Most AI coding content you see online is theatre. Someone types a one-shot prompt, the model spits out a polished React component, and the post goes viral. Then those same people open a real codebase, ask Claude to "fix the bug", and watch it confidently rewrite three files that weren't broken.
This is the workflow I actually use to ship code at 47 Industries. We run it across BookFade, MotoRev, the 47 Industries platform itself, and a stack of client projects. It's boring on purpose.
The CLAUDE.md is the workflow
The single biggest leverage point — bigger than any prompt — is the CLAUDE.md file at the root of your repo. It's the first thing Claude Code reads. Get this right and you skip 80% of the bad output.
Mine has six sections:
- Project overview — one paragraph. What this is, who uses it, where it deploys.
- Important rules — the non-negotiables. "A task is not done until it's pushed." "No emojis." "Read before write." Boring rules. Make them impossible to misread.
- Tech stack — versions matter. Claude will pick the wrong React syntax if it doesn't know you're on 19.
- Conventions — TypeScript strict, no
any, named exports, where tests live. - Commands — dev, build, test, typecheck. Claude needs the actual command for your project.
- What "done" means — typecheck clean, tests pass, manually verified, committed, pushed. Define it explicitly or it'll skip steps.
That's it. Five hundred words. It does more for code quality than any clever prompt.
Read before write
The single instruction I find myself repeating: read the file before you change it. Claude is a confident author and a lazy reader. If you don't force it to read, it'll hallucinate the contents.
I keep a saved prompt for this:
Before you write any code, list every file you need to read to understand this task. Read each fully. Then state back to me, in 3-5 bullets, what you understood. Wait for me to confirm before writing code.
This sounds excessive until you skip it on a tricky task and watch Claude rewrite the wrong file.
Plan, then execute
For anything more than a one-line change, I make Claude write a plan first. Files to touch. Files to create. Approach in 2-4 sentences. Risks.
Then I read the plan. This is where the value is. 30 seconds of reading saves a 20-minute debugging detour. If the plan is wrong, the code will be wrong.
Half the time the plan reveals that Claude misunderstood the task. We fix that before any code is written.
Surgical diffs only
I tell Claude, every session: smallest possible change. Don't refactor surrounding code. Don't rename variables. Don't "improve" formatting. If you spot something to clean up, list it at the end — separately.
This isn't pedantry. Big diffs hide bugs. Small diffs are reviewable. A 30-line PR is a 5-second review. A 300-line PR is a 30-minute review and you'll skim half of it.
Verify before done
Claude will tell you the task is complete before it's complete. Not because it's lying — because "complete" is fuzzy in its head.
So I make "done" mechanical. Before claiming a task is finished, Claude must:
- Run the typechecker. Paste the output.
- Run the tests for the affected module. Paste the output.
- For UI: open the dev server, describe what they see.
- For an API: hit the endpoint with curl, paste the response.
If any check fails, fix the underlying issue. Don't skip the step.
The unsexy truth
The tweets you see ("Claude wrote my entire SaaS in 4 prompts") aren't lying. They're just leaving out the part where the SaaS doesn't actually work in production.
Shipping with AI looks a lot like shipping without AI. Read the code. Plan the change. Make a small diff. Test it. Push it.
The leverage from AI isn't that it skips those steps. It's that it does each one 10x faster — once you set up the discipline.
Where to go next
If you want the actual prompts I use daily, the 47 Industries Prompt Library has them. The full set, including premium prompts and CLAUDE.md templates for different stacks, is part of the Vibe Coding Playbook.
