All Insights
#SpecDrivenDevelopment
August 2026

Sunday Coffee & Code: Spec Driven Development in Claude Code Web

I spent today working on getting Spec Driven Development via Speckit, deployed into an existing repo to use via Claude Code Web. The premise seems simple: write the specification first, treat it as the thing that actually matters, and let the code be derived from it. My worry being, that not capturing the sort of detail that SDD drives leaves coding agents room to make poor decisions, which can be expensive. They will happily build the wrong thing, correctly, very fast :)

By Steve Harris

I spent today working on getting Spec Driven Development via Speckit, deployed into an existing repo to use via Claude Code Web.

The premise seems simple: write the specification first, treat it as the thing that actually matters, and let the code be derived from it. My worry being, that not capturing the sort of detail that SDD drives leaves coding agents room to make poor decisions, which can be expensive. They will happily build the wrong thing, correctly, very fast :)

Spec Kit (https://github.com/github/spec-kit) formalises the appropach into phases. First a constitution: your project’s non-negotiable principles. Then a specification - what and why. Then a clarification pass that forces the ambiguities into the open. Then a plan, reconciled explicitly against the constitution. Then tasks traced back to requirements. Only then, implementation.

Each phase gates the next. That’s the whole idea. Discipline you can’t skip past because you’re in a hurry.

𝗧𝗵𝗲 𝘄𝗿𝗶𝗻𝗸𝗹𝗲

Spec Kit installs through a Python CLI. I wanted to run it in Claude Code’s web client - no terminal, just a browser, a sandbox, and a branch handed back when it’s done.

Looks like the CLI is only an installer. Everything the workflow does at development time is markdown and scripts sitting in your repository. Install once, commit them, and the browser agent reads them like any other file.

So I pointed Claude at a real repository - existing code, existing documents, existing opinions - brownfield, not a clean slate. Quite a bit of back and fore with Claude on getting this in place and working - 72% of my session consumed - it’s in now and working. First step is done - constitution is built, onto the next phases.

Everything the day taught Claude and I went back into a single bootstrap prompt I can point at the next repository, or a brand new one, one that does its own reconnaissance and halts when reality contradicts its assumptions rather than improvising past them.

𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆𝘀

  • Speed without a specification is just faster rework.
  • Any check that can pass without evidence eventually will.
  • Retrofit into the messy repository - you learn more.

Want to Discuss This Topic?

Steve is always happy to have a direct conversation.