The Audit Deck Automation
Turn a recorded discovery call into a pre-populated audit deck at a URL you can review on your phone in under two minutes. Fireflies transcribes, Claude pulls the pain points and matches them to your past work, and n8n posts a finished deck. You stop writing audit decks from scratch.
Section One
What You're Building
You get on a discovery call with a prospect. Fireflies records it. The moment the transcript is ready, an automation reads it, pulls out the pain points the prospect actually described, matches them against a library of solutions you've already delivered, calculates the cost of doing nothing, and delivers a mostly-finished audit deck to a URL you can review on your phone in under two minutes.
You edit pricing, the roadmap, and drop in a hero image, then walk into the follow-up call already 80% ready. The point isn't that the deck is finished. The point is you never write an audit deck from scratch again — whether what you sell is an automation, a campaign, a retainer, or a rebuild.
Fig. 1 — system flow
Record
Fireflies
Call recorded, webhook fires on transcript-ready
Extract
Claude
Pain points, tools, headcount, revenue — from the call only
Match
Airtable Library
Verified past builds only
Deliver
Replit deck
Live URL + Telegram notification
The rule that makes this work
Every Claude step in this build is grounded on purpose. It's told to never invent tools, headcount, revenue, or dollar figures that weren't actually said on the call. A blank slide is correct behavior. A hallucinated one is the thing that gets you called out mid-pitch.
Section Two
What You Need
n8n
Orchestrates the whole pipeline: receives the Fireflies webhook, runs the three Claude reasoning steps, queries Airtable, and posts the finished payload to Replit. Self-hosted or cloud both work.
Anthropic (Claude) API key
Powers three separate reasoning steps: extraction, matching, and cost-of-inaction. One credential, reused across all three nodes.
Airtable
Your Workflow Library. This is the matching brain, and it's worthless empty. Seed it with builds you've actually shipped before you expect the pipeline to produce anything useful.
Fireflies
Records the call and fires a webhook the moment the transcript's ready. Needs a paid tier that supports GraphQL and outgoing webhooks.
Telegram bot
Pings you when the deck's ready to review, with the pain points, matches, and COI summary.
Replit
Hosts the deck itself: a React app plus an API endpoint your n8n workflow posts to. Needs a paid tier so the app doesn't sleep between calls.
Rough monthly cost floor if you're starting from zero: $70 to $100 a month, mostly Anthropic, Replit, and Fireflies.
Section Three
Step-by-Step Build
01
Step One
Deploy the Replit deck and API
Two apps, one pnpm monorepo: a React + Vite deck served at /audit-presentation/, and an Express API served at /api-server/. Storage lives in Replit Object Storage under a single key, not a database and not a file on disk, so it survives a cold start without losing state. Set two env secrets in Replit: PRESENTATION_API_KEY, the bearer token n8n will use to post payloads, and DEFAULT_OBJECT_STORAGE_BUCKET_ID, which sets itself once you attach Object Storage. Pricing, roadmap colors, and the hero image all get set later in the deck's own editor, not through the automation.
| Method | Endpoint | Auth | Purpose |
|---|---|---|---|
| GET | /api-server/api/presentation | None | Returns current deck JSON |
| POST | /api-server/api/presentation | Bearer token | Writes new JSON |
| PUT | /api-server/api/presentation/editor | None | Editor-only |
| DELETE | /api-server/api/presentation | Bearer token | Resets to sample data |
02
Step Two
Build your Solutions Library
This is the pattern-matching brain, and it's the one part that can't be handed to you pre-built. Whatever your business actually delivers, whether that's n8n builds, ad campaigns, financial plans, or security audits, seed the library with your own real, closed work. Seed it thin or with stuff you can't actually deliver, and the matcher will recommend exactly that.
Schema: Workflow Name, Problem It Solves, Source Type, Source, Source Link, Industry/Niche, Tools Used, Complexity, Deliverable, Tags, Date Added, Review Status, Entry Type. Rename fields to fit what you sell (Tools Used becomes Channels Used for a marketing agency). The matcher only queries records where Entry Type is Applied Solution and Review Status is Verified. Six real entries is enough to go live — add three more every time you close a deal.
03
Step Three
Adapt this build for your niche (optional)
Skip this if you're building for AI automation exactly as written. Everything in this build (transcript in, pain points out, matched against real past work, deck out) works for any business that runs a discovery call before a proposal, but the three Claude prompts are still written in AI-automation language. Adapt the prompts before you build, not after.
Copy this: the adaptation prompt (for Claude Code)
I'm adapting the Audit Deck Automation build for a different niche. My business: [YOUR NICHE]. What I deliver to clients: [YOUR DELIVERABLE]. Here is the build guide markdown describing the pipeline. Go through it and: 1. Rewrite the three Claude system prompts (extraction, matching, COI) to match my niche's language. Keep every grounding rule intact word-for-word: never invent tools, headcount, revenue, or dollar figures not stated on the call. 2. Rename the Airtable field references the pipeline depends on (Tools Used, Entry Type, etc.) and note every step downstream that references those fields. 3. Rename the Replit deck's slide labels to match what I sell. 4. Leave the pipeline mechanics untouched. Show me a diff before applying anything.
Why this stays manual
Auto-renaming fields and prompts without a human reading them risks breaking a node reference or softening a grounding rule you didn't mean to touch. Review the diff the same way you'd review pricing or a hero image: judgment doesn't run unsupervised.
04
Step Four
Build the n8n workflow
Connect Claude to your n8n account via the MCP connector, describe the ten-node pipeline below (or paste this guide in), and let Claude build it directly in your account. Every credential gets attached fresh in the next step.
| Node | What It Does |
|---|---|
| Fireflies Transcript Ready | Webhook, fires when transcript's ready |
| Normalize Webhook Payload | Handles field-name drift across payloads |
| Fetch Full Transcript | GraphQL call for sentences + attendees |
| Extract Pain Points | Claude step 1: industry, pain, tools, revenue |
| Search Workflow Library | Queries verified Applied Solutions only |
| Match Against Workflow Library | Claude step 2: fewer matches over weak ones |
| Calculate COI Estimate | Claude step 3: stated numbers only |
| Build Presentation Payload | Assembles the final JSON |
| POST to Replit Presentation API | Sends it live |
| Notify Telegram for Review | Pings pain points, matches, COI |
05
Step Five
Wire up credentials
Airtable nodes need a personal access token with data.records:read scope on your Workflow Library base. Anthropic nodes need one API key, created once and reused across all three Claude nodes. Fireflies needs a bearer token, Telegram needs a bot token from BotFather.
Three placeholders still need manual swaps
The POST node has your Replit URL and API key hardcoded as placeholders. The Telegram node has your chat ID and Replit URL the same way. Find-and-replace on REPLACE_WITH_ after import surfaces every spot that still needs your real values.
06
Step Six
Wire up Fireflies
Activate the workflow in n8n and grab the production webhook URL from the Fireflies Transcript Ready node. In Fireflies: Integrations → Webhooks → add webhook, paste the URL, set the trigger to "Transcription completed." Test with a 30-second recording and confirm n8n shows an execution.
07
Step Seven
Test the full flow
Two ways to do this. Run an actual test call, with someone playing a prospect describing two or three pain points that map to your library. Or skip waiting on Fireflies entirely: inject a saved transcript ID straight into n8n's Execute Workflow. When it works, Telegram fires 60 to 90 seconds after the transcript completes, and the deck shows real pain points, real matches, and a COI number that's either a genuine estimate or blank. Never invented.
What stays manual, and why
Pricing is yours to set per client based on scope and posture. Roadmap sequencing needs your judgment about what to promote from "opportunity" to "core." The hero image is where a real photo of the prospect's operation beats any placeholder. Automate these three and the deck looks automated. Leave them manual and it looks considered.
What'll break, and how to read it
No library matches means your library isn't seeded deep enough yet, not that the pipeline's broken. A blank COI means the prospect didn't state numbers on the call, which is correct behavior, not a bug. If you soften a grounding prompt to "always fill every field" to make either of these go away, you'll get a deck that invents facts instead. Don't loosen them.
Section Four
The Upgrade Path
Run this through a handful of real calls before you touch anything. Those first few decks show you where extraction gets sloppy or a match feels forced, and that's what tells you whether the problem is your prompts or your library. Tighten from evidence, not from guessing.
From there the same skeleton scales sideways. Point the matcher at proposal generation once the deck's proven itself. Feed the library into a second pipeline that turns verified work into a public case study page. Fork the deck template per vertical if you sell into more than one industry, since the slide order stays the same and only the framing changes.
| Your Niche | What Fills the Library | What the Deck Sells |
|---|---|---|
| Marketing agency | Past campaigns, tagged by channel and result | Channel mix and projected lift |
| MSP / IT provider | Security frameworks, past remediation projects | Risk score and remediation roadmap |
| Financial advisor | Planning packages, past client outcomes | Gap analysis and projection |
| Management consultant | Frameworks and closed engagements | Diagnosis and 90-day plan |
| Recruiter / staffing firm | Past placements by role and industry | Pipeline plan and time-to-fill estimate |
“A blank slide is correct behavior. A hallucinated one is the thing that gets you called out mid-pitch.”
— The Audit Deck Automation, p. 2
§ Fin.
Rather have it built for you?
This guide is one system from the Soulwire Studio workflow library. We build AI automations for businesses that run on process: intake, scheduling, proposals, compliance-heavy operations. If you'd rather spend your time closing deals than debugging automation, book an audit and we'll find the automation hiding in your business.
