grill-me-with-agents

name

grill-me-with-agents

description

Code-aware variant of grill-me-agents — interrogates a multi-agent design while continuously cross-referencing existing agent definitions, skills, prompts, and tool configs in the repo. Use ONLY when the agent stack is already implemented in the repo and you want to stress-test a *change* to it (versus designing from scratch — use `grill-me-agents` for greenfield). Triggers: "grill the existing agent stack", "stress-test the agent stack", "grill the implemented agents", "audit the agent topology against code". Do NOT use for greenfield agent design (use `grill-me-agents`) or for two agents grilling each other (use `peer-grill`).

Interview me relentlessly about every aspect of the proposed agent collaboration until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies one-by-one. For each question, provide your recommended answer.

Ask the questions one at a time.

Before asking any branch, scan the repo for ground truth and cite what you found. Look in (in priority order):

• .claude/agents/, .claude/shared-agents/ — existing agent definitions
• .claude/skills/, .claude/shared-skills/ — skills the agents may invoke
• .claude/settings.json, .claude/settings.local.json — permissions, hooks, env
• .claude/AGENT-STACK.md, CLAUDE.md, any AGENT_DESIGN.md — declared topology
• .github/workflows/ — CI hooks that gate or trigger agent runs
• Any orchestration code referencing the Agent SDK or MCP servers

If a branch can be answered from those sources, propose the answer for my confirmation rather than asking from scratch. Quote the file and line so I can verify.

If the codebase contradicts what I tell you, surface the contradiction and ask which is canonical before proceeding.

Cover the same 13 branches as grill-me-agents, in order, only descending once an upstream branch is settled:

1. Goal & success criteria
2. Agent roster — for each existing agent, confirm responsibility against its definition file
3. Topology — diff the proposed topology against the current one; flag every removed or repurposed agent
4. Context boundaries — check actual tools: allowlists in agent frontmatter
5. Handoffs & contracts — locate where messages cross agent boundaries today
6. Shared state — find where mutable state actually persists (files, MCP servers, env); distinguish from #4
7. Tool access — read each agent's tool list and flag privilege creep
8. Per-agent failure modes — check existing hooks and error paths in settings.json for each agent
9. Inter-agent disagreement & authority — find any precedence rules in agent prompts or orchestration code
10. Human-in-the-loop — locate AskUserQuestion, ExitPlanMode, and confirmation gates
11. Termination — find current stop conditions in orchestration code
12. Observability — check for logging/tracing infra already wired up
13. Cost & latency — note model assignments per agent (model: frontmatter)

Do not move to the next branch until the current one has a concrete answer, not a placeholder. If I give a vague answer, push back with the specific file or edge case that makes it vague.

When every branch is resolved, summarize the agreed design and offer to write it to AGENT_DESIGN.md using the template at ~/.claude/skills/grill-me-agents/AGENT_DESIGN.template.md (sister-skill template, intentionally shared). Include: roster (with file paths), topology diagram (ASCII), handoff contracts, failure/termination rules, observability hooks, and a diff against the previous design if one existed.

Sibling skills

• grill-me-agents — same 13 branches but greenfield (codebase-blind). Use when designing a new agent stack from scratch.
• grill-me — single-track plan/design grilling (not agent-specific).
• peer-grill — for two or more agents to grill each other (file-based reconciliation protocol).