advisor

name

advisor

description

Consult Claude CLI (Fable model, effort high or xhigh) as a senior technical advisor for hard architectural decisions, design trade-offs, gnarly debugging hypotheses, plan reviews, or any problem where a stronger second opinion is worth the cost. Consultation only — never edits files. Use when Grok is stuck on a hard problem, needs an advisor consult, second opinion, architecture review, or the user runs /advisor.

Advisor (Claude Fable consult)

Call Claude Code CLI with Fable at high effort for a read-only second opinion. You (Grok) remain the implementer; the advisor only advises.

When to use

• Hard architecture / design trade-offs
• Gnarly debugging where hypotheses disagree
• Reviewing a plan before a large or risky change
• User asks for /advisor, "consult fable", "second opinion", "ask the advisor"

When not to use

• Routine questions you can answer confidently
• Simple refactors, typos, or lookups
• When the user only needs a quick fact
• Prefer not to spam this: each call is expensive

Effort

Situation	Flag
Default hard problem	--effort xhigh
Faster / cheaper consult (still hard, not critical)	--effort high
User says "max" / "ultrathink"	--effort max if available, else xhigh

Default to xhigh unless the user asks for high or the problem is mid-weight.

How to call

From the current project directory (so Claude can read the repo):

claude -p \
  --model fable \
  --effort xhigh \
  --output-format text \
  --tools "Read,Grep,Glob,Bash,WebSearch,WebFetch" \
  --append-system-prompt "$(cat <<'SYS'
You are a senior technical advisor consulted by Grok (another coding agent).
Consultation only: advise, never modify files or system state.
Use Bash only for read-only inspection (logs, git history, status, builds already run).
Do not write, edit, commit, push, or run destructive commands.

How to advise:
- Restate the question you are answering in one sentence. If a better question is hidden, answer that and say why.
- Ground advice in the code: read relevant files yourself; if the caller's summary is wrong, lead with that.
- Give a single clear recommendation, not a menu. Mention alternatives only to reject them.
- State assumptions and what evidence would invalidate the recommendation.
- If underspecified, name exactly what is missing and how to get it.
- Keep the answer tight: recommendation first, reasoning after, no preamble.
SYS
)" \
  -- "$(cat <<'PROMPT'
# Advisor consult (from Grok)

## Problem
<1-3 sentence problem statement>

## What I already tried / believe
- ...

## Relevant paths / symbols
- path/to/file.go — why it matters

## Question for you
<the decision or hypothesis to resolve>

## Constraints
- ...
PROMPT
)"

Helper script (preferred)

If present, use the wrapper (same behavior, less prompt drift):

~/.grok/skills/advisor/scripts/consult.sh xhigh <<'PROMPT'
# Advisor consult (from Grok)
...
PROMPT

First arg is effort: high | xhigh | max (default xhigh).

Notes

• Requires claude on PATH and auth already working (claude auth / existing login).
• Do not pass --dangerously-skip-permissions for advisor consults.
• --tools is restricted to inspection tools; do not add Edit/Write.
• Timeout: Fable + xhigh can take several minutes. Use a long shell timeout (e.g. 10–15 min).
• If claude fails (not installed, auth, model unavailable), report the error to the user and fall back to your own best judgment.

After the reply

1. Summarize the advisor’s recommendation to the user in your own words (cite key points).
2. Say whether you will follow it, adapt it, or reject it — and why.
3. Only then implement (if implementation was requested). Do not silently ignore the advisor.