Cedric Builds

I build AI agents that help Kenyan businesses run on autopilot.
Currently building in public with n8n + Gemini.

Founder of Zana AI

[Cedric Builds @ceeddy492]

@alexmwanzo @Avugwii 💀🔥

[Cedric Builds @ceeddy492]

@KeKirwa @osoroKE Follow back

[Cedric Builds @ceeddy492]

Strict prompts help a lot in the beginning, but I’ve already seen them drift too.
Validating the response shape before booking makes total sense that way a format change fails explicitly instead of creating fake slots.
How are you doing the validation in n8n? Simple JSON schema check or something more robust?

[Cedric Builds @ceeddy492]

This week I finally shipped my first real 24/7 AI booking agent in n8n. It checks Google Calendar live, talks to customers, and books appointments completely on autopilot. But building it was way messier than I expected.

[Cedric Builds @ceeddy492]

word.......100% agree Matt.
Shipping it felt good, but I’m already thinking about week 6 Right now I’m fighting prompt drift and calendar format changes. Added very strict system prompts + tool grounding to keep it stable.Have you found any good patterns that survive after the honeymoon phase?
Also saw the Runner post – looks clean

[Cedric Builds @ceeddy492]

Haha exactly ......herding digital cats is the perfect description

The trickiest part by far was the Google Calendar tool + hallucinations.
Rewrote the tool 4 times because of weird API responses and time zone issues (Kenya +254).

Then the real battle was making the AI stop inventing fake slots.

What’s been your hardest part when building agents?

[Cedric Builds @ceeddy492]

Spot on fr
In testing everything looked perfect. The moment real customers started chatting, the hallucinations and edge cases came out of nowhere.

That’s exactly why I ended up adding very strict system prompts + heavy tool grounding.

How do you guys usually handle post-launch reliability on the tools you build for clients? Any go-to patterns?

Appreciate the insight though!

[Cedric Builds @ceeddy492]

Biggest lessons from the week: Google Calendar API is sneaky as hell (time zone issues + weird formats). Rewrote the tool 4 times.
Hallucinations were the biggest enemy the agent kept confidently inventing fake slots and customer names
A very strict system prompt is the best weapon.

[Cedric Builds @ceeddy492]

@droid254 @grok confirm this claim😆💀

[Cedric Builds @ceeddy492]

Thanks man!

Interesting to hear you guys are seeing similar patterns at Cognitix.

How do you guys handle the hallucination battle on the calendar/booking side? Do you also use very strict system prompts or do you have other tricks?

Would love to learn how you solve it on bigger agency projects.

[Cedric Builds @ceeddy492]

Behind the scenes of the 24/7 AI booking agent I shipped yesterday Everyone sees the clean final workflow… but building it was messy.

The hardest part wasn’t just making it work it was fighting the AI’s hallucinations the whole time.

[Cedric Builds @ceeddy492]

Exactly this 😂

The hardest part isn’t making the agent book appointments, it’s stopping it from confidently inventing fake slots, making up customer names, or confirming things that never happened.

My fix was adding a very strict system prompt that says:
- Only use data from the calendar tool
- If no slots available → say “No slots available”
- Never invent information

Still not 100% bulletproof, but it cut hallucinations dramatically.

How are you handling it on your end?

[Cedric Builds @ceeddy492]

@HelloVyom Let’s connect

[Cedric Builds @ceeddy492]

@dark_coderz @X Les connect

[Cedric Builds @ceeddy492]

@Cosmosfinite Let’s connect

[Cedric Builds @ceeddy492]

@Hardehsoji @wangai_johnn I can

[Cedric Builds @ceeddy492]

@sflorimm Automation ⚡

(AI chatbots for Nairobi businesses — https://t.co/547I2dFu1R)

[Cedric Builds @ceeddy492]

Moral of the story:
Building AI agents isn’t just connecting nodes. Half the work is taming the model so it doesn’t lie to your customers 😂Still early with Zana AI, but these painful lessons are making the tools actually reliable.What’s the wildest hallucination you’ve fought while building agents?#BuildingInPublic #n8n #ZanaAI

[Cedric Builds @ceeddy492]

Then came the real killer ........hallucinations.Even with Google Gemini, the agent kept: Inventing fake available slots
Making up customer names or confirmation details
Confidently giving wrong answers

I went deep into the docs and learned the hard way:
You have to explicitly tell the AI “If you’re not 100% sure, say ‘I don’t know’”
Ground it with direct tool output instead of letting it “think”
Use memory + structured prompts so it doesn’t freestyle