Teardown No. 109 · Lead magnet

12,271 DMs, 3,771 Emails: How Melvyn Malherbe has a Hermes agent turn Instagram comments into signups, and why his dashboard counted wrong

Melvyn Malherbe, a French developer who publishes AI coding tutorials as Melvynx, says his Instagram account went from about 1,000 to 10,000 followers in a little over two weeks. Under each reel he asks viewers to comment a keyword such as "prompt" or "brain". A Hermes agent running on a rented server answers each comment with a DM that links to a resource page for that keyword, where people leave a first name and an email to get the tutorial. His screen shows 12,271 DMs sent and 3,771 unique emails collected from Instagram in two weeks. He is careful to say that the DM count is messages, not people, and that one dashboard the agent built for him counts wrong.

Melvyn Malherbe · melvynxVerified Nov 13, 2026Markdown copy

Try this step

One keyword, one resource page. Each reel ends with "comment prompt" (or brain, audio). Each keyword maps to its own page at ressources/<word>, and he says the AI creates these pages automatically. The page asks for a first name and an email and offers an optional tip of 0, 3 or 10 euros.

The problem

He makes tutorials, not vlogs, and he writes that before this he could not make content that worked on Instagram because editing blocked him. Once an editing workflow fixed that, each reel could ask for a keyword in the comments, and each comment needs a DM with the right link. He skips ManyChat and similar paid tools: his own agent sends those messages.

The blueprint (as he describes it)

  1. One keyword, one resource page. Each reel ends with “comment prompt” (or brain, audio). Each keyword maps to its own page at ressources/, and he says the AI creates these pages automatically. The page asks for a first name and an email and offers an optional tip of 0, 3 or 10 euros.
  2. Let the agent send the DM. Hermes, which he calls the asynchronous half of the system, publishes the scheduled posts, answers DMs and sends the resource link. When someone gives an email in a DM, it looks them up in his email tool. It copies his tone: a “thanks” gets a casual “you’re welcome, bro” back (translated from French).
  3. Scout trends every morning. Every day at 7:26 the agent scrapes videos in his niche and lists the outliers, reels with far more views than usual. He picks one and remakes it his way.
  4. Rewrite, don’t copy. The AI keeps the original script’s structure and sentence order but rewrites it with his opinions and the tools he recommends. His prompt says not to copy the original creator’s interface: the design is rebuilt from scratch. He keeps these videos for the top of the funnel and says he does not overuse them.
  5. Edit with Claude Code, then give feedback. Claude adapts the script, gives him a prompter link to record with, runs his editing skill and reports that the short is ready. He reviews it in his studio, notes what is wrong, and a second version is built before he says “schedule it”. An AI voice is generated for the outro of the TikTok and Shorts versions.
  6. Check what every number counts. He added a prompt to the article for anyone copying this: ask the agent whether each dashboard figure counts unique people or events, where the data comes from, and whether any funnel step has more entries than the step before it.

Stated results

  • Instagram, after about two and a half weeks: 10,000 followers, about 8,000 gained (Instagram shows +562%), 1,020,000 views and about a hundred posts. TikTok, over about two weeks: 12,800 followers and 67,000 likes. He says these are screenshots of a moment, not live counters.
  • His funnel screen: 12,271 DMs sent (messages, not people), 6,917 site opens (the tool’s counter), about 5,000 sign-ups (probably not unique) and 3,771 unique emails from Instagram in two weeks.
  • The “40%” DM-to-signup rate in his video is about 5,000 divided by 12,271, so events, not people. With the 3,771 unique emails he puts it at around 30%, and says even that is approximate.
  • 31 euros in tips from the resource pages.
  • One short from this workflow reached 200,000 views on TikTok, and one Instagram video had 100,000 views three days after posting, 43 times his median.

Caveats

  • The dashboard is not reliable yet, by his own account. When he asked Hermes for conversion by keyword, it ran a Claude Code session from 8:47 to 9:13 and shipped the page. On one keyword that page shows 1,745 DMs, 107 site opens and 128 sign-ups, which is impossible. He thinks the resource system was added after the tracking and that PostHog work remains.
  • The follower and view figures are snapshots, and the Instagram and TikTok periods may not match. He notes that he already had a YouTube audience, a clear niche and products to offer, so the result does not prove each part of the system caused it.
  • He sold his AI video editing course afterwards and gives no revenue figure. The article ends with offers for a free masterclass and a pack of Claude Code editing skills.
  • He says the setup takes some technical skill, and that the hard part is the initial setup rather than each request after it.
  • His own rules for copying this: only message people who asked for something by commenting the keyword, no mass cold DMs, check each platform’s rules, reread the automatic replies at the start, and collect emails with clear consent and one-click unsubscribe.
  • The article is in French; quotes here are translated.
  • Handles: X @melvynx (bio links melvynx.dev and names Lumail, which melvynx.com lists as his product), Bluesky melvynx.dev (domain handle), Instagram @melvynxdev (linked from melvynx.com) and the matching Threads account, verified 2026-11-13. No TikTok account verified for him.

FAQ

How does Melvyn Malherbe get emails from Instagram comments?

Each reel asks viewers to comment a keyword. A Hermes agent sends each commenter a DM with a link to that keyword’s resource page, where they leave a first name and an email to get the tutorial.

Is the 40% conversion rate real?

Not as a rate of people, he says. It is about 5,000 sign-ups divided by 12,271 DMs, and both count events. Using the 3,771 unique emails he puts it around 30%, and calls even that approximate.

What does the agent do on its own?

It publishes scheduled posts, answers DMs and sends resource links, looks up subscribers in his email tool, and scrapes his niche every morning at 7:26 for outlier reels. He still records, reviews and gives feedback on each edit.

Did the agent build anything that went wrong?

Yes. The per-keyword conversion page it built in one session shows more sign-ups than site opens on one keyword, 128 against 107, which he says means the data is wrong.

Sources

Want agents like these working your distribution? Get It Built. Get new teardowns weekly: subscribe (double opt-in).

Get one teardown a week.

Same format, a new operator every issue. Double opt-in.