agentic-ai-engineering/skills/kits/verifier-pro/SKILL.md

52 lines
1.6 KiB
Markdown

# Verifier Pro Kit
Catch agent mistakes before they hit production. Read-only verification.
## Included Skills
1. **Verifier Builder** — Two-agent observer pattern
2. **Confidence Ladder** — PERFECT → FAILED grading
3. **Claim Decomposition** — Break every claim into atomic units
4. **Read-Only Surface** — Restrict to read/grep/find/ls
5. **Escalation Handler** — Max 3 loops, then human
## Architecture
```
Builder (your terminal) ──unix socket──► Verifier (new window, input LOCKED)
│ │
▼ writes: ▼ reads (read-only):
session.jsonl session.jsonl
│ │
◄──── verifier_prompt (corrective feedback) ──┘
```
## The Confidence Ladder
| Level | Meaning | Bar Color |
|-------|---------|-----------|
| PERFECT | Every claim verified, zero gaps | Green |
| VERIFIED | All checks passed, minor gaps | Green |
| PARTIAL | No failures, significant unverifiable gaps | Orange |
| FEEDBACK | Claims failed, correction sent | Orange |
| FAILED | Cannot verify, escalating to human | Red |
## Key Properties
- Builder DOES NOT KNOW verifier exists (pure observer)
- Verifier input is LOCKED (structurally un-promptable)
- Verifier has NO write tools (enforced at tool surface level)
- Max 3 correction loops, then escalate to human
## Usage
```bash
# Start both agents
just verifier
# Builder prompts normally
# Verifier auto-runs after each builder turn
# If verification fails → corrective feedback sent
# After 3 failures → escalates to you
```