Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Send a task or question to Claude Code and/or Gemini as external subagents and report back what they said. Use when the user asks to ask Claude, ask Gemini, ask both, get a second opinion, split reviewer roles, or compare external agents.
.claude/skills/chrisblattman-ask-agents/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -32% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 37% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 30% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 59% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 132% | 0% |
When the user asks to ask Claude, ask Gemini, ask both, split work between them, or get an outside opinion, do it directly with the bundled runner. Do not turn it into an integration discussion.
Sending a task to Claude Code or Gemini sends the prompt, listed context files, and any quoted material to another AI service. Before every run, check whether the prompt or context includes confidential records, personal contact details, human-subjects or research-participant data, private transcripts, unpublished sensitive work, or anything the user would not paste into a public website. If so, do not run the external agent. Offer a local Codex-only answer or a redacted prompt for the user to approve first.
If the privacy check passes, ask for explicit send confirmation before every external run: "Send this prompt to Claude/Gemini as an outside opinion? (yes/no)" Proceed only if the user answers yes. If they answer no or do not answer clearly, stop and offer a Codex-only answer.
node on your PATH)--agent claude: the Claude Code CLI (claude) installed, signed in, and on your PATH--agent gemini: the Gemini CLI (gemini) installed, signed in, and on your PATHIf a required tool is missing, the runner stops with a plain-language message naming it. Never substitute your own answer while implying the external agent ran.
The runner is bundled inside this skill at scripts/run-external-agent.mjs:
textnode ~/.codex/skills/ask-agents/scripts/run-external-agent.mjs --agent claude --prompt "..." node ~/.codex/skills/ask-agents/scripts/run-external-agent.mjs --agent gemini --prompt "..." node ~/.codex/skills/ask-agents/scripts/run-external-agent.mjs --agent both --prompt "..."
(If your skills live somewhere other than ~/.codex/skills, adjust the path — the script is in this skill's own folder.)
Defaults:
cwd: the folder the command is run from (your current project)External runs are read-only: the runner snapshots the working folder before and after, reports any file changes, and instructs the external agent to propose patches rather than apply them.
Useful options:
text--cwd /path/to/workspace --context relative/path/under/cwd --timeout 600 --max-output 100000 --max-turns 10 --max-budget 4 --claude-prompt "implementation/reliability role..." --gemini-prompt "strategy/alternative role..."
Claude implementation review:
textReview this as Claude Code. Focus on implementation risk, missing tests, operational failure modes, and the smallest robust next step.
Gemini outside view:
textReview this as Gemini. Focus on alternative approaches, strategy/product tradeoffs, hidden assumptions, and what we may be overcomplicating.
If an external run fails, say it failed and include the details that matter (state, error, missing tool). If the command was blocked by sandboxing or approvals, ask the user to approve that specific command and rerun it.
textExternal agents: Claude: Gemini: Synthesis: Next action:
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-20 | pass→pass | 16,843 | 12,230 | -27% | 1 | 1 | 0% | 2,598 | 2,889 | +11% | 0 | 0 | — |
case-21 | fail→fail | 3,819 | 2,250 | -41% | 1 | 1 | 0% | 605 | 1,317 | +118% | 0 | 0 | — |
case-01 | fail→pass | 16,665 | 4,243 | -75% | 1 | 1 | 0% | 2,450 | 1,655 | -32% | 0 | 0 | — |
case-02 | fail→fail | 7,251 | 3,151 | -57% | 1 | 1 | 0% | 281 | 1,458 | +419% | 0 | 0 | — |
case-03 | fail→pass | 6,761 | 5,972 | -12% | 1 | 1 | 0% | 1,054 | 1,440 | +37% | 0 | 0 | — |
case-04 | fail→pass | 8,004 | 3,904 | -51% | 1 | 1 | 0% | 1,224 | 1,597 | +30% | 0 | 0 | — |
case-05 | pass→pass | 13,895 | 8,913 | -36% | 1 | 1 | 0% | 2,056 | 2,280 | +11% | 0 | 0 | — |
case-06 | fail→fail | 22,960 | 8,042 | -65% | 1 | 1 | 0% | 3,911 | 2,467 | -37% | 0 | 0 | — |
case-07 | fail→pass | 16,851 | 6,124 | -64% | 1 | 1 | 0% | 1,343 | 2,131 | +59% | 0 | 0 | — |
case-08 | fail→pass | 5,596 | 7,397 | +32% | 1 | 1 | 0% | 994 | 2,304 | +132% | 0 | 0 | — |
case-09 | fail→pass | 36,506 | 6,095 | -83% | 1 | 1 | 0% | 2,721 | 1,946 | -28% | 0 | 0 | — |
case-10 | fail→pass | 15,526 | 8,914 | -43% | 1 | 1 | 0% | 2,189 | 2,463 | +13% | 0 | 0 | — |
case-11 | fail→pass | 16,302 | 7,437 | -54% | 1 | 1 | 0% | 2,399 | 2,248 | -6% | 0 | 0 | — |
case-12 | pass→fail | 9,032 | 4,207 | -53% | 1 | 1 | 0% | 1,464 | 1,618 | +11% | 0 | 0 | — |
case-13 | fail→pass | 11,823 | 3,598 | -70% | 1 | 1 | 0% | 1,556 | 1,481 | -5% | 0 | 0 | — |
case-14 | pass→pass | 15,084 | 10,845 | -28% | 1 | 1 | 0% | 2,462 | 2,681 | +9% | 0 | 0 | — |
case-15 | fail→pass | 6,754 | 6,022 | -11% | 1 | 1 | 0% | 984 | 1,939 | +97% | 0 | 0 | — |
case-16 | fail→fail | 11,892 | 7,181 | -40% | 1 | 1 | 0% | 1,855 | 1,983 | +7% | 0 | 0 | — |
case-17 | pass→pass | 3,934 | 4,324 | +10% | 1 | 1 | 0% | 483 | 1,641 | +240% | 0 | 0 | — |
case-18 | fail→pass | 10,790 | 91,345 | +747% | 1 | 1 | 0% | 1,124 | 1,551 | +38% | 0 | 0 | — |
case-19 | fail→fail | 1,976 | 2,733 | +38% | 1 | 1 | 0% | 253 | 1,407 | +456% | 0 | 0 | — |
case-22 | fail→pass | 8,511 | 2,274 | -73% | 1 | 1 | 0% | 1,312 | 1,291 | -2% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 21 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +50 percentage points is the difference between those two pass rates over the 21 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.