The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
I built a lightweight framework inspired by the Schema harness, a system that pushed AI models to ~99% on ARC-AGI-3 just by changing how the model thinks, not the model itself.
Here is how to get frontier models to solve complex, ambiguous problems without hallucinating or running in circles, tell your agents to follow these rules:
1. Grounded: Define what exists first. Name objects, variables, and state based strictly on past context before guessing where things are going.
2. Ruly: Make rules explicit. State exactly what happens next and why. No vague vibes.
3. Consistent: No ad-hoc excuses. If a rule doesn’t match past history, don't patch it with weird exceptions. The rule is wrong.
4. Careful: Don’t guess in the dark. If you’re uncertain, pick the one test that eliminates the most confusion with the fewest moves.
5. Clear: Call your shots. State the expected result before acting. If reality doesn’t match your prediction, immediately drop the plan, step back, and re-ground your assumptions.
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Trending
retrofuturism
Share your favorite designs from the past exploring the distant future. Or your own take on the aesthetic 🚀
halftone
Dots. Texture. Grain. Let's see your designs using the halftone technique.
Midjourney
Creatives are showing off what they're making with Midjourney.
colorgrade
Video editors sound off: How do you enhance the mood of your video in color grading?
Challenges
View all
Related posts
One thing I’ve learned from vibecoding: don’t break what already works.
When I’m building a new feature, I usually create a separate test area instead of modifying the live version directly.
If the current version is v3, I’ll experiment with v3.1 there first, test it, iterate, and only move it into the main app once it’s stable.
For bigger rewrites, I keep both versions locally and compare them side by side. The old one stays untouched while I work on the new one.
It’s a simple workflow, but it saves me a lot of headaches when AI starts losing context.
There are many ways to approach this. This is just the one that works best for me.
What’s your workflow when building with AI?
The isolated v3.1 workflow is such a practical guardrail—especially when AI starts drifting. Comparing the stable and experimental versions side by side makes iteration feel much safer.
Working on DragSet involves coaching an AI coding assistant through features and bug fixes, then running QA. One failure mode I watch for is a growing pile of small patches without meaningful progress.
An AI engineer should make the path from input to output inspectable. In Tenderseal, I shaped an AI-assisted workflow where users READ source material, ASK focused questions, and DRAFT a response while keeping the underlying material close to the work. I build with approvals around consequential actions, read-back after integrations, and a clear boundary between model output and human judgment. If your AI feature needs a dependable workflow, message me.
Great Design :)