Multi-AI Brainstorm for Muse

--- name: brainstorm-ia-gratuites-sans-navigateur description: Brainstorm a topic with several free AIs, no browser, through direct API calls. Same question asked to each AI, answers collected, duplicates removed, synthesis as top ideas. --- # Free multi-AI brainstorm, no browser needed ## The goal, in one sentence You give a topic, the assistant asks the same question to several free AIs through direct API calls (no browser, no clicks, no sign-in to a chat website), then hands you back a synthesis: the ideas that come back often, the best unique ideas, and the wild ideas kept apart. ## What "no browser" means, honestly With no browser, you CANNOT have the real free ChatGPT or the real free Claude. Their official APIs have no free tier. For those, you need the browser. With no browser, you can have: 1. **With no key at all, right now: Pollinations.** Anonymous endpoint, no sign-up, no key. A single model when anonymous: `gpt-oss-20b`. Slow: about 1 request every 15 seconds. Quality: small model, decent but not brilliant ideas. 2. **With a free key, set up once, no browser after that:** - Gemini API (Google AI Studio): free key, no card. Flash-Lite, about 500 requests per day. Real Gemini. - Groq: free key, no card. About 1000 requests per day on `gpt-oss-120b` and other open models. Very fast. - OpenRouter, `:free` models: a single free key, several models behind it. About 50 requests per day, 1000 after a one-time $10 credit purchase. - Pollinations with a free Seed key: unlocks more models (GPT, Gemini, DeepSeek, etc.), 1 request every 5 seconds. 3. **As a bonus, no browser needed: the PC's local LLM (Muscle, Qwen) if the PC is on, and Muse herself as the synthesis voice.** They are not AIs from the web, but they add different voices at no cost. Free keys are never pasted in a thread, never in this file, never in plain text in a script. They live in the assistant's secure vault, and the assistant uses them through its connector, without ever displaying them. Never send sensitive or personal data to a free tier: Google says it may use free-tier content to improve its products, and anonymous relays offer no confidentiality guarantee. ## Test verified on 2026-11-08 (no browser, from the VM) Test topic: "5 short name ideas for a skill that makes several free AIs brainstorm together", in French. - **Pollinations anonymous, POST https://text.pollinations.ai/openai, model `gpt-oss-20b`: WORKS.** Answer in about 13 seconds, 453 tokens in total. Result: 1. Éclat IA, 2. CréaCollectif, 3. BrainSync gratuit, 4. Fusion d'idées, 5. Conflux Cogité/Compréhensif. Average quality, one idea in five is vague, but the mechanism works. - **LLM7 anonymous, key `unused`: DEAD, tested the same evening.** `401 invalid_api_key`, despite August 2026 articles saying it worked with no account. Do not include it. - **Duck.ai by direct curl: BLOCKED, tested the same evening.** The status endpoint returns an anti-bot challenge, no easily usable token. It is a browser service, not a reliable API for this skill. Do not include it in the no-browser version. Test conclusion: with no key, there is a single reliable external voice today (Pollinations). To have 3 different external voices with no browser, you need at least one free key (Gemini, Groq or OpenRouter). ## Global rule — check everywhere, in every chat Max's decision, 2026-11-08: this rule applies to ALL chats, present and future, not just this skill. At every brainstorm, whatever the topic, the assistant tries to check everywhere: 1. No browser first: Pollinations with no key, then Gemini / Groq / OpenRouter if a free key is installed. 2. Muse herself, as synthesis and filtering voice. 3. The PC's local LLM (Muscle, Qwen) if the PC is on. 4. Browser only if Max asks for it explicitly: ChatGPT, Gemini web, Claude, Duck.ai. Each source is reported with what worked or not, never invented. A failure (401, anti-bot block, timeout) is noted as-is, and the brainstorm goes on with the voices that remain. Never present an invented idea as coming from an AI that was not asked. ## How it works, step by step ### 1. The user gives the topic The assistant asks, if it is not already clear: - the topic, in one sentence - the language of the answer - how many ideas (default: 10) - the constraints (budget, audience, tone, things to avoid) - which AIs to use (test default: Pollinations alone; default with keys: Pollinations + Gemini + Groq) If the topic is vague, the assistant restates the topic in its own words and waits for the yes before launching the calls. Do not launch 3 time-costly calls on a misunderstood topic. ### 2. Same prompt, everywhere The assistant writes ONE prompt, the same for every AI, for example: ```text Brainstorm on: <topic>. Give <N> short and different ideas, in <language>. Constraints: <constraints>. Format: numbered list, one idea per line, no long explanation. ``` Same prompt = honest comparison. If the prompt changes between two AIs, you no longer know whether the difference comes from the AI or from the prompt. ### 3. Calls, one by one, no browser - Pollinations anonymous: `POST https://text.pollinations.ai/openai` `{"model": "gpt-oss-20b", "messages": [{"role": "user", "content": "<prompt>"}]}` Wait at least 15 seconds between two anonymous calls. - Gemini / Groq / OpenRouter: same OpenAI-compatible format, with the key from the vault, never in plain text. Base URL depends on the provider, model depends on the key. - Each answer is kept as-is, with its source (provider + model + time). Do not rewrite an answer before the synthesis, or you lose track of who said what. If a call fails (401, 429, timeout), note it as-is, do not invent it, and carry on with the others. A 2-voice brainstorm is better than an invented 3-voice one. ### 4. Synthesis The assistant, as synthesis voice, does this: - removes duplicates (same idea, different words = one idea) - counts the ideas that come back across several AIs - keeps the best unique ideas, even if only one AI said them - puts wild or risky ideas apart, without throwing them away - returns a top 5 or top 10, with the source of each idea - proposes a single concrete next step Default output format: - Top ideas (with source) - Ideas that come back everywhere (consensus) - Unique ideas worth keeping - Wild ideas, to keep - Proposed next step ### 5. Second round, only if the user wants it If the user picks 2 or 3 ideas, the assistant can run ONE second round: same refinement prompt to every AI, on those ideas only. No automatic second round: each round costs time and free quota. ## Limits to tell the user, without hiding them - Free tiers change often: model, limit, price. Check live before promising a number. - Anonymous Pollinations is slow and small. It is a test, not production. - Ideas are not verified: a free AI can invent a fact, a price, a name already taken. Check before using an idea as a fact. - No official ChatGPT or Claude without a browser. - Free quotas are per day and per account, not infinite. A brainstorm with 3 AIs, 2 rounds, is 6 calls. ## What this skill does not do - It opens no browser. - It does not sign in to any ChatGPT, Claude, Google or other account in the user's place. - It does not paste any key, password or secret in a thread or in this file. - It does not publish anything, on Skill Harbor or anywhere else. - It does not promise an exact number of free messages per day: limits change, they are checked. ## Test checklist - [x] Pollinations anonymous: works, 2026-11-08 - [x] LLM7 anonymous: dead (401), 2026-11-08 - [x] Duck.ai by curl: blocked (challenge), 2026-11-08 - [ ] Test with a real topic from Max, 3 voices - [ ] Test with a free key (Gemini or Groq), if Max wants it - [ ] Reviewed and approved by Max before any other use