consulting-ai-assistants

Imported from secondtruth/agent-skills (plugins/agent-workflows/skills/consulting-ai-assistants/SKILL.md). Install upstream with npx skills add secondtruth/agent-skills --skill consulting-ai-assistants. Copyright stays with the author (MIT).

A consultation hands another model a brief and brings its answer back with provenance. This skill decides who gets the brief and through which door; the doors themselves are other skills: oracle-advisor and oracle for ChatGPT through oracle, and driving-ai-chat-websites for the sites, each when it is among your available skills.

Routing

Second opinions, reviews and design critiques follow this table. A YouTube video follows the youtube-video-ingestion skill when available.

"Ask the other AIs", a second opinion from "the others", a review round: that means ChatGPT, Gemini and Kimi, at least, each by its row above. Every one of them gets the same brief and the same context, and every answer is kept unabridged beside the others, each with its provenance: where the project keeps its reviews (docs/reviews/ is the usual place), and outside the repository when it keeps none. An assistant that cannot be reached — a login wall, a transport that fails twice — is reported as missing from the round, with the error; the round goes on without it.

Update the transport first

Before a round of runs, update oracle and every CLI the round uses, through the package manager that installed it:

brew upgrade steipete/tap/oracle    # Homebrew; likewise codex, kimi-code, antigravity-cli
scoop update oracle                 # Scoop, on Windows

The sites change their pickers and the vendors their models faster than an installed version ages well; oracle failing to find a model picker is the usual symptom. The package manager goes before a tool's own updater (kimi upgrade, agy update), which leaves the manager's record behind. Note the version that ran as part of the provenance.

A directory review runs on a disposable worktree

A CLI without a read-only sandbox of its own gets a copy of the tree instead:

git -C <directory> worktree add --detach <review root>/review HEAD
# the CLI runs in <review root>/review
git -C <directory> worktree remove --force <review root>/review

<review root> is one fixed directory outside every repository, reused for each review. The worktree protects the repository: it holds the committed files of HEAD, so .env files and untracked work stay out of the CLI's reach, and whatever it writes there vanishes with the worktree. When the review concerns uncommitted work, copy the named files into the worktree after checking them for credentials. The rest of the machine is protected by the prompt alone, so state in it that the review is read-only. kimi -p and agy --print take the prompt on the command line, where the process list shows it.

ChatGPT: oracle

The oracle MCP server (a consult tool) or the oracle CLI pastes prompt and files into ChatGPT itself, selects and verifies the model and thinking time, and stores every run as a session that survives a timeout. Follow the oracle and oracle-advisor skills for flags, model choice and the evidence to report when they are among your available skills.

Verified September 2026 with oracle 0.20 and 0.21:

• Pass files as absolute paths — the MCP server resolves relative ones against its own working directory.
• For long runs, start with waitForCompletion: false and block on the wait tool instead of holding the turn open.
• When oracle refuses to submit because it cannot confirm the requested model or tier, the account most likely lacks it. Pick one the account has.
• The browser engine is fragile about context: file uploads fail when the automated browser window is in the background, inline bundles truncate past roughly 15k tokens, and a multi-line prompt can arrive cut at its first paragraph. Before a long run, a mini run ("name the headings of the attached file") proves that context arrives. When it does not, report ChatGPT as unavailable for the day instead of burning attempts.
• One oracle run at a time; the same prompt twice needs --force.

chatgpt.com by hand, through driving-ai-chat-websites when it is among your available skills, is for the cases oracle leaves: oracle absent, or the chat itself the deliverable — a link the user continues in, a button they press there.

Gemini: oracle for text and files, agy on a directory, the site for the rest

oracle reaches gemini.google.com through its cookie client and takes long inline bundles without complaint (27k tokens in September 2026). Its Gemini support covers text, images and YouTube URLs. oracle 0.21.3 resolves gemini-3-pro to Gemini 3.1 Pro and reports the selection as unverified; it falls back to Flash-Lite unless --no-gemini-fallback is set. Notebooks, the newest models and the user's Google account need the website — as does a text consultation when oracle is absent or refuses the model: gemini.google.com through driving-ai-chat-websites when available.

On a directory, the Antigravity CLI (agy) answers headless and reads the files itself, run on the disposable worktree described above; agy models lists its models with their effort tiers:

cd <review root>/review && agy --model <model> --print "<prompt>"

Verified September 2026 with agy 1.2.8 and 1.2.13:

• Headless mode auto-denies every tool without an allow-rule, read_file included. A denied turn prints no answer and still exits 0, so an empty answer means a missing rule.
• One rule under permissions.allow in ~/.gemini/antigravity-cli/settings.json opens the review root for reading, subdirectories included: read_file(<absolute path of the review root>). Writes stay denied, which makes the review read-only by enforcement. Adding the rule changes the user's configuration: ask once.
• --dangerously-skip-permissions answers without a rule, on the footing Kimi's -p mode always has: the worktree protects the repository, the prompt the rest of the machine. --sandbox confines terminal commands only; the file tool still wrote into the home directory with it. Prefer the allow-rule.

Kimi: the CLI on a directory, kimi.ai otherwise

When the task names a directory — a codebase to review, a repository to critique — run the Kimi Code CLI (kimi) on a disposable worktree of it. It reads the files itself:

cd <review root>/review && kimi -m <model> -p "<prompt>"

Verified September 2026 with kimi-code 0.41 and 2.1.1 (default model k3-256k):

• -p - reads nothing from stdin; the prompt has to be the argument.
• -p mode edits files without asking, and refuses --plan; the worktree is the isolation.
• The CLI's plan runs on a five-hour quota. A 403 naming the usage limit means the window is spent; the website keeps working on the same account in the meantime.

Every other Kimi task goes to kimi.ai through driving-ai-chat-websites when available: a review that reads knowledge-base pages through Kimi's plugins, a prompt without a directory, a chat link the user continues in.

Codex: the CLI reads the files itself

For a second opinion from Codex, keep the prompt file and the answer file in a scratch directory and run, in the directory under review:

codex exec --sandbox read-only --skip-git-repo-check --ephemeral \
  -o <scratch>/answer.md - < <scratch>/prompt.md

Codex reads the files itself. The prompt arrives on stdin, so it stays out of the process list and out of shell interpolation, and --ephemeral keeps the session off disk, leaving the answer file as the only output. When it reports that the configured model requires a newer version, upgrade the CLI first (September 2026: 0.147 to 0.154 for gpt-6-astra). Record the model and reasoning effort from its header as provenance.

claude.ai: only for what lives there

A skill hosted on claude.ai is updated in a claude.ai chat — the install button lives in the resulting chat, and the user is the one who presses it, so the chat link is the deliverable. A skill installed from files is edited at its source, the repository it was installed from; the browser has no part in that.

The brief

A brief carries what the task needs: the question, the answer format, the constraints already decided, and the files and pages it names. A transport that takes no files gets them inline in the brief. Credentials, API keys and tokens stay out of every prompt. For a review, name the knowledge-base pages it should read; the request for the review covers sharing them, Kimi's Notion plugin fetches them itself, and a reviewer who knows the decisions already taken catches contradictions with them that a reviewer working from the brief alone misses.

Take the strongest tier the user's plan includes; tiers that cost extra credits stay unused unless the task names them.

Provenance

Store every answer with: the chat URL or session id, the version of the transport, the model and effort as requested and as observed (a picker label or an oracle selection log proves the UI selection, never which model served the answer), the transport, and the plugins or tools the assistant used. A consultation that returned text under an unverified model constraint is reported as exactly that. The answer is advice: verify it against the repository and the user's decisions before acting on it.