Consult Claude specialist agents during implementation for codebase understanding, pattern checking, security review, debugging help, and more. Use this skill whenever you're unsure about conventions, stuck on a failure, or need expert input before writing code. Does not replace the formal review gates in agent-loops — this is for mid-implementation consultation.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Consult Claude specialist agents during implementation for codebase understanding, pattern checking, security review, debugging help, and more. Use this skill whenever you're unsure about conventions, stuck on a failure, or need expert input before writing code. Does not replace the formal review gates in agent-loops — this is for mid-implementation consultation.
["unsure about codebase conventions","stuck on a test failure","need security or architecture guidance","before writing code in an unfamiliar area"]
Claude Consult
Get expert help from Claude specialist agents during implementation. Each agent
has domain-specific skills pre-loaded and tools scoped to its role.
This is NOT a replacement for the review gates. The agent-loops skill
defines formal code review (specialist-review) and test audit
(test-review-request) gates that run after implementation. This skill is for
asking questions during implementation — before you've written code, while
you're writing it, or when you're stuck.
When to Use
"How does this module work?" — before implementing
"Does my approach match existing conventions?" — while implementing
"Is this pattern secure?" — before committing to a design
"Why is this test failing?" — when stuck
"What do the testing standards say about X?" — before writing tests
"Is this component accessible?" — during UI work
How to Invoke
claude -p --agent <agent-name> "<your question>"
All agents run headless (-p / print mode). They read the codebase, apply their
loaded skills, and return an answer to stdout. Capture or pipe as needed:
# Direct invocation
claude -p --agent debugger "Why is test_auth_flow failing in src/auth/tests.py?"# Capture to variable
ANSWER=$(claude -p --agent security-auditor "Is the input validation in src/api/handler.py safe?")
# Save to file
claude -p --agent architect-reviewer "Does adding a cache layer between api and db match the existing architecture?" > /tmp/arch-advice.md
These agents advise — they don't write your code. Read the answer, apply it
yourself. The agent may suggest approaches, but you implement them.
2. Scope your questions
Bad: "Review everything in src/"
Good: "Is the input validation in src/api/handler.py:45-60 safe against injection?"
Narrow questions get better answers and use fewer tokens.
3. Don't replace review gates
Consulting security-auditor during implementation doesn't skip the formal
specialist-review gate afterward. The review gates exist for independent
verification — consulting during implementation is for getting guidance early.
4. Use the cheapest agent that works
For factual lookups (where is X, what does the standard say), prefer
search-specialist or the language-specific agents. Reserve
security-auditor, debugger, and performance-engineer for questions
that need deeper analysis.
5. One question per invocation
Each claude -p --agent call is a cold start. Don't try to batch multiple
unrelated questions in one call — they'll get shallow treatment. Ask one
focused question per invocation.
Relationship to agent-loops
agent-loops claude-consult
───────────── ──────────────
Formal review gates Mid-implementation help
Runs AFTER implementation Runs DURING implementation
Produces reports with P0-P3 Answers questions directly
Required before merge Optional, use as needed
specialist-review + test-review Any agent from the roster
Both skills are designed to work together. Consult agents while implementing
to prevent issues, then run the formal review gates to verify nothing was missed.
