GHL Conversation AI Bots . Agency OS

What it does

Produces three prompt blocks you paste straight into GoHighLevel Conversation AI: Personality (who the bot is), Intent (what it does and who it disqualifies), and Additional Instructions (objection scripts, boundaries, fallbacks). Around them it produces Workflow AI prompts for Flow Builder branching, five test conversations to run before launch, the GHL setup steps for where each piece goes, and a post-launch monitoring checklist. Five bot types are supported: Lead Qualifier, Appointment Booker, General Support, Re-engagement and Full Pipeline. In your own workspace it builds for your agency's pipeline; in a client workspace it builds for the client's business.

Say this to start

This skill has no button. You start it by saying what you want. Any of these will do it:

/create-bot

> build a lead qualification bot

> conversation ai prompt

> ghl bot

> appointment booking bot

> re-engagement bot

When to reach for it

lead form enquiries sit unanswered for hours
you want inbound leads qualified before they reach your calendar
a client wants a bot booking appointments on their calendar
a cold database needs re-engaging without you texting each contact
the current bot rambles or asks three questions at once
building a phone Voice AI script rather than a text bot

When NOT to use it

If you actually want	Use this instead
the human sales conversation the bot hands off to, on inbound calls	`sales-conversion-engine`
your agency's own outbound sales conversations	`agency-sales-system`
the GHL pages, pipelines and drip sequences the bot sits inside	`funnel-architect`
the Meta lead-form intake that feeds the bot	`meta-lead-system`
the client delivery plan that installs a bot as one line item	`client-delivery-engine`

Before you start

What you need	Why
`epistemology.md` in the workspace	the skill refuses to ship a bot without it and sends you to `/deepen-profile` first. It is injected into the system prompt verbatim, turning your invisible assumptions into the bot's premise checks and your diagnostic process into its question flow. Without it the bot parrots.	Required
`policy.md` with a `never_touch` list	that list becomes the bot's hard refusal list, the topics it must decline cleanly. Without it the bot has no defined boundaries in front of real leads.	Required
A GoHighLevel account you can log into and edit Conversation AI on	this skill writes configuration, not a deployment. Somebody has to paste each prompt section into GHL and wire the workflow, so in client mode you need edit access on the client's account before you promise a launch date.	Required
Business hours, timezone, and a named human the bot escalates to	after-hours behaviour and every escalation trigger depend on these three, and the skill asks for each if the workspace does not answer it	Required
A calendar link, for booking bots	an Appointment Booker or Full Pipeline bot has nothing to book against without it	Optional
`ica-output.md` and `offer-output.md` for the product the bot serves	these pre-populate qualifying questions, disqualifying criteria, avatar language, pricing and CTA framing. Missing them means answering all of that by hand in the interview.	Optional
`voice-profile.md`	calibrates the bot's tone to the brand. Without it the bot defaults to professional and friendly, which is safe and generic.	Optional
No MCP connection is needed for this skill	the output is text you paste in yourself, so nothing is pushed into GoHighLevel automatically and no API key is involved	Optional
If the workspace holds multiple products, know which one the bot is for	the skill asks before resolving which ICA, offer and funnel data to read	Optional

How it runs

Mode detection and gateDetects whether you are in _default (a bot for your agency's own pipeline) or a client--* workspace (a bot for the client's business) and says which. Checks for epistemology.md and policy.md, and stops with a prompt to run /deepen-profile if epistemology is missing.
Bot type selectionPresents the five bot types with purpose, best fit and complexity, and recommends one based on what it found: an Appointment Booker if your offer is high-ticket, a Lead Qualifier if you are running ads, Re-engagement if there is a database and no ad spend. It then stops and waits for your choice.
Business context gatheringAuto-loads business name, services, ideal customer, pricing, calendar link, hours, timezone, tone, disqualifying criteria, common objections and handover contact from workspace files, then asks only for the gaps. It reads back a one-paragraph summary of what it is about to build and waits for your confirmation.
Prompt generationWrites the three sections using the CRISP framework and hands them back as marked blocks ready to copy: === PERSONALITY ===, === INTENT ===, === ADDITIONAL INSTRUCTIONS ===. Formatting follows what GHL parses well: bullets throughout, section delimiters, merge fields such as {{contact.first_name}}, one question at a time, and a 20-word ceiling on every bot reply. It stops for your revisions.
Flow Builder configurationFor Appointment Booker and Full Pipeline bots it generates the supporting GHL Workflow AI prompts: question-and-branch patterns, qualification gate logic, data collection flows and conditional routing. Each AI step is written to output a single clean word that GHL can branch on by exact text match. It stops again for confirmation.
Test conversationsProduces at least five scripted scenarios to run before launch: the ideal qualified lead through to booking, a disqualified lead exiting gracefully, an objection (price, timing, spouse, competitor), an edge case such as gibberish or three questions at once, and a re-engagement scenario where relevant.
Deployment guideGives the setup steps in order: where each prompt section is pasted, workflow trigger configuration, calendar connection, pipeline stage mapping and notification setup for human handover. Adds a common-fixes table for the issues that show up most often.
Monitoring and loggingHands you the post-launch checklist: review the first 20 conversations manually, check qualification accuracy, monitor booking completion, track how often the bot hands to a human, and identify failure points to feed back into the prompts. The bot configuration is appended to assets.md and feedback logged to learnings.md.

What you get

Three copy-paste prompt blocks marked === PERSONALITY ===, === INTENT === and === ADDITIONAL INSTRUCTIONS ===
Objection handling scripts, re-engagement sequences and fallback responses for when the bot does not understand
Qualifying criteria, disqualifying criteria and explicit escalation triggers
GHL Workflow AI prompts for Flow Builder, each producing a single-word output that maps to a branch
Multi-bot orchestration structure where the pipeline needs specialised bots handing off to each other
Five or more test conversations covering the qualified path, graceful disqualification, objections, edge cases and re-engagement
A GHL setup guide: where each prompt goes, trigger configuration, calendar connection, pipeline stage mapping, handover notifications
A post-launch monitoring checklist and a common-fixes reference table
Voice AI adaptations if the bot runs on phone calls, including the tighter 10 to 15 word limit and rules against reading out links
Timeout handling and follow-up branches for every wait-for-reply step
A bot configuration summary appended to assets.md

Honest limits

Read this before you rely on it

This produces configuration, not a live bot. Nothing is pushed into GoHighLevel for you: you paste each section in, connect the calendar, map the pipeline stages and set the handover notification by hand.
Run all five test conversations inside GHL before a single real lead reaches the bot. A bot that fails on its second message costs you more than having no bot.
Human review does not end at launch. Read the first 20 live conversations yourself, then check qualification accuracy, booking completion rate and handover frequency, and correct the prompts against what real leads actually said.
Every bot needs a named human to escalate to and a defined trigger for doing it. Handover is part of the design, not a failure of it, and complex or sensitive conversations must reach a person.
The bot only knows what you wrote into the prompts. It is instructed never to fabricate, so anything you left out becomes a stall or an escalation rather than an answer, and the qualifying criteria you set loosely will let unqualified leads onto your calendar.

Where people go wrong

The mistake	Do this instead
Asking two or three things in one message to speed the qualification up	One question at a time. Stacked questions confuse leads and drop response rates.
Letting the bot write paragraphs	Twenty words maximum per reply for text, 10 to 15 for Voice AI. The bot is texting, not writing.
Letting the bot reference the qualification logic	It qualifies silently. It never mentions scoring, thresholds or internal criteria to the lead.
Writing qualifying criteria without disqualifying criteria	Define both, plus the escalation triggers. Loose criteria fill your calendar with people you cannot help.
Deploying straight from the generated prompts	Run the five test scenarios first, including the gibberish and off-topic edge case, and fix what breaks before any live traffic.
Cutting off unqualified leads bluntly	Graceful disqualification with a useful resource or alternative. They still talk to people who are a fit.

Worth knowing

GoHighLevel parses bullet points better than prose and branches on exact text matches, so keep every instruction a bullet, separate sections with ---, and make each Workflow AI step return one clean word and nothing else. A step that returns a polite sentence instead of the branch keyword breaks the routing silently.