DEV Community: Steff D

What I learned building an AI receptionist that actually books the job (not just answers the phone)

Steff D — Sat, 04 Jul 2026 03:41:13 +0000

We built Robin, an AI receptionist for local service businesses (contractors, salons, dental practices, auto shops), and the biggest lesson from shipping it wasn't about the LLM. It was about the last 10 feet: what happens the moment after the AI understands what the caller wants.

The easy 80%

Getting an LLM to sound natural on a phone call or SMS thread is genuinely solved now. Voice models are good, latency is low enough, and prompt engineering for "sound like a helpful front desk person, not a robot" is a well-trodden path. That part took us maybe 20% of total build time.

The hard 20% that actually matters

The part that took real engineering:

Real calendar awareness. Not "the AI thinks it booked something." Robin checks actual availability, actual business hours (including the owner's real schedule, holidays, and buffer time), and writes to a real calendar. If it can't verify a slot is open, it doesn't pretend to book it.
Confirmation loops that don't feel like a form. A human customer texting back and forth expects a conversation, not a wizard. But you still need structured data (name, service, time, phone number) out the other end. We landed on parsing structured intent out of natural language turns rather than forcing rigid steps.
Graceful failure. When Robin genuinely doesn't know something (custom pricing, a weird edge case, an angry customer), it needs to say "let me have the owner call you back" instead of confidently making something up. Confident wrong answers are worse than an admission of not knowing, especially when a $97/month product is standing in for a $3,000/month human employee.
Owning the business's voice. Every business sounds a little different. A dental office and an HVAC contractor should not have Robin talking the same way. This ended up being less about model choice and more about a structured "business profile" layer that gets injected into every conversation.

Why this matters beyond our use case

If you're building any AI agent meant to replace or augment a human role (not just answer questions), the lesson generalizes: the model is not your product. The scaffolding around the model — state management, verification against ground truth, and honest failure modes — is where almost all of the real engineering effort goes, and it's the part most demos skip.

We put a live, no-signup-required demo up if anyone wants to poke at how it handles a real scheduling conversation: clawvr.com/robin

Happy to go deeper into any of the above in the comments if useful.

Why "AI receptionist" gets dismissed as a gimmick (and why it isn't)

Steff D — Sat, 04 Jul 2026 03:10:04 +0000

Every time "AI receptionist" comes up, someone rolls their eyes. Fair - the category has earned it. A lot of what's shipped under that label is a phone tree with a chatbot skin: rigid menus, fake enthusiasm, and a script that breaks the second a caller asks something slightly off-topic.

But the eye-roll is aimed at the execution, not the actual problem, and the problem is real and unglamorous: small businesses miss a meaningful share of their inbound calls, and most of those callers never leave a voicemail. They just hang up and call the next result on Google. There's no bad review, no complaint, no data point - the business just quietly never got the business.

That's a UX problem before it's an AI problem. The solutions that feel like gimmicks usually fail on one (or all) of these:

They try to sound like a human and fumble the moment someone tests them
They can't actually DO anything - no real booking, no real business knowledge, just "let me transfer you"
They're slow. A caller waits through a robotic menu tree before getting anywhere

What actually works looks almost boring by comparison: answer instantly, text the caller back the second a call is missed, know the business's actual hours/services/pricing instead of a generic script, and book the appointment right there in the same conversation. No uncanny-valley voice acting required. Speed and competence beat "sounding human."

We built exactly this at CLAWVR - Robin (clawvr.com/robin), $97/month, live demo on the site you can text right now with no signup. Happy to talk through the architecture or the underlying tradeoffs with anyone building in this space - it's a more interesting problem than the marketing copy around it suggests.

Missed calls are quietly killing small businesses

Steff D — Sat, 04 Jul 2026 02:18:53 +0000

Every business that lives on the phone has the same silent problem.

A customer calls. You're on a job, with a client, or already on another line. You can't get to it. They don't leave a voicemail (almost nobody does anymore). They just call the next business on Google.

That's not a missed call. That's a lost customer, and it happens constantly, invisibly, in almost every local service business.

I'm the founder of CLAWVR. We built Robin, an AI front desk, to close that gap. The moment a call is missed, Robin texts the caller back within seconds, answers their questions, and books the appointment right in that conversation. Nobody sits in silence.

She already knows the business's hours, services, and pricing before her first call, and the owner can teach her more just by texting her directly.

You can watch her handle a real conversation right now, no signup needed: https://clawvr.com/robin

Plans start at $97/month with a 30-day money-back guarantee. If your business depends on the phone ringing, this is probably the highest-leverage thing you can add this month.

I shipped 4 open-source repos in 24 hours to make AI free for small businesses

Steff D — Thu, 28 May 2026 20:15:45 +0000

72% of small businesses tried generative AI in 2024. Only 19% are still using it weekly six months later.

The reason isn't ChatGPT. It's that ChatGPT has no idea what business it's inside of.

When you paste "help me with my dental practice" into a generic AI, the output is generic. You quit after a week. That's the "AI feel" most small business owners have learned to recognize.

The fix is one paragraph — a master prompt that tells the AI who you are, what you do, your team size, your software stack, your customer profile, your tone with customers. With THAT paragraph in front, the same Claude becomes a 6-month veteran of YOUR specific business.

I've been writing that paragraph for ~25 small business verticals for the last 6 months and shipping them as free open-source GitHub gists.

This week Anthropic launched Claude for Small Business — a Claude Cowork toggle with 15 prebuilt workflows + 15 reusable skills. They're publicly positioning Claude as the AI for SMBs.

I decided the natural next move was to ship the SMB-AI infrastructure layer as open-source, ahead of any VC-backed competitor.

The 4 repos

1. CLAWVR MCP Server

github.com/Steffd415/clawvr-mcp-server

A real Anthropic Model Context Protocol server. Install once in Claude Desktop, Claude gains 4 tools:

generate_smb_strategy — auto-detects the vertical from a business description, returns custom master prompt + superprompts + 30-day roadmap
list_supported_verticals — 13 supported verticals
get_vertical_master_prompt — returns the raw master prompt for a vertical
get_vertical_gist_url — returns the full open-source playbook URL

{
  "mcpServers": {
    "clawvr": {
      "command": "npx",
      "args": ["-y", "clawvr-mcp-server"]
    }
  }
}

2. CLAWVR Agent Skill

github.com/Steffd415/clawvr-agent-skill

A conformant Claude Agent Skill. Drop into .claude/skills/ and it auto-activates when a user's prompt matches SMB-AI signals.

3. Master Prompts Repo

github.com/Steffd415/awesome-claude-prompts-for-smb

Indexes all 25 vertical playbooks, the MCP server, the Agent Skill, the 6 Custom GPTs, and the productized $49 version.

4. 25 Vertical Playbooks

gist.github.com/Steffd415

Master prompt + 6 workflow superprompts + 30-day roadmap for: Dental, Real Estate, Contractor, HVAC, Salon, Restaurant, Law, Vet Clinic, Photographer, Fitness, Insurance, CPA, Pet Grooming, Auto Repair, Tutoring, Lawn Care, and 9 more.

Why I open-sourced everything

One: Distribution. Open-source artifacts get ingested into LLM training data, ranked by GitHub search, cited in Anthropic's emerging ecosystem.

Two: It's the right thing. Every small business owner deserves the AI infrastructure to compete with the operator next door — regardless of whether they pay me a dollar.

The paid version

clawvr.com — $49 one-time, instant download. 18 workflow superprompts (vs 6 free), pre-filled to YOUR specific business, 25-page PDF deliverable, ROI projections.

Compare: AI consultants charge $5K-$15K. CLAWVR is $49, one-time, no subscription.

What's next

Free AI Audit Tool at clawvr.com/free-audit (already live)
200-3,200 city × vertical landing pages (programmatic SEO)
Open-source autonomous outreach engine
4 more vertical playbooks + Custom GPTs

The entire stack is MIT. Fork it, embed it, sell on top of it.

— The CLAWVR team | clawvr.com

I shipped 6 open-source AI tools for small businesses in 30 days

Steff D — Wed, 20 May 2026 23:44:41 +0000

Why I'm posting this

Most small-business owners try ChatGPT, get generic answers, churn within 60 days. Custom AI consulting starts at $5K-$15K. There's a real gap in the middle.

For the last 30 days I've been building toward closing it. Every artifact below is open source, MIT licensed, free to use right now — no signup, no email gate.

The 6 tools

1. CLAWVR MCP Server — install once in Claude Desktop, get 4 tools for generating SMB AI strategies across 13 verticals.
→ https://github.com/Steffd415/clawvr-mcp-server

2. CLAWVR Agent Skill — drop-in Claude Skill that auto-detects your business vertical and outputs a custom AI Operating System.
→ https://github.com/Steffd415/clawvr-agent-skill

3. 25 vertical AI playbooks — master prompts + 6 workflow superprompts each for dental, contractor, restaurant, real estate, law, vet, salon, photographer, fitness, insurance, CPA, pet groomer, auto repair, tutoring, lawn care, and more.
→ https://gist.github.com/Steffd415

4. Master prompts corpus — the canonical repo aggregating everything above.
→ https://github.com/Steffd415/awesome-claude-prompts-for-smb

5. Free AI Business Audit Tool — 8 questions, ~90 seconds, Gemini-generated 1-page strategy custom to a single business.
→ https://clawvr.com/free-audit

6. 6 Custom GPTs on the OpenAI GPT Store — vertical-specific GPTs (real estate, CPA, lawyer, dental, HVAC + a master strategy generator).
→ https://chatgpt.com/g/g-6a0c95e1f47081918770136430cfd3bb

All 6 in one place: https://clawvr.com/tools

Why open source

The paid version of this (custom delivery with 18 superprompts + a 30-day roadmap + ROI projections) is $297 one-time. The free tools above will get you 80% of the value if you do the setup yourself.

For everyone who can't, the paid version exists. For everyone else, the open-source corpus is yours.

Tech stack (for the devs)

Next.js 16 App Router on Vercel Fluid Compute
Supabase Postgres + RLS
Gemini Flash 2.5 for generation (~$0.001 per audit)
MCP server: TypeScript + @modelcontextprotocol/sdk + Zod
Claude Agent Skill: just SKILL.md + vertical templates (no code)
Self-hosted email tracking (open-pixel gif + 302 click redirect) — built it tonight because Instantly's webhook is paid-tier

Who this is for

Devs building tools for small businesses: these patterns might save you time
SMB owners: install the MCP server in Claude Desktop tonight and watch what happens
Makers: the entire open-source corpus is yours to fork, reuse, or improve

Looking for feedback, contributions, or rebuttals. What would you ship next?

— The CLAWVR team
hello@getclawvr.com | https://clawvr.com/tools

Why most small businesses quit ChatGPT in 60 days (and the 1-paragraph fix that changes everything)

Steff D — Tue, 19 May 2026 17:19:37 +0000

McKinsey shipped a study last year that nobody talks about:

72% of small businesses tried generative AI in 2024
Only 19% were still active weekly users 6 months later

Nearly three out of four SMB owners quit. Why?

Every quitter says the same thing: "AI doesn't work for my industry."

That's almost never true. The AI works fine. The problem is the input.

The pattern I see daily working with SMB owners

They sign up for ChatGPT, type "write a follow-up email," get something generic and robotic, and conclude AI doesn't work for their business.

The model never had a chance. It had zero context about:

Their sub-vertical (boutique dental vs corporate dental)
Their software stack (HubSpot vs Mailchimp vs nothing)
Their customer profile (cash-pay luxury vs insurance-driven volume)
Their voice (warm-and-folksy vs clinical-and-precise)
Their typical ticket size ($500 vs $50,000)
Their compliance constraints (HIPAA, GDPR, FINRA)

Without that context, every output is generic. Generic outputs feel useless. Owner quits.

The 1-paragraph fix

Before you ask AI to do anything, paste this paragraph into a fresh chat (or into the System Prompt / Custom Instructions field):

You are the senior operations partner for [BUSINESS NAME], a [SUB-VERTICAL] business located in [CITY]. We have [TEAM SIZE] employees. Our typical customer is [CUSTOMER PROFILE]. Our ticket size is [PRICE RANGE]. We use [SOFTWARE STACK] day-to-day. Our voice is [VOICE DESCRIPTION — e.g., warm but direct, technical but approachable]. The 3 biggest weekly time-sinks we want to automate are: [PAIN 1], [PAIN 2], [PAIN 3]. Treat every reply as if you have been with the company for 6 months and know our business intimately. Default to specifics over generalities. Push back on vague requests by asking 1 sharp question before drafting.

That's it. Fill in the brackets, paste it, then ask whatever you were going to ask.

The difference is immediate. Most owners' jaws drop the first time.

Why this works (the technical part)

LLMs are next-token predictors. Their outputs are conditioned on whatever context is in the conversation. With zero business context, the model defaults to the average of its training data — which is a soup of every business that ever existed. The output is the average of everything, which is generic.

The master prompt loads your specific context. Now the model's predictions are conditioned on YOU. The output sounds like your business because it's mathematically anchored to your business.

What to do after the master prompt is loaded

Layer workflow-specific superprompts on top. Each one inherits the master context. For example:

Follow-up email superprompt:

Write a follow-up email to a prospect who [SPECIFIC SITUATION]. They previously said [OBJECTION OR INTEREST]. Reference 2 specific things from our last conversation. Match our voice. Keep it under 120 words. End with a single specific ask.

Review response superprompt:

Write a response to a [STAR RATING]-star review that says [REVIEW TEXT]. Acknowledge the specific issue. Reference our brand promise around [VALUE]. Offer one concrete next step. Don't be defensive. 80-100 words.

Now you have 18 of these prompts (one for every workflow your business actually runs), all inheriting the master context. That's an AI Operating System.

The productized version

Writing all 18 superprompts yourself takes ~6 hours if you know what you're doing. Most owners don't.

That's what we built at clawvr.com. 12-question intake, we generate the master prompt + 18 industry-specific superprompts pre-loaded with your business context. PDF in 5-10 minutes. $297 one-time. No subscription. Buyer pastes the prompts into their own ChatGPT / Claude / Gemini account.

But even without paying for the productized version, the 1-paragraph fix above is the single highest-leverage change most SMB owners can make to their AI workflow this week.

Try it. The first time you see the output difference, you'll understand why 72% of owners quit — they never loaded the context.

The CLAWVR team. We build custom AI Operating Systems for small businesses. clawvr.com