"AI-assisted developer" is a phrase that means different things depending on who's saying it. Here is what it concretely means in my day, broken down so you can decide whether any of it is useful to copy.
The IDE: Claude Code
Claude Code is my primary editor. Not VS Code, not Cursor, not a hand-rolled CLI. The single most useful thing about it for the way I work is that it runs entire projects under one context — so when I ask it to refactor a Zod schema, it remembers every file that imports the schema and updates them in the same pass.
I pin different models for different tasks:
- Opus 4.7 with the 1M context window: codebase-wide refactors, architecture decisions, any task where holding the whole shape of the project in head matters. Costs more per token, pays for itself when the alternative is a multi-day refactor.
- Sonnet 4.6: the daily driver. Single-file edits, route additions, content tweaks, test writes. The price-performance sweet spot.
- Local Ollama (Qwen2.5-Coder-32B): mechanical fixes — typecheck errors, lint cleanup, JSON transforms. Free, fast, no network, never the bottleneck.
- gpt-4o-mini through the AI SDK: runtime chat endpoints in client apps. Not for editing; for inference inside shipped products.
The factory picks the cheapest capable tier per task class. I don't make the choice manually most days; a router does.
MCPs are the surface area
If Claude Code is the cortex, MCPs are the limbs. The ones I actually use, in approximate order of how often they fire:
- Firecrawl: site mapping, content extraction, competitive scrapes. When I take on a new client I crawl their old site (or their competitor's) into the project so the model can ground every copy decision in real text.
- Playwright / chrome-devtools-mcp: headless QA, accessibility tree inspection, real-Chrome screenshots. Tests run against the real DOM, not a synthetic snapshot.
- Sharp Image: bulk image conversion (PNG → WebP, hero AVIFs, dimension extraction). Saved ~80% bandwidth on one project's gallery alone.
- GitHub: PR review automation, commit-message generation. The default for VCS work.
- Vercel: deploy, inspect, log retrieval. The release flow runs through it.
The factory pre-checks each MCP's health before any session starts. A dead MCP gets routed around silently; a critical one missing aborts the run with a clear error. That alone has saved me hours of debugging "why didn't the script ping IndexNow."
The release gate
pnpm release on every project is:
``bash
vercel --prod --yes && pnpm indexnow && pnpm verify-canonical
``
vercel --prod deploys. pnpm indexnow notifies Bing, Yandex, Naver, Seznam that URLs changed — Bing's index also powers ChatGPT search and Microsoft Copilot, so this directly accelerates AI visibility. pnpm verify-canonical fetches the live URL, follows redirects, and exits non-zero if the canonical declaration doesn't match the served host. That last check exists because two of my sites shipped with the bug; it now cannot ship again unobserved.
Pre-merge, every PR runs Lighthouse mobile. Performance ≥ 95, Accessibility = 100, SEO = 100. Below threshold = block. I've cut releases for less.
Where the model still doesn't think
Two places I refuse to outsource:
1. Architecture decisions. Whether to use Supabase auth or roll my own with Auth.js. Whether to put the RAG store in Sanity or in a flat JSON file checked into git. These calls have downstream taste consequences the model can't see ahead. 2. Voice. Every copy edit goes through me. The model proposes; I rewrite. Brand voice across my three sites looks consistent because one person owns it, not because Claude has a "professional but warm" preset.
Everything else — typing, transforming, validating, deploying — is automated.
What this isn't
It's not "AI ships my code." The model lowers typing cost. The human still owns the system shape. The day someone confuses those two is the day they ship something they can't debug.
If you want the actual setup file by file, the /uses page lists every model, every MCP, every backend service in current rotation.