Building an Eternal AI Workflow for Solo Operators

Last month I needed to send a personalized follow‑up to every lead that downloaded my free guide, but I didn’t want to sit at my laptop all day copying and pasting. I wanted something that would run on its own, keep the tone human, and not break after a few dozen messages. I built a loop that pulls new leads from a Google Sheet, runs them through a Claude prompt, and sends the email via the Make platform — all without writing a line of server code.

Why does the AI agent keep forgetting context?

One of the first things I noticed was that after about ten runs the output started to sound generic. The model was losing the thread of each lead’s industry and pain point. I traced it to the way I was passing the previous output back into the prompt. If you just append the whole conversation, the token balloons and the model truncates the oldest bits.

What fixed it was to compress the history into a short summary before each call. I added a tiny preprocessing step that asks Claude to give me a two‑sentence briefing of the lead’s profile and the last email sent. That summary stays under 150 tokens, so the model never drops the important details.

Here’s the exact prompt I use for the summary step (plain text, no special formatting):

Give me a two‑sentence summary of the following lead: {lead_name} works at {company} in {industry}. Their last email said: {last_email}. Focus on the main pain point and the product they showed interest in.

That little tweak kept the personalization sharp even after two hundred runs.

What most guides get wrong about chaining prompts

Many tutorials tell you to daisy‑chain prompts by feeding the full output of one step directly into the next. They act like the model has infinite memory. In reality, each call has a hard token limit and the model will start dropping the earliest parts of the context once you go over.

The guide‑writers also tend to ignore the cost of each extra call. If you run three prompts per lead and you have a thousand leads, you’re suddenly paying for three thousand API calls. That can blow a modest budget fast.

What works better is to keep each prompt laser‑focused and to pre‑process data outside the model whenever you can. Use a spreadsheet formula or a Make.com module to clean up dates, strip HTML, or format a list. Let the AI do the creative heavy lifting — writing the email body — and let the automation handle the mechanical bits.

Concrete example: weekly outreach with Make.com, Claude 3.5, and Google Sheets

Below is the exact flow I run every Monday morning. First mention of each tool is bolded.

1. Google Sheets – a sheet called “Leads” with columns: Timestamp, Name, Email, Company, Industry, Last Contact, Status.
2. Make.com watches for new rows where Status is “New”.
3. When a new row appears, Make.com calls a Claude 3.5 endpoint with the summary prompt shown above.
4. The summary is stored in a temporary variable.
5. Next, Make.com calls Claude again with the email‑writing prompt: “Write a friendly follow‑up email to {Name} that references their interest in {product} and mentions the pain point from the summary: {summary}. Keep it under 150 words.”
6. The generated email is inserted into the sheet’s “Email Draft” column and the Status is changed to “Ready”.
7. Finally, a separate Make.com scenario runs every hour, picks rows with Status “Ready”, sends the email via Gmail, and updates Status to “Sent”.

Cost breakdown (as of 2026):

• Make.com free tier allows 1 000 operations/month — more than enough for a solo operator with a few hundred leads.
• Claude 3.5 API: $0.008 per 1 000 tokens. My average call uses about 600 tokens, so each lead costs roughly $0.005 for the two calls. For 500 leads a month that’s $2.50.
• Google Sheets is free with a personal Google account.

I think the $29/mo Make.com Pro plan is overpriced if you only need the basic webhook and email modules; the free tier handles my volume just fine.

How to debug when this workflow breaks

When the automation stops sending emails, the first place I look is the Make.com scenario log. If you see a “401 Unauthorized” error from the Claude module, double‑check that the API key hasn’t rotated. Some vendors silently expire keys after 30 days.

If the emails look blank, check the temporary variable that holds the summary. A common mistake is to map the wrong output from the previous module — Make.com will happily pass an empty string if the field name is misspelled.

Another failure mode is the Google Sheets update step. If the sheet is temporarily locked because you have it open in another tab, the write operation will queue and eventually time out. The fix is to keep the sheet closed during automation windows or to use the “Append Row” module which is less prone to locking conflicts.

Finally, watch your token usage. If you start seeing “max_tokens exceeded” errors, your summary prompt is probably too long. Trim it down or increase the max_tokens setting in the Claude call (just remember that raises the cost).

One concrete gripe: I got frustrated when Make.com changed the UI for the webhook trigger without warning, breaking my test scenario for an afternoon. The new layout hid the “Response” field under a collapsible panel, and I kept mapping to the old invisible field.

One concrete love: I love how the built‑in scheduler lets me set cron‑like triggers without touching a server. I can set the email‑send scenario to run at 9 am every weekday and forget about it.

Price mention with opinion: $0.008 per 1 000 tokens for Claude 3.5 is fair for the quality you get; the free tier of Make.com is enough for solo work.

We cover this in more depth elsewhere — deeper coverage of AI agent platforms.

If you’d rather skip the build and deploy a working version in an afternoon, we’ve packaged this workflow as a blueprint at deepusecase.com/vault.