Client stories: the question each business couldn't answer, what we built and what changed.
All workTools we built, for clients and for ourselves, when nothing we could buy did the job.
AI moves fast. Here's how we use it in our own work, and where to start in a business your size.
The AI hub
The AI we trust most to change our code. It plans the work, builds the hard parts, briefs Cursor's agents on the rest and reviews what comes back. A person decides what goes live.
It reads the repository, makes the change on a branch, runs the tests and opens a pull request for a person to read.
A CLAUDE.md file in each repository holds the house rules. Ours is generated from the same source as Cursor's, so both tools behave the same way.
Our own checks run before it acts. One blocks email, chat and publishing unless someone asked for that in the same message.
A lead session splits the work across agents, each in its own copy of the repository, then reviews and merges what they send back.
It briefs Cursor's agents on clear-cut jobs, watches them, steers them mid-run and reviews their pull requests.
Terminal, VS Code or Cursor, JetBrains, the desktop app and the browser, all with the same settings and tool connections.
Pick
The AI we trust most to change our code. We run it inside Cursor, with our own rules and a person reading every pull request.
Last known verdict
Another coding agent that sticks to our rules and review habits better.
Claude Code's the AI we trust most to change our code. It runs as a panel inside Cursor, our pick for now. It reads the repository, makes the change on a branch and opens a pull request on GitHub, and a person reads it before anything goes live. Which models do the work is on the Anthropic page.
We write our rules once, for Cursor. A generator turns them into Claude Code's own rules, agents, skills, hooks and tool connections, and a drift check (--check) fails in any repo where the two disagree. Both tools follow the same house rules: British English, and no email or chat message goes out unless someone asked for it in that message.
In our operations repository, a hook blocks Claude Code's outbound email and publishing unless a person asks for that action there and then. A pre-commit hook blocks any commit with credential-shaped text in it. AI builds it, AI reviews it, the checks run, and a person decides what goes live.
This website (its lead form posts to n8n) and our own tools, like TechStack Analyser.
As of 10 October 2026: 26 rules and a 1,001-line generator in our admin repository, 37 of the 45 repositories in our GitHub organisation carrying its CLAUDE.md, and 280+ pull requests merged on this website's repository since 22 September (780+ commits).
There's more on the wider setup on our AI page, and we set up the same thing for other teams: AI setup.
We pay for both and use them together. Pick the job and we'll tell you which one gets it.
| What matters | Claude Code |  CursorOur pick |
|---|---|---|
| Examples | The same jobs, using Claude allowance we'd rather keep for harder work | A bug fix, a small feature, a refactor, a docs update |
| Models | Claude's Sonnet and Opus, on our own Claude plan | Cursor's own models, like Composer, on the Cursor plan |
| Where it runs | Your terminal, editor, the desktop app or the browser | Cursor's cloud or a job on the laptop, each on its own branch |
| Long runs | Web and desktop sessions can run on without you too | Cloud agents carry on after you close the laptop |
Our pick: Cursor. Fixes, small features, refactors and docs go to a Cursor agent on Cursor's own models, which come with the plan. Claude Code writes the brief and checks the result.
| What matters | Claude CodeOur pick | Cursor |
|---|---|---|
| Best at | Working out the approach, then building and checking it | Doing a clear brief quickly |
| Big jobs | A lead session splits the job across sub-agents and merges what they send back | One agent per job, each on its own branch |
| In our setup | Writes the briefs, watches the agents and reviews their pull requests | Takes briefs and works on them |
| Models | Sonnet or Haiku for sub-agents, Opus only for the lead session when a job needs it | Cursor's own, plus others billed on usage |
Our pick: Claude Code. Working out an approach, design-heavy pages and anything with open questions go to Claude. It also reviews what Cursor's agents send back.
| What matters | Claude Code | CursorOur pick |
|---|---|---|
| What it is | An agent for your terminal, an editor panel, the desktop app or the browser | A code editor built on VS Code |
| How you work | It makes the changes and you review the diff | You edit, and the AI helps in the same window |
| GitHub | Commits, branches and pull requests on its own | Branches and pull requests from the editor |
| Using both | It reads the same rules, generated from one source | Claude Code installs as an extension, so both share one window |
Our pick: Cursor. It's a full editor, so you see the code, the AI and the terminal in one window, and Claude Code runs as a panel inside it.
Facts are from Cursor's and Anthropic's own pricing pages and docs, October 2026, and from how we run both today. Prices change, so we've kept them out.
It writes code that looks right whether it is or not. Someone still has to read it, run it and decide it's ready to ship.
CursorWhere we run it. It also briefs Cursor's agents and reviews their work, on the same rules.
AnthropicIt's the Claude family. Sub-agents run on Sonnet at most, Haiku for simple runs, and Opus is kept for the few lead-session jobs that need it.
GitHubIt opens pull requests, and they're reviewed there.
PlaywrightIt screenshots the pages it changes, at desktop and phone width, before a person sees them.
Astro + KeystaticWe build and change this website with it.
n8nWe used it to build the form handler that sends every lead to n8n.
FigmaCan read a Figma file through its MCP server, so the agent has the layout to build from.Our in-house scanner lists the tools and tracking a website runs. A quick scan takes about 5 seconds.
A small Python script that turns a repo's Cursor rules, agents, skills, MCP servers and hooks into the files Claude Code reads, so both editors follow the same guardrails.
Rule templates for Cursor, a pre-commit hook that refuses credentials, and a script that turns your Cursor rules into Claude Code config.
How we rebuilt pacing.agency in Astro, with Keystatic as the editor and one Cloudflare Worker, in three weeks, with AI agents opening 95% of the pull requests.
+4We don't sell it on its own. Each of these services uses it alongside the rest of your setup.
Claude Code is an AI agent from Anthropic that works in your code. It reads a project, makes changes, runs commands and checks its own work. It runs in a terminal, in VS Code, Cursor and JetBrains, in a desktop app and in the browser, and it comes with a paid Claude plan or an Anthropic API account.
Straight answers on how we work, what it costs and what to expect. If yours isn't here, just ask.
Still have questions?
Talk to usCursor's an editor with GitHub sync and a good base allowance, so the code, the agent and the terminal all sit in one window. It's where Claude Code works best for us so far. That's our pick for now, and we'll say if it changes.
Not the way we run it. A hook blocks its email, chat, invite and publishing tools unless a person asks for that action there and then. That's a guard rail, not a guarantee, so we review the work as well: AI builds it, AI reviews it, the checks run, and a person decides what goes live.
No. It writes code that looks right whether it is or not, so someone still has to read it, run it and decide it's ready to ship. It makes a small team faster, but you still need someone who understands what's been built.
If you write code yourself, start with Cursor. It's a full editor, so you see the code, the AI and the terminal in one window. If you want an AI to take a job from plan to pull request, add Claude Code. It runs inside Cursor, so you don't have to choose.
We use both. Claude Code plans the work, builds the parts that need judgement and reviews the changes before they merge. Cursor's agents take the clear-cut jobs, like fixes, small features and docs, on Cursor's own models that come with the plan (check Cursor's pricing page).
We'll tell you which tools you need and which connections are missing.
We'll tell you if the spend makes sense.
