Setuproll editorial
@setuproll·The Setuproll team·Since 2026
We assemble, verify and document AI coding setups so you can copy what works without relearning it from scratch. Every build we publish is labeled as editorial, not community-submitted.
Latest articles
How we built this Claude Code setup and how to actually work with it
Our reference Claude Code rig for a full-stack TypeScript app: the exact CLAUDE.md, settings.json, .mcp.json, subagents, hooks and slash commands, plus the day-to-day loop to run them through. Steal the whole thing.
I gave Claude Code four jobs and finally stopped doing all of them
How a planner/reviewer subagent loop, Postgres + Playwright MCP, and three boring hooks turned my Next.js app into something I can leave running.
I Refuse To Pay Opus Prices: My 14-Cent Claude Code Setup
How I run Claude Code on Sonnet 4.6 with a 1M context window, two cheap subagents, and a Stop hook that stops me shipping broken code, all for about a coffee a day.
Tab is doing 70% of my typing now, and I am not going back
My Cursor setup tuned for one thing only: editing at the speed I think. Tab, Agent mode, and Figma frames straight into components.
I tune Cursor like a workshop, not a slot machine
Max mode on Opus 4.8, Context7 docs attached on every hard task, background agents for the long jobs, and rules I actually enforce in review. The Cursor setup I use when correctness matters more than speed.
The Copilot setup that survives an audit
How I run GitHub Copilot Enterprise on GPT-5.3 Codex so the coding agent moves fast and still cannot do anything our auditors would flag.
I was scared of AI coding agents. Then Windsurf told me what it was about to do.
A complete-beginner setup for Windsurf and Cascade that explains every step before it touches your files, so you actually learn instead of just clicking accept.
I Code in the Terminal With Aider and My Own Key. The Bill Is Six Cents.
A grizzled holdout's no-GUI setup: Aider, Gemini 2.5 Pro, architect mode, a git lint hook, and a config that fits in one yaml file.
I Let Cline Drive My Editor For a Week. Here Is the Config That Made It Safe
How I wired Cline's Plan/Act split, checkpoints, and Playwright MCP into one VS Code setup I actually trust to run on its own.
I gave Claude Code a real browser and a strict no-deleting-flakes rule
My Claude Code + Playwright rig writes the test first, quarantines flakes instead of nuking them, and only reports specs I actually have to fix.
The agent doesn't touch my Postgres until it has read the schema
How I wired Claude Code to reason about the database first, scan for secrets before every write, and refuse to merge an endpoint without an integration test.
I let an agent run my notebooks. Here is how I stopped it from p-hacking me.
A reproducibility-first Claude Code setup with a Jupyter MCP, a leakage-aware eval reviewer, and hooks that log every experiment so the agent cannot quietly cheat.
I make Cursor build my Figma, pixel for pixel, or it does not merge
How I wired Figma, Playwright, and an accessibility reviewer into Cursor so the agent ships components that actually match the design.
I let an agent touch my Terraform. Here is the harness that lets me sleep.
An infra-as-code Claude Code setup where the model plans, a hook hard-blocks prod applies, and a Stop hook tells me what it costs before I do anything dumb.
How I let Claude Code build my RN screens without it breaking the safe area every time
My actual Sonnet 4.6 setup for shipping Expo apps solo: a screen-builder agent, a platform reviewer that nags about iOS vs Android, and a Detox smoke test on stop.
I let Claude touch a 14-year-old codebase. Here is the harness that kept me employed.
An Opus 4.8 setup that writes characterization tests first, refactors in tiny behavior-preserving steps, and hands me a behavior-diff every time it stops.
I Build the Whole Thing in 48 Hours. Here Is My No-Tests, Ship-First Claude Code Setup
A stripped down single-agent Claude Code config tuned for one thing: getting a working demo on screen before the coffee gets cold.
I code with AI for two cents a task and I am not even slightly sorry
Aider plus Gemini 2.5 Flash, bring your own free key, repo map on, auto-commit on. The whole setup, the config files, and the screenshot of a billing dashboard that still reads under a dollar.
I let Claude Code red-team my codebase, but it does not get commit rights
How I run Claude Code as a read-only SAST cell: Opus 4.8, the Semgrep MCP, a threat-modeler subagent that has to prove exploitability, and hooks that stop the auditor from exfiltrating the secrets it audits.
I let Cascade build my dbt pipeline, but only one stage at a time
How I wired Windsurf + Cascade, a Postgres MCP server, and a dbt-test pre-commit hook so an agent can extend my warehouse without quietly breaking the DAG.
I pay $10 a month for AI and I am not switching
My whole AI coding setup is GitHub Copilot, one MCP server, and a single format hook. Here is exactly how I keep it cheap and fast as a solo dev.
I let Claude loose in a 60-package monorepo. Here is the leash.
How I scope Claude Code to one package, rebuild only what changed, and review across boundaries before anything merges.
I made Cursor do the boring 80 percent of code review so I could be meaner about the rest
A blunt reviewer's Cursor setup: two review subagents on Opus, GitHub and Sentry over MCP, and rules that ban nitpicks so the bot flags the bug that pages you instead of your quote style.
I let a model draft my docs. Here is the leash I keep it on.
A Claude Code setup for writing documentation that stays accurate: Context7 for real API signatures, a fact-checker subagent that deletes anything unsourced, and link checks on every save.