HypothesisIf partners' coding agents follow skills written by Wise, partners reach their first sandbox transfer sooner and raise fewer integration tickets.
84% of developers use or plan to use AI tools.[SO] Wise's Docs MCP already gives their agents the API reference. Skills add what Wise's integration engineers know: what to check first, what order to build in, and where integrations break.
The Docs MCP says what an endpoint does. A skill says how to do the job in this codebase, and when to stop and ask.
“Add a Pay out button that pays for the transfer from the seller's balance.”
Writes the payment call. Misses that the app's existing code treats every 403 as an expired login, and Wise's first answer to a payment is a 403 asking the seller to confirm.
Payouts loop and never go through. The rule was in the docs. Nothing told the agent to look.
Reads the existing code first and raises the problem before writing anything. Fixes the login handling, then builds the button.
The payout goes through after one confirmation.
I picked this step because the confirmation check (SCA) is where public Wise integrations have broken.[n8n][sure]
Getting new partners live two to four weeks sooner brings forward roughly $50,000 to $1.4m of revenue a year.[FY26][CFO] The likelier value is capacity: Wise wants Platform at around 10% of volume within three years,[plan] and every new partner brings its questions to the same integration team.
Four checks, cheapest first. Any one of them can end the idea.
Build time against total time to launch, for recent partners.
Wise's own time-to-integration data. A lookup.
Contracts and approvals set the pace for nearly every partner.
Evals. The same agent on the same tasks, with and without the skills.
No partner and no production access.
There's no clear gap. Agents can also skip skills: in Vercel's tests they did in 56% of cases.[V]
Replay this year's 40 or so API changelog entries against the skills and count the edits.
The public changelog.[log]
Most entries force an edit.
A pilot. Tickets per partner and days to first sandbox transfer, with and without the skills.
A few partners who are mid-build.
Neither number moves.
From outside Wise I can run checks 2 and 3, so the evals are next. Inside Wise I'd start with check 1. It's a lookup, and it can end the idea before anything is built.
There are many ways to build up a skill library. Here's one example: a skeleton that maps a whole integration, with the funding domain built out and the rest as light samples. A real version would be shaped by Wise's support ticket profile, with skills written for the steps where partners actually get stuck.
65 skill slots: 10 built out, 53 light samples rephrased from Wise's public API docs, 2 empty stubs. None of it is production-grade. It opens on the skill behind the example above.