Installs into .claude/skills of the current project.
Are you the author of Advisor?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/nikolanovoselec-advisor)
---
name: advisor
description: User-invoked only; use advisor solely when the user explicitly requests advisor.
disable-model-invocation: true
---
# Advisor (Pi): User-Invoked Stronger-Model Review
The `advisor` tool (provided by the `@juicesharp/rpiv-advisor` extension) hands the **entire current conversation branch** to a stronger reviewer model and returns guidance.
## Hard gate — user request only
Only the user may invoke advisor. Do **not** call the `advisor` tool, run `/advisor`, or suggest `/advisor` unless the user's current message explicitly asks for advisor, for example:
- `/advisor`
- "ask the advisor"
- "use the advisor"
- "call advisor"
- "check this with the advisor"
The assistant must not invoke advisor proactively for:
- planning;
- being stuck;
- routine debugging or CI fixes;
- code review;
- risky or irreversible actions;
- before writing, committing, pushing, deploying, or declaring done;
- generic "second opinion" requests that do not name advisor.
If advisor might help but the user did not explicitly ask for it, continue without advisor or ask a normal clarification question. Do not mention advisor as a required step, and do not tell the user to run `/advisor` proactively.
## Configuration
Advisor authenticates through **Pi's model registry** and uses the model selected by the user through the user-only `/advisor` command. The selection persists across sessions at `~/.config/rpiv-advisor/advisor.json`. Only the user should open `/advisor`; the assistant must not run, simulate, or recommend that command unless asked.
## When explicitly requested
When the user's current message explicitly requests advisor, call `advisor` with no parameters. Relay the returned guidance and make clear what, if anything, it changes.
## Troubleshooting
If the `advisor` tool is unavailable or returns an error, report that directly. Do not retry repeatedly and do not substitute another external LLM unless the user explicitly asks for that.