- Startup Spells 🪄
- Posts
- AI Feedback Loops: The Secret Behind Spiral's Humanized Writing
AI Feedback Loops: The Secret Behind Spiral's Humanized Writing
PLUS: We Spent >$1 Million Sponsoring Creators. Here’s What We Learned About B2B Influencer Marketing
Startup Spells 🪄
November 19, 2025
AI Feedback Loops: The Secret Behind Spiral's Humanized Writing
AI feedback loops are the difference between a generic wrapper and a tool that writes like a human. If you are building a bot or an agent, you’ve likely hit the ceiling of what "better prompting" can achieve.
When ChatGPT asks "How would you rate this response?" or Claude shows thumbs up and down buttons, they're not just being polite.
AI Feedback Loops - ChatGPT feedback rating interface
They're building the dataset that makes tomorrow's model smarter than today's.
The companies building the models know this. That is why every major interface forces you to become a judge.
Gemini feedback rating interface
They don't just want you to chat; they want you to rate. Positive feedback tells the model "do more of this." Negative feedback says "stop doing that." This is how base models get fine-tuned.
Why AI feedback loops separate good models from exceptional ones
Training an AI model costs millions. But the real expense isn't compute. It's knowing what good looks like.
You can feed a model billions of words. It'll learn patterns. But patterns don't equal judgment or taste. That requires humans telling the AI: "This answer nailed it. That one missed."
This isn't new. It's reinforcement learning from human feedback, the technique that made ChatGPT conversational. Without AI feedback loops, you get a calculator that speaks English. With them, you get something that feels intelligent.
The feedback interface matters more than you think
Look at how the major players collect feedback. Claude and Gemini use thumbs down (negative feedback) with detailed follow-ups and their models deliver better outputs.
Claude negative feedback form with issue types
These aren't cosmetic choices. They're data collection strategies. Simple thumbs up and thumbs down give quick sentiment signals. Detailed forms explain why something failed. Both feed the training loop.
Open AI used Kenyan workers for <$2 per hour to train their feedback loop.
Claude positive feedback submission details form
The trick is to make feedback effortless. One click to rate, optional depth for users who care. This maximizes response rates while capturing nuance from power users.
How Spiral proves feedback loops create better outputs
Spiral from Every.to writes like a human because it trained on human judgments of good writing. Not just examples. Judgments.
The secret is DSPy and GEPA, frameworks that treat prompts like code you can optimize. Feed them examples of great outputs and terrible ones. Add a judge that scores each attempt. The system learns what works.
Mike Taylor explained this in 2 essential videos: his GEPA evaluator-optimizer session and the DSPy primer with Every.to.
The gist is you don't manually tune prompts anymore. You build systems that tune themselves based on feedback.
Every.to built a human judge model. It scores writing quality. Spiral generates text, the judge rates it, and the system iterates until output quality soars. This is fine-tuning in production.
Why your project needs AI feedback loops today
You're building something with AI. A chatbot, a content generator, a research tool. Whatever. Without feedback, you're flying blind.
Grok AI interface showing feedback options
Start basic. Add thumbs up and down. Track which responses users approve. Review failures weekly. This alone separates amateur projects from professional ones.
Then layer in detail. Ask what specifically was wrong. Too verbose? Factually incorrect? Off-brand? Each signal refines your prompts or training data.
The economics work in your favor. Collecting feedback costs almost nothing. Improving your model based on that feedback compounds forever. Month 1, you ship version 1.0. By month 6, you're running circles around competitors who never asked users what worked.
The fine-tuning advantage nobody talks about
Generic models are impressive. Custom models trained on your feedback are unstoppable.
When you collect feedback, you're building a proprietary dataset. Your users tell you what good looks like in your domain. Feed that back into training, and you create something competitors can't replicate by throwing more GPUs at the problem.
This is how niche AI tools beat giants. They know their vertical better because their users taught them. The giant model knows everything shallowly.
This works on small scale too. Give good and bad examples to your prompt and watch your output automatically improve.
If you are just prompting, you are guessing. If you are implementing AI feedback loops, you are engineering.
Top Tweets of the day
1/
You should definitely do this for D2C brands, but even SaaS companies can create fake co-workers who are good on camera, because most of them are nerds who are terrible on camera.
2/
Partner up with celebrities to grow the fastest brands because most celebs are broke but have the fame and familiarity to make up for it.
You don't need full-blown partnership either where you give away your equity but you can just partner up with a popular character and get the mindshare easily. Look at Acquire x Silicon Valley's Russ Hanneman collab.
3/
Above all, take action on imperfect plan. The biggest mistake is not taking action at all.
Rabbit Holes
What’d ya think of today’s newsletter? Hit ‘reply’ and let me know.
Do me a favor and share it in your company's Slack #marketing channel.
First time? Subscribe.
Follow me on X.
Reply