Skip to content
Agentmart
Submit a tool

consulting-cross-model

Use when a hard technical decision remains ambiguous after inspecting relevant evidence and choosing incorrectly would cause meaningful rework

Skill by basnijholt

IndexedMIT★ 1
Write a review

Install

Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.

npx agentmarket install skill/basnijholt-consulting-cross-model

Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.

Get the weekly diff

The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.

We use your email only for this newsletter. Privacy

About

Use when a hard technical decision remains ambiguous after inspecting relevant evidence and choosing incorrectly would cause meaningful rework

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

Permissions

What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.

Tools
No tools found
Secrets
None asked for
Install scripts
None found

Network hosts it talks to 1

  • codeload.github.com · found in install command

Files it touches

  • ~/.claude/skills/consulting-cross-model

What's inside

Skill name
consulting-cross-model
Description
Use when a hard technical decision remains ambiguous after inspecting relevant evidence and choosing incorrectly would cause meaningful rework
License
MIT
Instructions
110 lines in SKILL.md · view on GitHub

Files (1)

  • SKILL.md

SKILL.md (first lines)

# Consulting the Opposite Model

Use one independent model as an adviser, then own the decision yourself.
Codex consults Claude on Opus 5.5 (`claude-opus-5-5`); Claude consults Codex on
GPT-6 Astra (`gpt-6-astra`). Pin these models in every call. Consultation is
advisory, not an authorization and not a substitute for inspecting the
repository.

## When to Consult

Consult only when all four hold:

1. You already read the relevant code, tests, history, and instructions.
2. Two or more technically defensible options remain.
3. A wrong choice would cause meaningful rework or lock in an interface.
4. The ambiguity is technical. Missing product intent goes to your human partner.

Cheap reversible choices do not qualify. Decide those yourself.

## Stopping Rules

- **One consultation per decision.** Decide after one reply. No rebuttal round.
- If `BASPOWERS_CONSULT_DEPTH` is already set, you are the consultant. Never consult again.
- Advice never grants permission for destructive, irreversible, security-sensitive, or external actions.
- Bound the call to 300 seconds. Failure, timeout, or empty output means decide from existing evidence.
- If two decisions in one task already needed consultation, a third indicates an under-specified task. Ask your human partner.
  When executing a plan, each plan task counts as its own task.

End every prompt with:

> You are an adviser. Inspect and recommend; do not edit, commit, or run
> `agent-cli`. Do not consult another agent. Reply in under 300 words with:
> recommendation, strongest reason, strongest counterargument.

## Run the Consultation

Write a self-contained prompt under `.claude/consultations/`. Include the
decision, evidence, options, constraints, target repository path, and requested
output. Keep it git-ignored.

Prefer synchronous `agent-cli dev run`; it returns the answer directly and
does not create a task file:

Non-login shells can omit user-level executable directories. Before reporting
a CLI unavailable, resolve it from `PATH`, then check
`$HOME/.local/bin/agent-cli` and `$HOME/.bun/bin/{claude,codex}`. Pass `--`
after the worktree name so `agent-cli` does not parse the nested CLI's flags.
`codex exec` also reads stdin when it is not a terminal, so redirect it from
`/dev/null`; a background shell's open stdin otherwise hangs it until the
timeout.

```bash
# Codex -> Claude
consult_agent_cli=$(command -v agent-cli || printf '%s\n' "$HOME/.local/bin/agent-cli")
consult_claude_cli=$(command -v claude || printf '%s\n' "$HOME/.bun/bin/claude")
[ -x "$consult_agent_cli" ] && [ -x "$consult_claude_cli" ] || exit 127
BASPOWERS_CONSULT_DEPTH=1 timeout 300 \
  "$consult_agent_cli" dev run . -- "$consult_claude_cli" \
  -p --model claude-opus-5-5 --output-format text \
  --permission-mode plan "$(<.claude/consultations/question.md)"

# Claude -> Codex
consult_agent_cli=$(command -v agent-cli || printf '%s\n' "$HOME/.local/bin/agent-cli")
consult_codex_cli=$(command -v codex || printf '%s\n' "$HOME/.bun/bin/codex")
[ -x "$consult_agent_cli" ] && [ -x "$consult_codex_cli" ] || exit 127
BASPOWERS_CONSULT_DEPTH=1 timeout 300 \
  "$consult_agent_cli" dev run . -- "$consult_codex_cli" \
  exec -m gpt-6-astra --sandbox read-only --color never \
  "$(<.claude/consultations/question.md)" < /dev/null
```

When synchronous output is unavailable, use `agent-cli dev agent .` with the
opposite `--agent`, the same model flag in `--agent-args`, `--no-hooks`,
`-m tmux`, and `--prompt-file`. Require a
unique report path. Remove generated `TASK-*` and consultation files after
reading the report.

`agent-cli dev path .` can resolve to a superproject when called inside a
submodule. Compare it with `git rev-parse --show-toplevel`; if they differ,
name the exact nested repository path in the prompt so the adviser inspects the
right tree.

## Review Mode

baspowers:merging-when-approved uses the same call for an independent code review of a pull request.
In review mode these rules replace When to Consult, Stopping Rules, and the adviser footer:

- Review the full diff at the pinned head SHA, not a delta since an earlier review.
- End the prompt with this footer instead of the adviser footer:

  > You are a code reviewer. Inspect only; do not edit, commit, or run
  > `agent-cli`. Do not consult another agent. Report real defects with
  > file:line and a failing scenario, plus over-engineering and scope creep.
  > End with `VERDICT: APPROVE` or `VERDICT: CHANGES REQUIRED`.

- There is no word limit.
- Use `timeout 900` instead of `timeout 300`.
- Each push starts a new review round; review rounds don't count toward any consultation limit.
- `BASPOWERS_CONSULT_DEPTH` still applies, and the verdict gates the merge but authorizes nothing else.

## Decide and Record

Evaluate the advice against repository evidence. Record one line:

`Consulted <model>: <question> — <advice> — <decision>`

Then proceed. Escalate only if the remaining issue is missing human intent or
requires authority you do not have.

From basnijholt/baspowers, MIT.

More from basnijholt/baspowers

Signals

Stars
1
Forks
0
Open issues
0
Last commit (this folder)
Oct 8, 2026 (20 days ago)

From GitHub (fetched Oct 24, 2026).

Details

License
MIT
First seen
Oct 24, 2026
Last checked
Oct 24, 2026

More from basnijholt

Cite this page

consulting-cross-model skill. Agent Mart. https://agentmart.dev/skills/basnijholt-consulting-cross-model (last updated 2026-10-24)