feat(sales): subscription gap audit + four workflows

[bpalmer77fluid]

One skill and four workflows, all built and run against a live Fluid company today.

sales/subscription-gap/SKILL.md
sales/subscription-gap/references/replenishment-classes.md
sales/subscription-gap/references/measuring-cadence.md
workflows/subscription-gap.workflow.json           8 steps
workflows/setup-tiktok-shop.workflow.json          8 steps
workflows/setup-instagram-affiliate.workflow.json  7 steps
workflows/subscription-sms.workflow.json          10 steps

The skill

Subscription gap audit finds the products a merchant sells which run out and get
rebought, but that no customer can put on a schedule. On the personal-care catalog I built
this against, 15 of 26 products were replenishable with zero carrying a subscription
plan — $151 of list price no customer could reorder.

Folder shape because the term lists and the SQL are lookup material the agent doesn't need
in context on every turn.

The workflows

setup-tiktok-shop — screen the catalog against TikTok commerce policy, pick the category
gate and its real timeline, set the creator qualification gate and commission tiers from live
prices, build the enrollment page and merchant console, close with an honest blocker list.

setup-instagram-affiliate — Meta commerce policy is a genuinely different ruleset, not
TikTok's with the names swapped. Meta ended native checkout in August 2025, so every tagged
tap lands on the merchant's own domain and the work is link attribution, not order
reconciliation.

subscription-sms — detect the messaging rail the merchant already has, clear the two
blockers that actually stop SMS launches (sender registration and consent), scope the
actions, deploy a branded droplet.

subscription-gap — the 8-step build that pairs with the skill.

What the live runs taught, and why the prompts look the way they do

Most of these were expensive to learn because nothing errored — the output just came back
confident and wrong.

Pin /api/v202604/company/products + /{id}. The shop-host /api/products payload
embeds the entire country atlas — roughly 25KB per product — and the admin list endpoint
carries no variants[] at all. Only the detail endpoint has everything. Choosing wrong
doesn't fail, it just quietly gives you nothing usable.

description is an object with a .body, not a string. String() coercion yields
[object Object], which silently reduces classification to title-only. A 26-product catalog
classified cleanly on half the evidence and looked fine.

price is null at product level on the shop-host payload; the real value is in
variants[].variant_countries, a dict keyed by country code. Reading the product field
returns $0.00 for an entire catalog and a gap report quoting "$0 of list price."

Eligibility and reviewability must stay separate numbers. A content-review flag should
never reduce the count of what's eligible to list — they answer different questions.

Weight-management products are prohibited outright on TikTok Shop. Not restricted.
Banned. For a book full of weight-management merchants that's a screening question, not a
footnote.

Drop rank-based creator gating when the company has no distributor ranks rather than
offering a qualification path that doesn't exist. Check is_mlm and the distributor count
instead of assuming.

Don't recommend on-platform autoship after proving the plans it needs can't be created.
Every documented v2026-04 subscription path is a customer-instance operation — cancel, pause,
skip, retry. There is no endpoint to create a product subscription plan.

Enrollment belongs on the storefront, never behind admin login. Creators have no admin
session. The merchant console is the droplet; the enrollment page is a theme page.

A launch checklist that reads clean while a 21-day approval sits unstarted is worse than no
checklist. The closing step's acceptance criteria enforce that.

Design decisions worth a reviewer's eye

Structure beats text. is_bundle / product_bundles / bundle_config are authoritative
when populated. Text classification is only the fallback — necessary because Fluid returns
category: null and tags: [] on a real import.

Text matches are scored, not hand-ordered. Title evidence outranks description evidence;
longer terms outrank shorter. This isn't premature generality — a hand-ordered version failed
in both directions at once on a real catalog. It classified two wax kits as bundles, hiding
genuine recurring revenue, while a real multi-product set carrying no bundle word fell through
to treatment and would have been offered as a subscription. Subscribing a bundle re-ships a
durable the customer already owns, which is how a subscription earns a cancellation.

Word boundaries, case-sensitive acronyms. Substring matching finds TENS inside
"softens", "brightens" and "moistens", which told a skincare merchant their lip balm was a
medical device.

Coverage is checked at product AND variant level. A plan attaches at either. Checking one
under-reports coverage and manufactures a gap that isn't there.

It refuses to invent revenue. The estimate needs an attach rate and a buyer count, both
from order history. With no reporting database connected the skill says so and gives the
defensible version — N products and $X of list price outside any schedule — rather than
borrowing an industry attach rate.

Add-on before skip. Only adding an item creates incremental revenue. A flow that opens
with "want to skip?" reduces the merchant's revenue very efficiently.

Where these apply

Any merchant selling consumables — personal care, supplements, food and beverage, pet,
household, cosmetics. The subscription-gap trigger is computable from the catalog alone:
a replenishable product with no subscription plan. Weak fit on apparel and hardgoods,
which are mostly durables — the skill correctly returns NO GAP rather than inventing one.

The channel workflows need no comp plan and no distributors, which is why they port to
another company in minutes.

Verification

• python3 scripts/validate_catalog.py → 43 skills, 11 workflows, passes
• The 14 contract tests pass: python3 -m unittest discover -s scripts -p 'test_*.py'
• All four workflows validated for stored-shape parity against the known-good
  subscription-sms.workflow.json; DAGs sort clean
• Classification logic separately unit-tested (17 tests) and mutation-tested outside this
  repo: removing title weighting fails 2, removing ambiguity flagging fails 3, removing the
  structural bundle check fails 2
• Exercised against two live Fluid catalogs — a 31-SKU and a 26-SKU personal-care book

Open question for maintainers

Four workflows is a lot to land at once. If you'd rather take the skill alone and have the
workflows as a follow-up PR, say so and I'll split it — they're independent.