API"       "this customer is furious — sort it out"
                        │                                         │
                        ▼                                         ▼
 ┌────────────────────────────────────────────────────────────────────────────────────┐
 │  parse  →  resolve  →  typecheck  →  emit                    one engine, one spec  │
 ├────────────────────────────────────────────────────────────────────────────────────┤
 │  PROBE   deps, routes, git history, ADRs   │  order + ticket log, entitlements,    │
 │          — you are not interrupted         │  the policy in force                  │
 │                                            │                                       │
 │  ASK     serve stale, or serve uncached?   │  refund, or replace?                  │
 │          the one answer no file holds      │  the one that can't be undone         │
 └────────────────────────────────────────────────────────────────────────────────────┘
                        │                                         │
                        ▼                                         ▼
              IntentSpec  ──────►  your planner / agent / workflow / human

Same three passes. Same stopping predicate. Only the places it looks change.

What a real run looks like — one sentence in, it reads the repo itself and asks only what no file can answer. From the measured run of 2026-09-23 behind the numbers below: 8 of the 10 cases the suite held that day, 0 hallucinated evidence.

No install, no API key, no dependencies — it's a skill:

npx skills add angel291592/Intent-Router

Manual install, or an agent the installer doesn't know: Quick start.

Contents

• How it compares
• The problem · How it works · The four decision states
• Two walkthroughs — one in a codebase, one in a support queue
• The artifact · Where it works · Quick start · Evals
• Design principles · Limitations
• Backends · Prior art · Contributing

How it compares

Intent-Router is the layer between a vague request and the work: look it up, ask, or act — then check what got built. Your harness's own plan mode already prefers looking things up over asking, and grill-me converges beautifully and states the right principle — but neither one checks a stated premise against the evidence, and neither one checks the finished work against what was decided. spec-kit keeps everything and makes you buy its whole workflow to get it. Kiro's specs come closest on paper — requirement-level conflict analysis, a correctness check after delivery — at the cost of a full three-file spec workflow and an IDE. Jev and Laya make the same bet this skill makes — typed output instead of prose — one layer further down: both turn an already-clear input into a typed, calibrated decision in a single pass, and neither one converges a vague request into that clear input to begin with. Intent-Router is that upstream layer as a single skill: no workflow to adopt, no files to keep in sync by hand, no IDE, no separate decision service to run.

Read the row gaps, not the checkmarks. Every cell below is sourced to the tool's own docs or repo — quoted cells are verbatim, and an unquoted ✅/⚠️/❌ still traces to a specific page.