A free Claude Code skill that designs your whole customer referral program: who to ask, when to ask, and what to offer, backed by real published benchmarks.
Referral Programs is a Claude Code skill: a markdown playbook that walks your agent through building a full customer referral program from scratch, not just a list of tips. It picks the customer segment most likely to send you good referrals, matches the reward type to your product's actual value metric, builds a three-tier commission structure, writes the fraud-proof mechanics (pay on collection, not sign-up), recommends a tool, lays out a week-by-week launch sequence, and gives you a measurement dashboard benchmarked against real published B2B SaaS referral data.
The video's claim is that "this agent builds the referral system for you: who to ask, when to ask, and what to offer." That's true and tested. Run against an invented business (a B2B warehouse software company, made up for this test, not a real company), it produced a specific eligible customer cohort instead of "ask everyone" (customers live 6+ months, using it heavily, multiple locations, happy in their last survey), a specific trigger moment instead of "ask on day one" (right after the customer hits their own first real win, not during onboarding), and a specific offer matched to what the product actually does (product credit first, cash-equivalent credit only once someone refers enough people that credit alone isn't enough). All three parts of the claim held up concretely.
One honest caveat: the skill ships a small self-check script that is supposed to confirm your finished program is complete. In testing, it failed a genuinely complete, well-built program because it just checks for a handful of literal words rather than actually checking the content. Don't rely on that script to tell you if your program is good, read the output yourself, it's the actual guidance that does the work, not the checker.
Any business owner or marketer who knows their happiest customers would send more business their way, but has never actually built the system: who qualifies, what moment to ask them, what to offer, and how to stop people gaming it. This replaces the work of a paid growth strategist sitting down and designing that program with you.
This copies the whole gtm-skills library onto your machine. You only need one folder out of it.
git clone https://github.com/LeadMagic/gtm-skills.gitThis drops just the referral-programs skill into the folder Claude Code checks for project skills.
mkdir -p .claude/skills
cp -r gtm-skills/skills/growth/referral-programs .claude/skills/referral-programsNote: the repo also documents a curated single-skill installer, scripts/cc-gtm.py. On the test machine's default Python (3.9), that script crashed outright with a syntax error, it needs Python 3.12 or newer. The plain copy above is the path that actually worked on the first try, and it is also the exact fallback method the repo's own installer uses internally.
Sample input used in testing: an invented B2B SaaS business, made up for this test, with 400 paying customers, a $12,000/year average contract, and a clear "first real win" moment (their software successfully syncing data for the first time).
Abbreviated real output, following the skill's own six-phase process:
Segment: customers live 6+ months, heavy usage, multiple locations,
happy in their last survey. About 70 of 400 customers qualify.
When to ask: right after their own first real win, and again after
a great survey response. Not on day one.
What to offer: product credit that matches what the product actually
does, three tiers (10% / 12% / 15% of year-1 contract value as the
program scales), reward pays only once the customer actually pays,
never on sign-up.
Result: estimated cost per referred customer came in about 58% below
a typical paid-acquisition cost, in the same range the skill's own
published benchmarks predict.scripts/cc-gtm.py (the "curated installer") crashes with a SyntaxError about an f-string and a backslash. That script needs Python 3.12 or newer. If you're on an older Python 3, use the cp -r copy method in Step 2 instead, it lands in the same place.
The bundled scripts/check-output.py self-check says FAIL on a program that looks complete. It only checks for a short list of literal words, not whether your program is actually good. A finished, correct referral program can still fail it just because it never uses the word "affiliate," for example. Ignore the checker, read the actual output and check it against the skill's own Quality Check list inside SKILL.md instead.
You're not sure which incentive type to pick. The skill has a built-in decision tree in references/framework-notes.md, ask Claude to walk through it with your product's actual value metric (storage, seats, credits, whatever you charge for) and it will pick for you.
You don't know your real numbers yet (CAC, ACV, NPS). The skill will still build the full structure and flag exactly which numbers you need to plug in before it's launch-ready, it doesn't block on missing data, it tells you what's missing.
Built by LeadMagic (https://github.com/LeadMagic/gtm-skills), MIT license, 53 stars at test time. The video's claim held up exactly as stated: it produces a specific "who to ask, when to ask, what to offer" program, not generic advice. Only correction: the bundled self-check script is weaker than it sounds, it checks for keywords, not quality, so don't lean on it as proof your program is finished.
This is one agent. The 7-Day AI Native Challenge is how you learn to build the next ten yourself, live, with Dany.
Join the Challenge →Too busy to build it yourself? Focentra AI will build, install and train your team on it. Priced on the call.
Talk to Focentra AI