Harsh Makadia (@MakadiaHarsh) header
Harsh Makadia (@MakadiaHarsh) profile picture

I build custom AI + software systems that automate operations, save time and turn more opportunities into revenue. Founder https://nitter.cf/t.co/FHMNUJ3oB0, https://nitter.cf/t.co/KnKHlRawx3

Book an AI Systems Review
Joined April 2012
Photo by Harsh Makadia (@MakadiaHarsh)Photo by Harsh Makadia (@MakadiaHarsh)Photo by Harsh Makadia (@MakadiaHarsh)Photo by Harsh Makadia (@MakadiaHarsh)Photo by Harsh Makadia (@MakadiaHarsh)Photo by Harsh Makadia (@MakadiaHarsh)Photo by Harsh Makadia (@MakadiaHarsh)Photo by Harsh Makadia (@MakadiaHarsh)Photo by Harsh Makadia (@MakadiaHarsh)
what businesses ask me to automate vs what i end up building first: 1/ asked: "connect everything to our old system" 
built: a daily export from the old system into one sheet, so we can see what's actually in there 2/ asked: "an AI that answers every customer email" 
built: a sorter that tags emails into 5 piles, and a person still replies 3/ asked: "a dashboard for the whole team" 
built: one message every morning with the 3 numbers the owner already asks about 4/ asked: "replace our 15 year old software" 
built: nothing. we work around it. it's ugly, and it's the only thing everyone knows how to use. the first build is the one they judge every later build by. keep it small enough to be right.
6
2
6
1,948
Harsh Makadia retweeted
The life cycle of almost every vibe-coded MVP i get sent: 1. "It works." on the founder's laptop, logged in as the founder 2. "A friend tried it and loved it." the friend clicked two buttons and closed the tab 3. "We got our first signups." Most never opened it a second time 4. "Can you look at why it's so slow?" every page loads every row in the database "Should we just rebuild it?" No. Go back to the step everyone skipped: One user, one job, written down before the first prompt. The AI will build anything you ask for. It just has no idea who it's for.
20
1
18
7,816
The life cycle of almost every vibe-coded MVP i get sent: 1. "It works." on the founder's laptop, logged in as the founder 2. "A friend tried it and loved it." the friend clicked two buttons and closed the tab 3. "We got our first signups." Most never opened it a second time 4. "Can you look at why it's so slow?" every page loads every row in the database "Should we just rebuild it?" No. Go back to the step everyone skipped: One user, one job, written down before the first prompt. The AI will build anything you ask for. It just has no idea who it's for.
20
1
18
7,816
Harsh Makadia retweeted
if a builder offers you a one-hour "ai audit" and a pdf by the end of the call, i'd walk. an hour with the founder is enough to hear what they think is slow. it's not enough to see what the team actually does when a job comes in, or when two people keep a different version of the same record. the useful audits start with watching one live job move through the business. intake, the handoff, the invoice. then they talk about tools. if the first deliverable is a slide deck and not a walkthrough of a real process, you're buying a document, not a diagnosis. the question i'd ask before hiring anyone: who on the team are you going to sit with? if the answer is just the founder, keep looking.
7
2
10
3,487
Using AI feels cheaper with treg.to🔥
2
1
3,106
if a builder offers you a one-hour "ai audit" and a pdf by the end of the call, i'd walk. an hour with the founder is enough to hear what they think is slow. it's not enough to see what the team actually does when a job comes in, or when two people keep a different version of the same record. the useful audits start with watching one live job move through the business. intake, the handoff, the invoice. then they talk about tools. if the first deliverable is a slide deck and not a walkthrough of a real process, you're buying a document, not a diagnosis. the question i'd ask before hiring anyone: who on the team are you going to sit with? if the answer is just the founder, keep looking.
7
2
10
3,487
a taxi dispatch company launched an AI receptionist yesterday. the CEO's line: "taxi operators lose fares every time the phone rings out at 2 a.m. or on a busy saturday night." everyone builds for 2am. saturday night is the one i'd fix first. at 2am the office is closed and the caller half expects it. at 9pm on a saturday the office is open, both lines are busy, and the third caller listens to ringing until they hang up and open uber. what i'd set up before any AI: 1. the third call rolls to something that answers, even if it only takes the pickup address and time. 2. regulars get recognized by number so they never repeat their address. 3. anything it can't book goes straight to whoever is dispatching that night, by name. most of the fares you lose on a busy night already called you.
12
10
3,581
Harsh Makadia retweeted
Firecrawl is now FREE! Run it locally on your own hardware and use as much of it as you like.
Introducing Firecrawl Local. We rebuilt the Firecrawl repo so agents can scrape locally. - 37x faster rust based scrape primitive (vs self-hosted) - 230x less memory (11 MB idle) - 100% open source Compatible out of the box with all of our MCP & SDKs. github.com/firecrawl/firecra…
34
84
5
2,487
258,816
most service businesses track one speed: how fast someone calls the lead back. there are 3, and the other two lose more jobs: 1. lead to first reply: the call came in at 9pm and nobody answered until 8am. 2. reply to estimate: you answered in 5 minutes, then offered thursday. 3. yes to job date: they approved the quote, waited 3 weeks, and quietly went with someone else. every gap gives the customer time to get distracted, call around, or cancel. when i build for a contractor, i time all three on last month's leads first. the slowest one is the build, and it's rarely the callback.
9
10
3,224
a roofer told me he never misses calls. i asked to see his phone log from last week. found 4 missed calls at random time of day and night. no voicemails. no callbacks. just numbers. he said "if it was important they'd leave a message." we called two of them back together. one had already booked someone else. the other said "oh, i thought you were closed." a missed call never shows up on any report. nobody writes down the job that didn't happen. the first thing we set up wasn't AI. a text back inside a minute, and a short list he reads every evening before he goes home. stupid simple trick works all the time.
9
7
4,413
Harsh Makadia retweeted
10 coding agents don’t give you 10 developers if you can only review one change at a time. before adding more agents, split the work into: INDEPENDENT can ship without another task finishing. DEPENDENT needs a decision or change from somewhere else. SHARED touches the same files, schema or public interface. parallelize the first group. sequence the second. assign one owner to the third. then limit work in progress to what you can actually review. measure merged, tested changes. not how many agents are running.
10
1
13
5,414
A practical way to prioritize feature requests: Put each request into one of four buckets: Can't start → onboarding or setup is blocking the first useful action. Can't finish → the core workflow breaks before the outcome. Can't trust → the result needs checking or correction. Wants it faster → the job works, but takes too much effort. Describe the last time it happened. The requested feature is a proposed solution. The bucket helps you find the problem.
8
1
10
4,271
5 signs nobody is using the automation you handed over 30 days ago: 1. someone still updates the old spreadsheet "just in case" 2. the alerts channel is muted 3. people message you for the report instead of opening it 4. the first weird case got handled by hand and nobody told you 5. the person who asked for it changed roles and nobody owns it now day 30 is when you find out if you built a system or a demo they were polite about.
1
1
1
4,720
Harsh Makadia retweeted
a business paid an "AI automation guy" $5,000 then they called me to fix it. their customers had stopped getting pickup emails. nothing crashed. no errors. someone had renamed one word in a dropdown. here's what happened: the automation sent a pickup email whenever an order's status said "ready." one day, someone on the team renamed "ready" to "packed." made sense. that's what the team actually called it. but the automation was checking for the word "ready." no "ready," no email. $5,000 broken by a dropdown edit. the fix wasn't more code. it was separating 3 things that should never have been one: 1. the internal state: ready_for_pickup 2. the visible label: "packed" 3. tags: "priority", "fragile" the automation listens for the state. the team can rename the label whenever they want. tags add context without touching either. before you make any dropdown editable, ask one question: is this just a label, or does it decide what happens next?
4
1
9
5,529
a business paid an "AI automation guy" $5,000 then they called me to fix it. their customers had stopped getting pickup emails. nothing crashed. no errors. someone had renamed one word in a dropdown. here's what happened: the automation sent a pickup email whenever an order's status said "ready." one day, someone on the team renamed "ready" to "packed." made sense. that's what the team actually called it. but the automation was checking for the word "ready." no "ready," no email. $5,000 broken by a dropdown edit. the fix wasn't more code. it was separating 3 things that should never have been one: 1. the internal state: ready_for_pickup 2. the visible label: "packed" 3. tags: "priority", "fragile" the automation listens for the state. the team can rename the label whenever they want. tags add context without touching either. before you make any dropdown editable, ask one question: is this just a label, or does it decide what happens next?
4
1
9
5,529
4 builds i turn down now: 1. "automate everything"
 nobody knows what everything means. it ends with a 40 page scope and nobody happy. 2. processes that change every week
you're not automating a process, you're chasing one. 3. "make it work like big billion company system" 
they had 50 engineers and 3 years. you have a budget and a deadline. 4. anything nobody owns after launch
if no one checks it, it breaks quietly and i get the blame. saying no to these made the yes projects better. less rework, faster launches, happier clients.
2
1
4,291
10 coding agents don’t give you 10 developers if you can only review one change at a time. before adding more agents, split the work into: INDEPENDENT can ship without another task finishing. DEPENDENT needs a decision or change from somewhere else. SHARED touches the same files, schema or public interface. parallelize the first group. sequence the second. assign one owner to the third. then limit work in progress to what you can actually review. measure merged, tested changes. not how many agents are running.
10
1
13
5,414
a useful structure for a 90-second automation demo: 0–15 seconds: show the messy input. 15–30 seconds: show what someone currently has to do with it. 30–60 seconds: run one example through the workflow. 60–75 seconds: show the finished output and where the team receives it. 75–90 seconds: show one limitation and what happens when it occurs. example: an emailed supplier price sheet goes in. changed prices come out, with the previous price, new price and affected products side by side. the buyer reviews the changes before anything is published. keep the tool diagram for questions. the buyer should understand the job you’re taking off their plate before they learn which tools you used.
9
12
4,128
a simple way to decide what to automate first: score each task from 1–5 on: 1. frequency: how often does it happen? 2. friction: how much time does it take? 3. rework: how often do errors send it back? 4. clarity: can someone explain the rules? multiply frequency × friction. use rework to break ties. then check clarity before you build. if nobody agrees on the rules, document the process first. example: a 5-minute task done 100 times a week costs over 8 hours. a 2-hour task done once a month may feel painful, but it’s rarely the best first target.
7
8
4,117
a founder wanted an admin dashboard in his MVP. charts, filters, user list, export button. i asked one question: 
"what would you open it to check every morning?" he thought for a second. 
"how many people signed up yesterday." so instead of the dashboard, i set up one email. 
7am. one line. "yesterday: 4 signups." 3 months later he still hasn't asked for the dashboard. he didn't want a dashboard. he wanted one number before his coffee.
7
9
4,388