In partnership with
When you are stuck on a hard AI task, what do you do?
WTF is an AI council?: A router picks one model. A council asks many models.
Then one judge model reads all answers and writes the final version.
For a few days, some builders felt like they had a much smarter model in their hands.
X called it "Fable-level intelligence."
Then access changed.
Limits changed.
The usual reaction is:
Wait for the lab to bring the smart model back.But there is another move:
Do not wait for one genius model.
Build a small room of models.Not as a metaphor.
As an actual setup.
3 models answer
1 judge compares
1 final answer comes outI am Alex, welcome to ShortCu8 by Innov8.
Lets Dive Deep 🐰
⭐Today's Shortcut
Use AI councils only for questions where the answer is expensive to get wrong.
Use it when you would normally open the most expensive model and hope.
The simple rule:
Easy work = one cheap model
Normal work = one good model
Hard decision = councilThe council decides.One agent executes.
Tests or sources verify.
Try it in OpenRouter Fusion
This is the easiest place to feel the idea.
OpenRouter has a Fusion page where you can run a fused answer without building your own setup.
Use this when:
you want to test the idea today
you do not want to install anything
you are okay paying per requestTry a question that has real tradeoffs:
I am building a small SaaS for Indian creators.
Should I use Supabase, Firebase, or a normal Postgres backend?
Compare:
- cost
- speed to build
- lock-in
- auth
- payments
- long-term migration risk
Give me one final recommendation.That is a council-shaped question.
The value is in the disagreement.
One model may care about speed. One may care about scale. One may care about lock-in.
The judge has to read all three and make a call.
Do not waste Fusion on:
Rewrite this sentence.
Make this title shorter.
Summarize this note.
Give me 10 names.That is like calling a meeting to choose a font size.
Try openfusion if you want control
There is also an open-source version of this idea called openfusion.
Its README describes a compound-model proxy:
your prompt
parallel model panel
judge model
one synthesized answerThe nice part is that it works like an OpenAI-compatible API.
So another app can call:
model: "openfusion"and the proxy handles the panel and judge behind it.
It also has presets like:
quality
budget
customUse this when:
you want to tune the model panel
you want to choose the judge
you are comfortable with API keys
you do not want a black-box setupKeep the warning from the README:
For easy tasks, it can cost more and perform worse than one normal model.
That is good to know.
The council earns its cost only when the first answer is likely to miss something important.
Try the manual version first
You can also do this without any tool.
Open three chats:
ChatGPT
Claude
GeminiOr replace one with Kimi, DeepSeek, or another model you use.
Ask all of them the same hard question.
Then paste the answers into your best model:
You are the judge.
Read these answers.
Do not average them.
Find:
- where they agree
- where they contradict
- what each answer missed
- which answer is strongest
- what the final decision should be
Then give me one final answer.This is slower.
It is messy.
But it teaches the pattern fast.
That disagreement is often the useful part.
Router vs council
Do not confuse this with model routing.
Router:
Pick the best model for this task.Council:
Ask multiple models.
Use one judge to compare the answers.Example:
Router:
This is a coding task, send it to Claude.
Council:
Ask Claude, GPT, and Gemini for the migration plan.
Ask one judge to compare the plans.
Then create the final plan.If there is no synthesis step, it is not a council.
It is just a switchboard.
When to use it
Use a council for:
architecture decisions
database choices
migration plans
pricing strategy
serious research
debugging with multiple theories
article angle selection
business decisionsDo not use a council for:
formatting
small rewrites
basic summaries
simple code edits
tiny bug fixes
caption writingAsk this before using it:
Would I normally burn my best model on this?If yes, try a council.
If no, use one model and move on.
The best setup
For most builders, the useful setup is:
cheap models = panel
strong model = judge
coding agent = executor
tests = final refereeFor writing and research:
judge reads and synthesizesFor code:
judge should care about tests, not vibesThat last line matters.
A council can explain bad code beautifully.
So for code, the judge should ask:
Does it run?
Do tests pass?
Is the diff smaller?
Did it break fewer files?The council should decide the plan.
The coding agent should do the work.
Now go build something great..🙂🙂
The ShortList
🛠️Cool Tools of the Week:
Sakana Fugu: An LLM trained to call on other models in an agent cool, designed to orchestrate other models in complex, multi-step tasks.
Grok Imagine: xAI's flagship model now has multitasking agents.
HappyHorse 1.1: Alibaba's latest model for production-ready video synthesis.
ElevenLabs' Ad Engine: Users can now connect Google, Meta, and LinkedIn ad accounts to ElevenLabs
📩 Innathe Shortcu8 engane undarunnu 👇️?
We read every reply - just reply to this email and let us know how we can improve !
Appo adutha Shortcu8il kanaam bie…👋
If you read till here, you might find this interesting
#AD1
Your membership list is full. Your calendar isn't.
You have hundreds of members needing tune-ups twice a year.
Scheduling them means multiple calls and weeks of manual outreach that never fully gets done.
The ones who don't get reached fall through the cracks. Those are members and recurring revenue quietly walking out the door.
Podium's AI Membership Coordinator texts every member, books the appointment, follows up automatically, and handles renewals.
It works your entire list proactively so your slow season fills before it starts.
No manual outreach. No dropped members. No empty shoulder season.
#AD2
AI/Tech Angle A, June - Secondary
Claude vs Gemini. GPT-7 vs Llama 5. Which AI lab ships AGI first. These are live Kalshi markets with real money on both sides, updated in real time as releases land. The person who follows model cards and tracks evals has a genuine edge here. If that's you, trade it.