Claude Friends

Before the calls that matter, Claude asks AIs from other families — and tells you where they disagree.

/plugin marketplace add awrshift/claude-friends
/plugin install claude-friends@awrshift-claude-friends

Then say "sanity check this", "this is important" or "help me choose" in any session.
Claude picks the review style from your phrasing, runs the reviewers, and reports where they
agree, where they split, and what it decided — reviewers are input, never a vote.

Ask your friends

Why more than one model

A single model's blind spots are systematic: it will make the same class of mistake confidently,
twice, when you ask it twice. The fix is not a second pass — it is a different vantage point.

Reviewer	What it is	What it is good for
Gemini	A different vendor and training distribution (Google)	Concepts, current practice, blind spots the Anthropic family shares
A second Claude, fresh eyes	Same family, isolated — no memory of your conversation	File-level facts. It reads the actual code and cannot anchor on the framing you already talked yourself into
GPT (opt-in)	A third family (OpenAI), needs a paid key	Widening the room on the highest-stakes calls only

The external reviewers see the brief you send, not your repository. That is deliberate: it keeps
private code with one vendor and keeps the honest split — outside families for judgement and
currency, the isolated Claude for anything that requires reading your files.

Three styles, picked from what you type

You say	What runs	Typical cost
"sanity check", "am I missing something"	One friend — a single critique	~3¢
"this is important", "before I ship"	The whole table — reviewers in parallel, then adjudication	~7¢
"help me choose", "brainstorm options"	Round-table — several rounds converging on one path	~25¢

The full list of phrases that trigger each style

One friend: ask a friend · sanity check this · am I missing something · stress-test this ·
critique this · give me a second opinion · cross-check this · review this · thoughts? ·
is this right? · devil's advocate · poke holes in this · what could go wrong

The whole table: this is important · run a full review · ask everyone · gather the friends ·
check before I send · before publishing · big decision · high-stakes review · boardroom debate ·
two independent opinions · don't let me ship something dumb · this can't be wrong

Round-table: help me choose between · brainstorm options · I have several paths ·
I have 3 angles on this · multiple options to weigh · diverge and converge ·
multi-round brainstorm · compare these approaches · what's the best direction

By name: ask Gemini · ask Opus · ask GPT · what would Gemini say ·
let's get a second model on this

You do not need an exact phrase — describe what you are stuck on and Claude picks the style.

Setup

The plugin calls two small CLI scripts. No MCP server, nothing running in the background.

pip install google-genai

# stored once, read by every shell — no export needed
mkdir -p ~/.gemini && printf '%s' 'YOUR_KEY' > ~/.gemini/api_key && chmod 600 ~/.gemini/api_key

A free key from aistudio.google.com is enough to start.
GOOGLE_API_KEY in the environment also works and takes precedence. A bare line in ~/.env
does not — nothing sources it in a fresh shell.

GPT is optional: pip install -U openai, a paid key in ~/.openai/api_key, and the details in
references/third-family-gpt.md.

What it will not do

It is not a fact lookup. "What is the latest X" is a job for WebSearch. This is critique.
It is not a vote. Three reviewers agreeing is not evidence; one dissenter holding a
file:line outranks three abstract concurrences. The skill adjudicates on merits and says so.
It does not run itself. It fires when you signal uncertainty, not on every answer.
It cannot see what you did not send. External reviewers get the brief; if the brief is
wrong, the review is confidently wrong with it.

Docs


SKILL.md	how the three styles run, anti-patterns, CLI reference
references/models.md	the model whitelist, with the date it was last verified against the live API
docs/DECISIONS.md	the open decisions board
docs/CHANGELOG.md	version history, including the rename from Don't Ask Me