---
title: 72 Hours Building AI Client Tools - 1AM Doubt Spiral
canonical: "https://filipsardi.com/building-ai-client-tools/"
pubDate: "2026-02-10T00:00:00+00:00"
author: Filip Sardi
description: "What worked, what broke, what hallucinated - and why I'm looking for 12 founders to test the diagnostic before it goes public."
---

For a moment during the sprint, I felt it again.

That old familiar energy from the launch days - the creative hyper-focus where hours disappear and you can see the whole thing coming together in front of you, like a painting that's finally showing you what it wants to be.

And I thought: maybe I do miss this.

But by the end of the 72 hours, I realized what I actually missed was the building - the nerding out, the problem solving, the satisfaction of getting a tool to do something useful for real clients.

**What I don't miss is launch hype, deadlines and "all or nothing" stakes.**

The sprint wasn't supposed to have that energy, but I guess old patterns die hard:

I stayed up way past the plan because I was "in the flow." I tested "just one more iteration" when I had already said enough two iterations ago. I switched between context tasks too often, and kept forgetting that brain work is more intense and tiring than physical work - you need more time to reset, not less.

All of which I'm sharing because this letter isn't a highlight reel, but what really happened.

## And the reason behind why I built any of these agents in the first place?

Every consultant and mentor I know wants the same thing - deep, personalized guidance that makes each client feel like you see their full picture.

The problem is that doing it 100% manually takes so long that at some point the math breaks.

I stopped offering my diagnostic VIP sessions for exactly that reason, even though people were willing to pay significantly more for them.

## 1/ How the sprint started

Last week I wrote [about having fun (Letter 015)](https://clientflow.substack.com/p/015-are-you-having-fun) and the question from my friend Phil that planted the seed for this whole thing:

The unfiltered answer was: building AI tools for client delivery and documenting the messy process in real time.

So I decided to have fun, cleared the schedule and committed to 72 hours.

Missed [my main FlowOS Sprint progress page](https://clientflow.substack.com/p/flowoslab)?

**Here's the short version of the plan - three tools, each solving a different part of my retention diagnostic workflow:**

**BusinessDNA** - captures the full business picture (offers, model, pricing, audience, resources) through a dynamic conversation, so I always have context before we ever talk. This is the onboarding layer (and Activate milestone).

**FlowMapper** - turns my 30-40 question diagnostic framework into a guided conversation that surfaces where clients lose momentum, where the delivery breaks down, and where upgrade opportunities are hiding. Originally planned as two separate tools (Map Your Flow + Upgrade Paths), but that changed fast - more on that in a minute.

**Client Flow Pulse** - weekly check-in ritual that show me who's actually moving and who's drifting, in real time while we're working together. Think of it as the client momentum intelligence layer.

That was the plan. Here's what actually happened.

## 2/ Day 1: The warmup (and the prep I forgot)

The night before the sprint I was so excited I couldn't sleep - which is a great sign for the work and a terrible sign for day one energy.

Next morning, ready to go, and I immediately realized I'd skipped an important preparation step.

Since I wanted to share live updates on socials, I should have prepped my content templates and checked my recording setup the night before. Instead of building, my first morning was spent writing intro posts and notes.

Lesson before the sprint even started: **if you're going to build in public, prepare the public part before you start building.**

Once I finally got into it, BusinessDNA came together fast.

_(BusinessDNA starting point…)_

The core challenge was getting the agent to not overcomplicate the intake with a flood of follow-up questions, and to pull as much context as it can from links the client provides first — their website, their sales page, their existing materials — before asking anything.

By end of day one: version 0.4, operational, usable. A good warmup.

Here's a demo example based on my business - this is what client sees in the end, and I get more in depth business profile as their mentor.

But it was just that. A warmup.

## 3/ Day 2: Where the plan broke (and got better)

The real challenge started on day two, and it was a completely different beast.

**Taking my 30-40 question diagnostic and turning it into a meaningful conversation, not just a chat version of a long form, required a very different kind of thinking.**

The agent needed to ask follow-up questions when someone gives a vague or incomplete answer, but without making them feel interrogated.

It needed to know when "we do a welcome email" means "we have a solid activation sequence" versus "we send one generic message and hope for the best."

That kind of judgment is what makes a diagnostic valuable, and teaching it to a tool is harder than it looks.

On top of that, the original plan was to build Map Your Flow and Upgrade Paths as two separate agents. But once I started building, it became obvious that splitting them was the wrong architecture.

You can't map upgrade opportunities without seeing the full client flow first, and you can't see the real client flow without understanding where the next steps are missing.

So I scrapped the two-agent idea, merged everything into FlowMapper, and rewrote the workflow logic from scratch.

At that point it was clear: there's no way the third tool (Pulse) is happening during this sprint.

**First big "fail", at least in my mind, because I had been overconfident enough to think about building** _**four**_ **tools before I even had two working properly.**

Classic old Filip from 2019.

But here's the thing that only became obvious in hindsight: consolidating Map Your Flow and Upgrade Paths into one tool wasn't a compromise.

It was the right call. The diagnostic is stronger because of it, the user experience is smoother, and the insights are more connected.

> _Sometimes the best design decision is killing the plan that looked good on paper._

## 4/ "Nobody is even commenting"

It was Thursday evening, somewhere around 1AM, and this was the first time the process stopped being fun.

I had been deep in the build for nearly two days, sharing progress as I went, and the doubt crept in with a very specific voice:

> _"Damn, nobody is even commenting or asking questions or giving me any type of feedback. Why am I even doing this?"_

**And for a moment I believed my tired mind's bullshit.**

Then I remembered two things.

One - people are interested and they are following along. They're just not on the hyper-focused level that I am at 1AM on a Thursday with five browser tabs of agent configurations open.

Two - the whole point of this sprint was the question [Phil Powis ❤️⚡️](https://open.substack.com/users/181219008-phil-powis?utm_source=mentions) asked. The thing that has juice for _me_, regardless of who's watching.

It was interesting to catch that pattern in real time, because it's the same one that stops so many founders from starting or finishing things:

**Waiting for public validation before they trust their own momentum.** That quiet need for someone to say "this is good, keep going" before you give yourself permission to keep going.

Also, it was obvious it was time for bed if I wanted day three to be anything other than a disaster.

## 5/ Day 3: When the tool started hallucinating

By 3PM on Friday I was having fun again.

FlowMapper v0.1 was deployed, and I did a little happy dance because the core conversation flow was actually working.

Then I started testing it properly.

And even with all my instructions, constraints, and a full knowledge library the tool kept doing two things that drove me nuts.

**First, it kept jumping into generic retention advice:** "you should improve your onboarding" - even though its only job was to diagnose, not prescribe.

It wanted to be helpful so badly that it skipped past the questions and went straight to answers nobody asked for.

**Second, when a test answer was vague, instead of asking a follow-up question, it would fill in the blanks on its own.** It assumed it knew what you meant, which is exactly the kind of "helpful" that makes a diagnostic useless.

And this is where one of the biggest lessons of the sprint showed up.

### Constraints are more important than instructions.

You can give an AI agent the most beautiful instructions in the world: "ask insightful follow-up questions, adapt to the client's situation, be conversational", and it will still wander off if you don't clearly define what it should not do.

Don't give advice.
Don't assume.
Don't skip the follow-up when the answer is shorter than two sentences.
Don't try to fill in context you don't have.

**The "don't" list ended up being the thing that made the tool actually useful,** because it forced the conversation to stay in diagnostic mode instead of drifting into generic coach-bot territory.

**The other big realization was about knowledge** - not my knowledge, but how the agent stores and accesses what it learns. Saving client entries isn't enough. If you want a tool to have real context about what someone shared three questions ago, you have to think carefully about how that information flows through the system and stays available in the right moments.

This was the "boring" part - testing iteration after iteration, tweaking the programming until the tool's questions actually resonated with how I think.

**If I saw an incomplete answer on my intake Google form, what follow-up would I send? And how many follow-ups are enough before someone starts feeling interrogated?**

Here's one simple example where agent proactively started doing better follow up after few iterations…

**Another important lesson:** **calibration took more time than the actual build.**

At one point I also had to do a hard reset because the AI started giving me praise for my "brilliant thinking" and overcomplicating the entire build process.

I stepped back, applied what I call "simplify to amplify", stripped out the clever parts, kept only what works, and the next version was immediately better.

Not everything has to be perfect in the first live version. That's what iterations and real client feedback are for.

## 6/ How I knew it was time to end the sprint

Apart from the obvious exhaustion, the Universe gave me a final exam.

Everything was ready to record what I believed was the last iteration demo. I hit record.

Wifi started lagging.
My ring light fell over.
Some mystery insect decided my face was the place to be.
And the grand finale - once I finally got through the whole thing, I checked the file and there was no audio.

### I had two choices: go ballistic, or laugh at myself and declare the sprint officially over.

I laughed it off, closed the laptop, and called it done - and have the final 10:50PM bloopers video to prove it.

## 7/ The diagnostic is open - 12 founding spots

Here's what the sprint was really building toward.

The goal was never to have shiny tools - it was to streamline my diagnostic workflow so more founders can get real clarity on where their retention revenue is leaking, without me spending hours on prep before we even talk.

The tools handle the first layer: mapping your business, surfacing your real client journey, flagging where momentum breaks and where upgrade opportunities are sitting untouched.

So, when we get on the call, we skip the context gathering and go straight to decisions, strategy and planning.

### Retention Multiplier Gameplan™

That's why I'm looking for **12 founders** who want to run their program or service through the full diagnostic and walk away with a concrete plan.

**Here's how it works:**

**Step 1 - Guided FlowOS Diagnostic** (on your own time, 20-30 minutes)

You go through BusinessDNA and FlowMapper before we talk. By the time you're done, you have a first draft of what's actually happening in your client journey: where clients lose momentum, renewals slowly disappear and where there's upgrade potential with no structure around it.

**Step 2 - 90-Minute Retention Multiplier Call**

We walk through your diagnostic together and I show you exactly where is your retention revenue potential. We pinpoint the 2-3 specific moments where momentum dies or people drift, and decide the safest retention plays you can install in the next 60-90 days (without adding more calls, or more complexity to your delivery).

**Step 3 - Your 60-90 Day Gameplan**

After the call you get a written plan with clear priorities in order (not a random list of ideas), simple changes your ops person or VA can run, and a shortlist of retention wins to go after first so your next launch feels safer even if fewer people buy.

No funnels to rebuild oe extra curriculum to design. Just a clearer path for the clients you already worked hard to enroll.

**Founding price: €444** (standard price is €750)

12 spots. Once they're gone, the price goes up.

If you've been feeling that familiar pressure after every launch:

- the revenue resets
- the starting over feeling
- the "I need more new clients" loop

…this is the fastest way to see what's actually fixable in your delivery without changing your offer.

**Reply or send me a DM and I'll send you the details.**

## 8/ What the sprint actually confirmed

A lot of people are writing about AI agents right now.

Theorizing about what to build, sharing frameworks for frameworks, debating which platform is best.

**Theory matters, but the real lessons, like constraints being more important than instructions, or that your value lives in the questions you ask and the follow-ups you choose, only show up when you're in the mess of it.**

And here's the part I want to leave you with - AI in client delivery is here to stay.

Not to prescribe what your clients should do, but to surface the signals you'd normally miss and help you do the kind of deep, personalized work you used to only be able to do with three or four clients - with ten, fifteen, or twenty.

### Not "scale" in the way the internet usually means it.

More like: the depth you already bring to your best client relationships, available to more people without you burning out.

So if you've been watching the AI conversation from the sidelines and wondering where to start - stop thinking about content tools and marketing automations for a moment. Ask yourself:

> _What part of your delivery is currently held together by your memory, your energy, and your hope that you'll notice problems before they become cancellations?_

That's probably where your first tool should live.

The sprint is done, but FlowOS Lab is just getting started.

More builds coming soon and I'll make sure to keep showing you the mess alongside the wins.

_- Filip "let's keep building" Sardi_