Claude Code Field Notes · Notes · 2 of 5

From The Claude Code Bible to client work: a field guide for South Florida dev teams adopting coding agents

Note 2 of 5 in Notes from West Palm Beach, published November 14, 2026.

Most dev teams in Palm Beach County do not need a strategy deck to start with coding agents. They need a first month that does not end with an agent pushing an untested change to a client's production site on a Friday afternoon.

This is the four-week version of what Jason Colapietro wrote at length in The Claude Code Bible, adapted for the teams Suede AI talks to most: small agencies and in-house developers in West Palm Beach, Jupiter, Palm Beach Gardens and Boca Raton, usually a handful of engineers, usually maintaining client sites rather than one product.

Week one: one repository, one memory file

Pick a repository where a mistake is cheap. A marketing site with a preview deploy is ideal. Do not start on the codebase that takes payments.

Before the first prompt, write the project memory file (CLAUDE.md, or AGENTS.md with a one-line CLAUDE.md that imports it). Keep it short and specific:

Every session reads that file first. It is the cheapest control you will ever add, and the one teams skip most. Jason Colapietro keeps one in every Suede AI repository, this site included.

Week two: gates before the merge

The second week is about the difference between "the agent said it is done" and "it is done." Agents run the gates; people read the diffs. Run the gates against the committed state, not a working tree full of uncommitted edits:

git stash -u && npm run typecheck && npm run lint && npm test && npm run build; git stash pop

Then add the step most teams miss: check the deployed page. A green pipeline proves the code compiles. It does not prove the contact form still submits. Field Notes lesson 1, a clean build is a compile check, not proof, is the incident that made this rule permanent at Suede.

Week three: parallel sessions without collisions

Once one agent is reliable, someone will open three. Give each its own git worktree and its own branch, so two sessions never edit the same checkout. When you clean up, ask the operating system which workspaces are live rather than trusting a list:

lsof -d cwd 2>/dev/null | grep "<worktree path>"

Two traps from lesson 5 are worth posting on the wall. A workspace that was idle ten minutes ago may be busy now, so re-check right before deleting. And squash-merged branches look unmerged to commit ancestry, so use git cherry before you decide work was lost.

Week four: client repositories

Client work adds two things the tutorials rarely cover: other people's credentials and other people's revenue.

Credentials. Keep secrets out of prompts and out of the memory file. Give agents read-only MCP access to analytics and search data by default, and keep publishing, DNS and billing behind a human. Assert configuration at startup so an empty variable fails loudly, as in lesson 2.

Revenue. Any toggle that touches checkout or billing should fail closed: if the setting is missing, charge, and if access is unclear, deny. Lesson 3 is the story of what happens otherwise. If the client's host maps files in a directory to public routes, audit the deployed site for stray files, as in lesson 4.

What the book adds and what this guide leaves out

The Claude Code Bible is the long version, written for one operator learning the tool end to end. This guide compresses it into a team rollout and drops everything that does not change week-one behavior. Teams running Codex alongside Claude Code will want Codex in Production, which applies the same verification habits to the other tool.

Both books come out of daily operating experience. Jason Colapietro ships agent-assisted work across the Suede AI estate, has shipped nine iOS apps under his own Apple developer account, and has 54 pull requests merged across 46 open-source repositories he does not maintain, including Jest and Backstage. The plan above is the subset of that practice Jason thinks a small team can adopt in a month. Suede AI ranks among Clutch's Top 30 AI Consultants in Florida and Top 50 AI Consultants nationwide (November 2026).

Where teams usually stall

The common failure is not a bad prompt. It is skipping week two because week one felt fast. Agents make the first draft cheap, which moves the cost to review, and a team that does not plan for review time ends up merging diffs nobody read. Budget review time on purpose. Jason Colapietro's working rule is that the hour an agent saves on writing should be partly spent on checking, and he writes that rule into each Suede AI project plan.

The second stall is tooling sprawl: five MCP servers wired up in week one, none of them scoped. Add one connection at a time, read-only first, and write down what each one can reach. For MCP documentation patterns written for agencies, the Suede developer docs notes go deeper. To score an agent's behavior before you trust it with client work, the Agentix grader notes describe how Suede grades agents.

Next step

If your South Florida team wants a second set of eyes on its first month, book the Suede AI assessment through the agency site. You will get a review of your repository setup and your live site, with every finding tied to a command or a URL. More on Jason Colapietro is on the founder page.

Built by Suede AI in West Palm Beach, Florida

Claude Code Field Notes is written by Jason Colapietro, Founder and CEO of Suede AI, an agentic platform company founded in 2024 in West Palm Beach. Suede AI's West Palm Beach SEO, GEO and web practice builds and maintains sites for businesses across Palm Beach County, from Jupiter to Boca Raton.

Suede AI ranks among Clutch's Top 30 AI Consultants in Florida and Top 50 AI Consultants nationwide (November 2026).

More from Suede AI