AI Guides › Workbench
By Nigel Guy · 8 min read
Most people chase leads from memory. You scroll your inbox, reply to whoever is at the top or whoever you liked talking to, and the person who asked for a price three weeks ago and then went quiet never hears from you again. It feels like follow-up; it is really answering the loudest thread.
The rule: get the conversations out of the app as a file, have an AI sort them against criteria you wrote down first, and send nothing it drafts until you have read the original thread yourself.
People describe this as a "lead follow-up agent", as if it were one product. I could not find a single verified product by that name. What you can verify is the kit underneath it: official message exports from LinkedIn, Instagram and WhatsApp, plus a free AI chat that reads the file and hands back a ranked list. That is what this guide covers.
| Item | What it does | Cost at time of writing | Best for | The catch |
|---|---|---|---|---|
| LinkedIn data export | Gives you a file of messages sent, received and archived, with dates, content and profile links | £0 | B2B and freelance leads | Desktop only; the download link lasts 72 hours |
| Instagram export (Meta Account / Accounts Centre) | Exports your information, including messages, in HTML or JSON, with a date range | £0 | Creators and small shops selling through DMs | Meta says the link can take up to 30 days; you then have 4 days to download it |
| WhatsApp "Export chat" | Saves one chat as a text file, with or without media | £0 | Service businesses booking by WhatsApp | One chat at a time, from the phone app only. I could not load WhatsApp's own help page to confirm the steps |
| Claude Free | Reads the exported file and ranks the threads | £0; Pro is $20 a month billed monthly on the US page. I couldn't confirm the UK £ price, so check it at checkout | Long exports and careful reasoning | Free usage limits reset in five-hour windows; a big export can use up a session |
| ChatGPT Free | Does the same job | £0 | Short exports, or if you already use it | Free uploads are limited, and the free input cap for the default model is about 12 pages of text |
| Third-party "DM qualifier" browser extensions | Connect to your account and score conversations automatically | Varies | Nothing, on LinkedIn | LinkedIn's rules ban browser extensions that scrape or automate activity |
You use three pieces: an export, a scoring card, and a drafting prompt.
LinkedIn. On a computer, click Me, then Settings & Privacy, then Data Privacy, then Download your data. Choose the data you want (messages only is quicker) and click Request archive. LinkedIn says a specific category arrives within minutes and the full archive within 24 hours. Download it within 72 hours.
Instagram. Go to More, then Settings, then Meta Account (some accounts still show Accounts Centre), then Your information and permissions, then Export your information, then Create export. Pick the profile, choose Export to device, and select only messages with a date range covering the last 60 to 90 days. Pick HTML if you will read it yourself and JSON if the file is going to the AI. Then click Start export.
WhatsApp. Open the chat. On Android, tap the three-dot menu, then More, then Export chat. On iPhone, tap the contact name and scroll to Export chat. Choose Without media. You get a text file.
Before uploading anything, open the file and delete threads that have nothing to do with business.
This is the mechanism. Write it before you run anything, because if you leave "worth chasing" undefined, the AI defines it for you, usually as "whoever wrote the most".
| Signal | What counts | Points |
|---|---|---|
| Asked about price, availability or next steps | Their own words, not your pitch | 3 |
| Fits your offer | Matches the customer type you serve | 2 |
| Went quiet after your last message | Ball is in their court, 5 to 30 days ago | 2 |
| Named a deadline or event | "Before March", "for the launch" | 2 |
| Said no, or asked you to stop | Any clear decline | Remove from list |
Change the points to suit your business. The removal row stays in.
Upload the file into a new chat in Claude or ChatGPT and paste this prompt. Fill in the bracketed parts: your offer, who your customer is, today's date and your scoring card.
You are helping a small business owner decide which message threads to follow up on. You are a careful analyst, not a salesperson.
Context:
- What I sell: [YOUR_OFFER]
- Who I sell to: [IDEAL_CUSTOMER]
- Today's date: [TODAY'S_DATE]
- Scoring card: [PASTE_YOUR_SCORING_CARD]
The attached file is an export of my direct messages from [PLATFORM].
Steps:
1. Identify each separate conversation and the other person's name or handle.
2. Drop any conversation where the other person declined, asked me to stop, or is clearly personal rather than business. List these separately by name only.
3. Score the rest against the scoring card. For every point you award, quote the short phrase from the thread that earns it, with its date.
4. Rank by score, highest first. Break ties by who spoke last: if they are waiting on me, they rank higher.
Output a table with these columns: Rank, Name, Score, Evidence (quoted phrases with dates), Who spoke last, Days since last message, Suggested next action in under 15 words.
Below the table, list conversations you could not read or score, and why.
Rules:
- Use only what is in the file. Do not guess intent, budget or company size.
- If the file is cut off or a thread is unclear, say so rather than filling the gap.
- If any context item above is missing, ask me for it before you start.
Before you answer, check that every point you awarded has a quote behind it and that no declined person appears in the ranked table.
Every score carries its reason, so you can check the top five in minutes instead of trusting a number.
Take the top few rows and draft each reply separately. Fill in the person's name, the evidence row from the table, and what you can offer next.
You are drafting a short follow-up message for me to edit and send myself.
Person: [NAME]
What they said, with dates: [PASTE_EVIDENCE_ROW]
Where we left it: [ONE_LINE_SUMMARY]
What I can offer next: [NEXT_STEP, e.g. a call slot, a quote, a sample]
My tone: [e.g. friendly and plain, no exclamation marks]
Write two versions, each under 60 words:
A) picks up their last question directly
B) offers the next step and gives them an easy way to say no
Do not invent discounts, deadlines, scarcity or anything I have not stated. Do not mention that you read an export. If you need a fact you have not been given, write [CHECK] in its place.
Before answering, confirm that neither draft claims anything not listed above.
The ranked table feels finished, and that is exactly the problem. The model reads text, not tone. It cannot see that someone left you on "seen", what you said on a call, or that "maybe next year" from this person is a polite no. Open the original thread before every send. If the evidence quote isn't in it, drop that row.
The second trap is volume: ten carefully read follow-ups beat forty pasted ones.