independent-verifierlisted
Install: claude install-skill RICK-ORPHEN/agent-discipline-skills
# independent-verifier — one strong model, read-only, on purpose
Launches a **READ-ONLY sub-agent on the strongest model available**, in one of two modes.
| Mode | Invoked by | Input | Output |
|---|---|---|---|
| **verifier** | "verify this" | deliverable + acceptance criteria | verdict (pass / warn / fail) + findings |
| **advisor** | "ask it", "get a second opinion" | a question + reference material | opinion + grounds + risks + one recommended action |
The advisor mode is a light second opinion **before** you build or decide. When the answer is
genuinely contested and high-stakes, escalate to a multi-model deliberation instead.
## How it reaches the model
- A top-tier model is often **not available through a credential-brokered API**. Design on the
assumption that you cannot reach it that way.
- The route that works is **launching a sub-agent with an explicit `model:` on the host agent
platform.**
- Therefore this skill only runs inside such a session. If the target model is unavailable in
the receiving environment, **stop loudly** and say so — do not silently substitute another
model and report "verified."
## Absolute rules
1. **READ-ONLY** — do not give the sub-agent Write or Edit. Use a read-only sub-agent type.
This structurally kills the "while I was in there, I fixed it" accident.
2. **No silent downgrade** — if the requested model is unavailable, **do not quietly fall back**
to a weaker one. Show the error and stop. (Same principle as never fal