I Built the Product. Then I Had to Learn How to Sell It.

How I turned my search for the first pilots into 16 AI skills, and how you can install them in ChatGPT, Claude, Gemini or any AI worker

30 September 2026 · By Elio Fernandes, founder of Flix Home · © 2026 Elio Fernandes. All rights reserved. Feel free to apply these ideas and quote short excerpts with credit and a link to this post.

I'm a builder. I built Flix Home, a tool that gives short-term rental managers control and visibility over their whole property portfolio. Building was the part I knew how to do.

Then came the part nobody teaches you: finding real customers and getting the first ones to say yes to a pilot.

This is not a victory lap. It is early evidence that a better process can change the outcome. With this new workflow, I built a list of 100 leads. In the first week, it produced five meetings and I closed two 60-day pilot agreements. The sample is still small, but the difference was useful enough that I documented every step.

This article covers that method, the mistakes that shaped it, and how to copy it for your own product.

1. From silence to a repeatable flow

Between late June and mid-July 2026, I sent 91 cold emails to 31 property management companies: three emails each, over two weeks.

I got zero replies. No bounces and no opt-outs either. Just silence.

Looking back, volume was not the problem. These were:

• I led with what the product does, not with what the manager worries about.
• Every email sounded like every other SaaS email.
• I had no way to learn from the silence. I kept no clean record of what was sent, from which account, with which subject and with what result.
• The work restarted from scratch every time, so mistakes repeated: a wrong email address, a lead treated as new when it had already been contacted, a meeting logged as "almost closed" when nothing had been agreed.

So I changed three things. I open with the manager's problem: control and visibility over the whole portfolio. I ask one question. I propose one small next step.

Then I rebuilt the whole flow around those rules. I prepared 100 leads and ran the new process. In the first week, it led to five meetings and two agreed 60-day pilots. These are still early numbers, not proof of a repeatable sales engine. But the workflow helped me make more progress in a short time, and I could finally see why each step existed.

That is why I designed and documented every part of it: to help other founders who have built the product and are now learning how to sell it without restarting from zero every week.

2. The idea: every repeated task becomes a skill

A skill is a written method: steps, rules, limits and the expected output. It is not a prompt you retype. It is a file the AI reads every time it does that job.

Once I had written down how I prepare a call, how I follow up after a discovery meeting and how I review my week, my AI workers could take over the preparation. That left me the conversations.

The most important design choice was what goes where:

If you put changing facts inside a skill, it quietly goes stale. If a suggestion slips into the decision log, it turns into a false commitment. Keep the three apart.

3. My workflow, end to end

This is the loop my AI workers and I run every week:

1. Find: pick leads from an approved list and check they have never been contacted.
2. Contact: write the message, have it reviewed, and send only after I say yes. Then log it in the CRM the same day.
3. Talk: prepare a one-page sheet before each call or meeting. After the meeting, turn the transcript into facts, proposals, decisions and open questions. These are four different things.
4. Propose: write a factual recap and, separately, a pilot proposal.
5. Measure: once a pilot starts, track it against a success metric agreed beforehand.
6. Review: every Friday, run one review that updates the scorecard, the risks and the decisions I need to make. Anything I decide goes back into the context, so the next run starts from the current picture.

4. Five lessons that changed how I prospect

4.1 Open with their problem, not your features

Every message follows the same order: context → relevance → one question → one light next step. If I have no real, sourced observation about the company, I leave the personalisation out. Invented personalisation reads worse than none.

Before anything goes out, the draft is scored on seven points: clear, relevant, natural, specific, one next step, proof, brief. If it fails two or more, it gets rewritten.

Follow-ups without a reply move through three steps, each lighter than the one before:

1. Bring back the context.
2. Add something new and useful.
3. Give an easy way out, and ask who the right person is.

4.2 Ask for the meeting early, and ask for the right person

From Sam Nelson's Cold Calling Essentials: a call has only three acceptable endings. A meeting booked, the name of the right person, or a clear no. "Send me some information" is usually a polite no. Send it, but propose 20 minutes to go through it together.

Asking "who should I be talking to?" is the most underrated question in early sales.

4.3 A human says "send"

My review skill can say a text is good. Only I can say it goes out, to whom, and from which account. "Editorially approved" does not mean "authorised to send".

When the AI is missing a tool, a permission or a source, it reports an honest pending item. It never claims it did something it couldn't do.

4.4 Define pilot success before the pilot starts

This one comes from Mark Roberge's The Science of Scaling and changed how I write pilot proposals. Success is a Percentage of customers reaching a concrete Event within a Time window.

The metric goes into the proposal as a hypothesis, and gets measured at fixed milestones, for example day 30 and day 60. When there is no data, the report says "evidence missing". Interest, a nice meeting or a letter of intent is not usage.

The proposed pilots run for 60 days at no cost, but I frame them as an exchange, not a freebie. The client gives access to their real process, names one responsible person, gives weekly feedback and joins a review on day 60. Once a pilot starts, I will track the agreed metric at fixed milestones. I never lead with "free". I lead with the problem.

4.5 Know your real competitor: the spreadsheet

Adapted from HubSpot's 30-Minute Battlecard Builder: one card per alternative, and the most common alternative is the status quo, spreadsheets and chat groups. Every objection gets a two-sentence answer: acknowledge the concern, then redirect with a question.

When you have no proof of your own yet, the card says "evidence missing". Don't borrow someone else's case study. Review the cards every two weeks, and mark any card older than three months as stale.

5. The 16 skills at a glance

Every template has the same structure: Objective · When to use · Inputs · Output · Human approval · Common mistakes · Steps · Copy-paste prompt.

6. Adapt them to your product in five steps

1. Fill in the product parameters below. Each field needs an owner, a source and a date.
2. Write your communication rules: allowed claims, banned words, and what you never promise.
3. Connect tools with least privilege: only the accounts and permissions the skill needs.
4. Name the approver: who says yes, and to what exactly.
5. Test on one real case, compare the result with the sources, then fix the method as well as the output.

Put this block at the top of every skill. An empty field blocks the actions that depend on it. It never gives the AI permission to guess.

## Product parameters
PRODUCT: {{factual name and description; source/date}}
ICP: {{segment and target decision-maker}}
PROBLEM: {{observed pain, or a hypothesis marked as such}}
VALUE PROPOSITION: {{allowed wording and limits}}
MATURITY: {{available / in validation / planned; evidence}}
SENDER: {{approved account and visible identity}}
APPROVER: {{who gives explicit authorisation}}
CRM: {{system, owner, write rules}}
FORBIDDEN WORDS: {{terms and promises that never appear externally}}
CLAIMS ALLOWED WITH SOURCE: {{claim | source | date}}
PRICE: {{validated / in validation; how to talk about it}}

7. Install them in any AI tool

A skill is plain Markdown, so it works anywhere a model can read instructions. Keep one file per skill:

---
name: commercial-messages
description: Writes one commercial message (email, follow-up, LinkedIn, WhatsApp) using context → relevance → question → next step, with a 7-point test. Never sends.
---

# Commercial Messages
## When to use
## Inputs
## Steps
## Rules and forbidden words
## Output
## Human approval

Where to put it:

• ChatGPT: create a Project or a custom GPT. Paste the skill into the instructions, or upload the .md files as project files. Then ask: "Use the Commercial Messages skill to draft a follow-up to…"
• Claude: add the files to a Project's knowledge or instructions. Or package each skill as a folder with a SKILL.md file (like the one above, with name and description at the top) and add it as a Claude skill. In Claude Code, put it in .claude/skills/<skill-name>/SKILL.md.
• Gemini: create a Gem and paste the skill as its instructions.
• Spinnable: this is an AI-worker platform that can keep reusable skills and work with connected tools such as email, Drive and a CRM. Send the .md file to your worker and say "save this as a skill". It can then run the method on demand or on a schedule, such as the Friday review.
• Any agent framework: load the file as the system prompt or tool instructions for that task.

Menu names change often, so check each tool's current docs. The rules don't change: connect only the access you need, and keep the human approval step.

8. Where I am now

I use the same format I use with my mentors:

• What I know: the first attempt taught me more than any course. With the new 100-lead flow, the first week produced five meetings and two agreed 60-day pilots. Opening with the manager's problem gets conversations. Writing the method down stopped the same mistakes from repeating.
• What I'm testing: once the pilots start, whether they turn into paid customers, whether my success metric is the right one, and whether referrals and partners can become a second channel beside cold email.
• What I don't know yet: the right price, and whether this works beyond my first city.

I'll write the next part when the first pilot has run its 60 days, whatever the result.

9. Start this week

1. Pick one task you repeat every week, like a follow-up email or a weekly review.
2. Write down how you did it last time, including the mistakes you fixed.
3. Copy the matching template, fill in the parameters and name the approver.
4. Run it once under supervision and fix the method, not just the output.
5. Only then schedule it, or hand it to an AI worker.

Get the 16 templates on GitHub.

Credits

These skills are adaptations, not copies. No long passages are reproduced. Thank you to:

• HubSpot, Cold Calling Essentials, by Sam Nelson: asking for the meeting early, three acceptable endings for a call, asking for referrals, handling objections.
• HubSpot / The Science of Scaling, The 30-Minute Battlecard Builder: fast competitive research and two-sentence objection answers.
• Mark Roberge, The Revenue Leader's Guide to Scaling (The Science of Scaling): P-E-T, readiness signals, and why scaling too early is dangerous.
• HubSpot, The Partnership That Built doola: the idea of testing partners as a growth channel.

Thresholds and conclusions in your own use depend on your own data.