JP | AI Workflow Lab

@thefarfar The jump from humans decide to AI operates probably needs its own proof step. I’d want to see how it handles missing facts and high-risk exceptions before widening access.

For the past several years, I’ve been building AI products for healthcare. One pattern kept repeating itself. Practices would buy an AI scribe. An AI receptionist. An AI chatbot. An AI scheduling assistant. Each tool looked impressive on its own. Yet somehow the organization itself wasn’t meaningfully different. That made me realize something: Healthcare doesn’t have an AI problem. It has an architecture problem. We’ve become very good at adding AI to individual tasks. We’re much less good at redesigning how healthcare organizations actually operate around AI. That’s the idea behind a framework I’ve been developing called Operations-First AI. Its central premise is simple: Healthcare becomes AI-native from the operations up, not from the algorithm down. Clinical AI is incredibly important, but it delivers its greatest value when it’s built on an operational foundation that can actually put those insights into action. This article is the foundation for a series I’ll be publishing over the coming weeks on topics like: The Healthcare AI Maturity Model Why organizations plateau at AI-Assisted Humans Decide. AI Operates. AI Sprawl Why some healthcare organizations will become AI-native much faster than others I’d love to hear where you agree—and where you think I’m wrong. Read the article here: faridfadaie.com/2026/07/13/why… #HealthcareAI #ArtificialIntelligence #HealthTech #Operations #DigitalHealth #HealthcareInnovation

@jamescoder12 Fast replies are useful, but I’d still separate drafting from sending. If price, timing, integrations, or expected results aren’t in the source material, a person should approve it.

The build order if you're starting from scratch this weekend. Saturday morning (2 hours): Set up the Typeform with fields: name, email, company, inquiry message. Create a Make.com account. Build the first automation: form submission → Claude API (draft response + lead score) → send email via Gmail. Test with 3 fake submissions. Fix the prompt until the AI responses sound natural. Saturday afternoon (2 hours): 4. Add the CRM node form data + AI score + enrichment → Airtable or HubSpot. 5. Add the router hot leads get a Slack alert + booking link. Warm leads enter the nurture sequence. 6. Set up Google Calendar Appointment Schedule. Add the link to the AI response template for hot leads. Sunday morning (2 hours): 7. Build the 5-email follow-up sequence in ConvertKit. Write the framework. Let Claude draft the personalized versions at runtime. 8. Build the proposal outline automation meeting booked → Claude generates draft → saves to Notion. Sunday afternoon (1 hour): 9. Test the entire flow end-to-end. Submit a form. Watch the response send. Check the CRM. Check the follow-up. Check the proposal. Total build time: 7 hours across one weekend. Total ongoing maintenance: 30 minutes/week reviewing AI outputs and adjusting prompts.

A marketing consultant was spending 8 hours every week on the same cycle. Lead fills a form. She reads it. She drafts a personalized response. She schedules a call. She adds the contact to her CRM. She writes a proposal outline. She sends a follow-up sequence if they don't reply. 8 hours. Every week. The same steps. Every lead. Her friend a developer who builds AI automation systems for startups watched her work for one afternoon and said: "You just spent 4 hours doing things an AI workflow does in 4 seconds. You don't need to hire someone. You need 6 automations connected to one AI model. They'll run 24/7. They'll respond faster than you can. And they'll never forget a follow-up." He built the system in one weekend. Her 8 hours dropped to 40 minutes the time she spends reviewing and approving what the AI prepared. Same leads. Same quality. Same close rate. 7 hours and 20 minutes back. Every single week. Here's every automation, every tool, and every connection 🧵

@NoemiTitarenco If I need a skill called /find-my-architecture, that's usually a sign my architecture isn't discoverable enough 😂

@jpbuildslab I have .md files in almost every directory that explains what is in the directory & how its used. I just realized maybe those count as "skills"? But it's great to not invoke anything. The agent should be able to know what to invoke.

I only use 3-4 skills. I don't use any memory system. I hate context pollution. Am I the only one?

@Adaonuoha2 That doesn’t sound unserious at all. Sometimes you need to step back to see a better path. What’s the first thing you want to change now?

Even though it looks like I've been unserious about running my business this year The best thing I did for myself was step back and learn marketing and sales Now I can see so many possibilities for scaling a business I once felt wasn't good enough.

@alex_vance We use AI images too, mostly for social visuals and motion graphics where small misses are easier to work around. Your case has almost no room for that because the anatomy has to stay exact. Why Grok as the fallback?

@jpbuildslab Right now it’s falling back to Grok after GPT fails 2 isolated retries (fresh agent, no previous context). Groks image quality is a lower than I need so some of those images are a little off.

Biggest learnings from this so far: 1) GPT is the better image generator for whole images, Grok handles specific edits better (heel drops, wrist flexion, rotation).

@noahiglerSEO Curious what the bad reviews show by service, location and where the job went wrong. There’s probably a second audit hiding in there.

Watch me live audit this large PE-backed home service brand. They have 17,000 Google reviews on ONE Google Business Profile. Surprisingly, they don't rank very well... Here's everything my audit found: --- THE BACKGROUND Radiant has been operating since 1999, got acquired by The Riverside Company in 2021, and runs two major markets, Austin and San Antonio, where they absorbed Schmidt Mechanical. Plumbing and HVAC, roughly 12,000 residential customers a year. Strong bones by any definition. The SEO tells a different story. --- THE MAPS CHECK The written audit below covers the website side, SERP rankings, on-page problems, citations. The maps deserve their own section, so that's where I started. I ran two mobile grid scans across Austin, plumber near me and AC repair near me, because a profile holding 17,000 reviews should be bullying both packs. Plumber near me is the decent foothold. Solid green through the urban core with an average rank around 8.5 across the metro, fading to yellow and orange out in the suburbs. AC repair near me is not too hot. Average rank sits in the mid-teens and they're outside the top 20 across almost half the scan. For a plumbing and HVAC brand in Texas, that's summer leads walking to competitors every single day. Here's what should sting. Reviews are the biggest bottleneck for most companies we audit, and this brand already owns that asset outright. There are companies with double this map visibility on 17x fewer reviews and just better SEO. The reviews aren't the problem, everything around them is. Now the website side, which I spent most of my time auditing. --- 1. INVISIBLE FOR THE MONEY KEYWORDS Searched from Austin: > plumber austin tx: not in the top 30 > austin plumbing company: #19 > ac repair austin: #21 > emergency plumber austin: #24 > water heater repair austin tx: #25 > drain cleaning austin tx: #31 The site ranks for 4,506 keywords nationally and only 24 of them sit at #1. Almost every one of those is branded, radiant plumbing, radiant austin, and so on. Their best non-branded position is #2 for "carbon monoxide furnace leak," an informational blog query that doesn't book jobs. So the brand carries them when people already know the name, and they vanish the moment someone searches for the service instead. THE FIX: a dedicated, deep service page for every money keyword in each market, with internal links and authority pointed at them. A 25-year-old domain with this much trust should hold positions 1 to 3 on these terms, not 19 to 31. --- 2. LOCATION PAGES GOOGLE SEES THROUGH The pages exist, Cedar Park, Round Rock, the usual suburbs, and each one runs 800 to 1,000 words with clean titles and H1s. That's where the effort stopped. No neighborhood specifics, no FAQ sections, no embedded reviews, barely any internal links to related services. They read like one template with the city name swapped in, and that approach stopped fooling Google years ago. THE FIX: the full treatment. Local housing context, area-specific FAQs with the city inside the answers, and real customer proof per suburb. With 17,000 reviews to pull from, filling these pages with local social proof might be the easiest content job in Texas. --- 3. A BLOG THAT PROVES THE POINT Roughly 15 posts total, which is stunning for a company operating since 1999. One of those posts, about house traps, ranks #8 for a 1,300-volume keyword with zero cluster around it and nobody feeding it. That single ranking is the tell. The domain has plenty of authority to rank, it's just being starved. THE FIX: topical clusters around each service line on a real publishing cadence, every post linking into the money pages. A domain like this would respond fast. --- 4. SCHEMA FROM A TEMPLATE Every page has schema, and all of it is generic Service and Organization markup. Missing: > LocalBusiness schema per location, with unique address, phone, and service area > FAQ schema, which is free SERP real estate > Review and AggregateRating schema. Seventeen thousand reviews and not a single star showing in the search results. > areaServed properties on the Service markup THE FIX is an afternoon of dev work that most companies this size never get around to. --- 5. THE DIRECTORY GAP BBB is handled, accredited since 2011. Yelp is neglected to the point that a brand search barely returns a clean profile. For a company doing 12,000 jobs a year that's a real hole, and it costs more now than it did five years ago. Language models lean on Yelp, BBB, and chambers of commerce when they recommend businesses, so a thin profile there means a missing answer when someone asks ChatGPT for a plumber in Austin. --- WHAT THIS ADDS UP TO Positions 19 to 31 on keywords a brand like this should own outright, a map presence coasting on review count alone, and hundreds of leads a month handed to competitors in two of the hottest home service markets in the country. The first 90 days write themselves. LocalBusiness and review schema across both markets, suburb pages rebuilt with real local content, the blog revived into clusters that feed the service pages, and the directory profiles built out for the AI engines that read them. Big brands coast on branded search until a competitor decides to take the unbranded terms.

@GergelyOrosz I wonder how much drift comes from context depth versus repeated compaction. I’d run the same repo task after 0, 1, and 3 compactions, then compare accepted output, missed constraints, and review time. A shorter context could still be worse if it’s been rewritten three times.

Trying to put my finger on how the more context you use in a context window (I called it context depth), over longer runs, the more errors can compound / the agent can drift. Aka if you want more reliablity, use less context + have shorter runs A quick sketch:

@BratDotAI I'd second the pro 🤟, especially if you're already exhausting codex

Finally adding Claude alongside Codex. 👀 For a part-time indie builder, which Claude plan makes the most sense?

Looking back, almost every meaningful improvement in my workflow came from improving the system—not switching models. Better context. Better architecture. Better constraints. Upgrading the model helped. Improving the system changed how I worked.

One thing I didn't expect after months of building with AI: I spend less time thinking about models. I spend more time thinking about systems. Better context produces better agents. Better design systems produce better interfaces. Better architecture produces better software. Better verification produces more reliable automation. The model matters. But it usually isn't the bottleneck.

That’s a pretty wild run. Arvid Kahl’s post got me wondering what held up after review: x.com/arvidkahl/stat…

AI ROI starts after the activity log. For every long agent run, I want a verification ledger: Output → accepted? Errors → caught? Review → minutes? Cleanup → minutes? Limit → consumed? Which number matters most before you’d say the run paid off?

@Fish_wid_a_V I like tying it to the actual constraint. I’d probably keep callbacks and compliance as hard gates, then use GM per supervisor hour to choose between the jobs that clear them.

In our sub contractor model, my constraint resource of choice is my field supervisor - gross margin per supervisor hour is something I take very very seriously & track religiously! It’s possible to get $750GM/hr and not worth it under $500/hr. We currently are executing 400 Supervisor hours/ month …

@FrontierBDesign That’s a useful distinction. I’d still want to know what confirmed cash flow was the root issue—supplier holds, aging receivables, or the timing of deposits against materials?

One of the metrics of a co. I'm working with is "on-time delivery" of their product. Stated goal is 95%. They're around 65% currently. What they thought the problem was: Inefficient labor What the uncovered problem was: Unpaid payables leading to shipping delays of materials What the ACTUAL problem was: Not enough cash flow coming in to pay all bills. It's not labor. It's not a process. It's cash flow. Solve that first.

@Roy_Banwell I’d still want to know what event the new campaign learns from—a booked job, a completed job, or something earlier. That choice could matter as much as the longer form.

First meta campaign. It got very good at giving me cheap leads, but we only sold 2/78 so quality was way off. Created a new campaign optimized for conversions (triggered off new booking webhooks) instead of optimized for form fills. Also made the form longer. Let’s see.

@JPRbiz Do you now model the longest payroll-to-collection gap even when AR comes with the deal? Seems like that timing mismatch is what the working-capital number can hide.

me back in 2020 when I miscalculated working capital in my first commercial plumbing M&A transaction.. ..and got hit with four straight payrolls - and $0 cash collected

@WilsonCompanies The $1M question is useful, but I’d still want to know what’s actually stuck before hiring. What should this person improve in the first 90 days, and what would tell you the hire isn’t working?

Every owner asks, "Who should I hire next?" Better question: "What's the next leadership hire that unlocks another $1M in growth?"

@sweatystartup The first delegation audit I’d run is a two-week decision inventory: label every owner touch as judgment, lookup, coordination, or approval. Rules handle repeatable cases; the owner keeps exceptions and accountability. “I can’t delegate” often hides four different problems.

I know excuse #1: "This is just a hard business. There is no way for me to delegate these important tasks. I'm in a field that is impossible to scale." Bullshit. You know what is hard? Building custom homes. I don't know if there is anything harder. Yet there are three national homebuilders that are publicly traded with more than 1,000 employees and more than $100 million a year of revenue.

If you are consistently working 60+ hour weeks you have one of three problems:

@_is_isaac Schema first, when the agent adds a route, what stops the generated server from accepting an over broad key contract tests, centralized policy middleware, or both? Encryption protects the key at rest, but the authorization boundary still has to live somewhere

How to launch an app in 3 days - without losing control of it: - Define your OpenAPI schema first. - Use Next.js as the proxy between your frontend and backend. - Have an AI agent build the server against that schema. - Generate an API key whenever someone signs up. Benefits: No separate API layer to build and maintain. A clean separation between frontend and backend. No lock-in to an auth vendor or payment provider. You can impersonate any user Disclaimer: Remember to encrypt the api keys in your database!