You export a report from some web app and it hands you a 40,000-line CSV. You need three columns out of it, only the rows from last Tuesday, with one misspelled city name fixed throughout. You could open it in a spreadsheet, wait for the beachball, scroll, sort, find-and-replace, and export again. Or you could describe the whole job in one line of text and get the answer before the spreadsheet finishes loading. ...
Three voices argue about how to extend a Claude agent — skills-first, MCP-anchored, complementary. The debate dissolves once you separate procedural knowledge from live reach.
A walkthrough of the surface an SDK exposes, in seven parts, with idiomatic Go for every piece: the client constructor with functional options, resource groupings that mirror the caller’s mental model, typed requests and responses, typed errors with request IDs and errors.As, first-class pagination via iter.Seq2, streaming and retries and instrumentation, and the docs/samples/changelog trio that finishes the surface. Companion to the SDK fundamentals post.
An eight-point rubric for judging whether an SDK is any good, each criterion paired with a real currently-shipping SDK that gets it right — Stripe (idiomatic hand-crafted per language, typed errors with request IDs, exceptional docs), AWS SDK for Go v2 (boring auth via the credential-provider chain, paginators as a distinct type), Anthropic and OpenAI (sensible retry defaults, streaming), Stainless-generated SDKs (types from a single spec), GitHub Octokit (honest versioning). Followed by the two-sided value proposition — what building on those axes actually buys for the vendor and the caller.
The question is almost never ‘how do we add this method to the SDK?’ — it’s ‘where should this capability live?’ A companion to the SDK fundamentals post, covering capability-first SDK design (with the caller-sketch discipline as the practice), the five-way routing framework for new capability requests (pure SDK / additive API change / new endpoint / new API service / push down for generation), signals for extending an existing API vs. creating a new one (with Stripe Terminal and AWS Lambda as named examples), and the version-coordination process when a capability requires changes on both sides.