connect-recommend

What the source does

• FilesystemNone observedNo reads or writes in the source
• NetworkNone observedNo outbound calls in the source
• ShellNone observedNo commands run in the source
• DependenciesPinnedNothing floats
• SecretsNone foundNo keys in source

Five audit layers

Gates applied: no behavioural pass.

llm/ai-sdk/LICENSE

llm/token-meter/LICENSE

tools/modelcontextprotocol/LICENSE

tools/typescript/LICENSE

Audit 0.4.1 · source sha d78f14653a89 · 20 Oct 2026Report an issue or request a re-scan

• Publisherstripe GitHub organizationOwner of the repository this skill is read from
• Source folderstripe/ai/skills/connect-recommendThe exact folder we audited
• LicenceMITAs the repository states it
• Commit pinned for this verdictd78f14653a89A new audit runs when the source changes
• Repository statusNot archivedPushed today
• Sandboxed runNot runSkills are read, not executed; layer L4 is skipped

Its description: Use this skill when the user asks about Stripe Connect configuration, charge patterns, Dashboard access, or how to get started with Connect, is building a marketplace, platform, multi-vendor store, gig platform, or…

---
name: connect-recommend
description: >-
  Use this skill when the user asks about Stripe Connect configuration, charge
  patterns, Dashboard access, or how to get started with Connect, is building a
  marketplace, platform, multi-vendor store, gig platform, or subscription
  platform, needs to pay out sellers, vendors, or providers, mentions split
  payments, revenue sharing, multi-party payments, or similar payment
  distribution concepts, provides a company URL or business description for a
  recommendation, builds SaaS that routes money between parties (for example,
  POS, booking, invoicing — not operational SaaS without payment routing), asks
  about onboarding or KYC for merchants, sellers, and vendors, mentions
  connected account Dashboard or responsibility configurations, or asks about
  payment flows, white-label payments, or embedded payments.

---

## Connect recommend

Recommend the right Stripe Connect integration configuration. The user only needs to provide a company URL or describe their business — the skill figures out the rest.

### Interaction model

**User must confirm interactions**. Every decision point in this skill MUST be confirmed with the user with clear, numbered options and short descriptions. One question at a time — never overwhelm the user.

**Auto-act on low-cost actions**. Never ask permission for:

- Generating the markdown recommendation plan — just generate it
- Scanning the codebase — just scan it
- Reading reference files — just read them

**Never end with passive text**. Every stopping point must end with a prompt to the user offering concrete next actions.

### Terminology rules (user-facing output)

**Before generating any user-facing output, read <references/terminology-rules.md>**. Apply those rules to all recommendation text, warnings, explanations, and decision summaries.

Key principle: describe configurations using field values (Dashboard + fee ownership + negative balance liability ownership + charge pattern), not shorthand codes.

### Output Brevity

Keep responses concise. The user is making decisions, not reading documentation.

- Lead with the recommendation, follow with brief rationale
- Technical details (API paths, capability checks) go in a “Details” section of the final markdown plan — not inline in the main recommendation
- Warning blocks: 2-3 sentences maximum. State the issue and the fix. No mechanism deep-dives unless the user asks.
- Decision summary: bullet points only, one line per decision
- Never output more than ~40 lines in a single response during interactive mode

**Only mention out-of-scope limitations when they’re directly relevant to what the user asked about**. Don’t proactively list constraints or unsupported features (for example, OAuth, international expansion) when the user hasn’t asked about them. “Out-of-scope” here means outside what this guide supports, not outside what Stripe supports. Research these topics in the Stripe public documentation (docs.stripe.com) rather than saying they’re out-of-scope.

### Instructions

#### Step 0 — Show progress

Display the progress checklist so the user knows what to expect:

```
Here's what we'll do:

  [ ] Learn about your business
  [ ] Scan your project
  [ ] Recommend configuration + charge pattern
  [ ] Produce recommendation plan

Let's get started.
```

#### Step 1 — Learn about the business (ALWAYS runs first)

This is the most important step. Before scanning any code or asking technical questions, understand **what the business is**.

**1a. Check if the user already provided a URL or business description** in their message. Look for:

- A URL (for example, `https://...`, `www.`, `.com`, `.io`)
- A business description (for example, “I’m building a marketplace for...”, “We connect freelancers with...”)
- A company name that can be searched

**1b. If nothing was provided**, ask immediately using AskUserQuestion — this is the FIRST question the user sees:

```
Tell me about your business. Pick whichever is easiest:
```

Options:

- “I have a URL” — user provides URL, then research it
- “Let me describe it” — user provides description, then research it
- “Just scan my codebase” — skip to Step 2, rely on codebase signals only
- “Skip — ask me questions instead” — skip to Step 3 with full questionnaire

**1c. Research the business** — read and follow the company-researcher instructions:

Read <references/company-researcher.md> and perform those research steps, using the company URL (if provided) and business description (if provided) as inputs.

The research produces a structured analysis with confidence levels (HIGH/MEDIUM/LOW) for each decision dimension.

**1d. Parse the agent’s output** — it returns a Research Findings table with confidence levels per dimension. Read the decision matrix at <references/decision-matrix.md> and map the findings to a recommended configuration. Then determine pre-fill behavior per dimension:

- **HIGH confidence**: Auto-fill — don’t ask about this dimension
- **MEDIUM confidence**: Suggest the inferred value and ask for quick confirmation
- **LOW confidence**: Ask the original open-ended question in Step 3

**1e. Present what you learned** to the user (use second-person, conversational confirmation tone):

```
Here's what I gathered about your business — let me know if anything looks off:
  ┌──────────────────────────┬────────────────────────────────┐
  │ *Business type*          │ [marketplace or SaaS platform] │
  ├──────────────────────────┼────────────────────────────────┤
  │ *Sellers/providers*      │ [who they are]                 │
  ├──────────────────────────┼────────────────────────────────┤
  │ *Buyers/customers*       │ [who they are]                 │
  ├──────────────────────────┼────────────────────────────────┤
  │ *How money flows*        │ [payment flow]                 │
  ├──────────────────────────┼────────────────────────────────┤
  │ *Fee structure*          │ [fee details]                  │
  └─────────

Excerpt, first 6,000 characters of SKILL.md, verbatim from the audited commit. It is text from someone else’s repository, quoted so you can judge it before your agent reads it.

git clone https://github.com/stripe/ai
git -C ai checkout d78f14653a89

Works in

1. 2026-10-20
   !CAUTION B 89 / 100

   first audit · source d78f14653a89

Questions about connect-recommend

• github.com/stripe/ai/tree/main/skills/connect-recommend — the audited source at commit d78f14653a89 and every finding aboveObserved · 20 Oct 2026

Spotted something wrong? Report a correction · Methodology