The Claude Code Workflow You Can Copy

AI & Data Science
, ,
1

1400×783 286 KB

The Claude Code Workflow You Can Copy

Chris Dunlop

This is the workflow I use when I want useful code fast without creating a mess I’ll regret later. It’s opinionated, repeatable, and built around Claude Code’s strengths (agents, short feedback loops, clean diffs).

Not a Medium member? Keep reading for free by clicking here.

0) One-time setup

/docs/01-scope.md # single source of truth for goals & constraints
/docs/02-decisions.md # architecture decisions log (ADR-style, <10 bullets)
/docs/03-tasks.md # running task list with checkboxes
/PLAYBOOK.md # “how this project works” for future-me / team
/CONTEXT.md # compressed summary Claude reads first
/src/ # code
/tests/ # tests (even if thin)
/scripts/ # one-off utilities (seed, migrate, import)

Keep CONTEXT.md tiny (≤ 200 lines). Describe modules, data shapes, APIs, and non-negotiables. Update it after every milestone.

1) Create five focused agents

In Claude Code: claude update → /agents.

Make these once; reuse everywhere:

mvp-planner

Role: Turn a vague goal into a MVP with a crisp scope.
Always output:

  • “We will NOT build” list
  • Risks with mitigations
  • JSON backlog [{id, title, acceptance}]

ui-stylist

Role: Restyle components to a design token set (typography, spacing, colour).
Constraints: no library swaps without approval; return a patch diff only.

bug-fixer

Role: Reproduce, write a failing test, fix, and return a minimal patch.
Always include root cause in 1–2 sentences.

modular-architect

Role: Propose directory structure, boundaries, and interfaces.
Output an ASCII module map + reasons for each boundary.

reviewer-readonly

Role: Code review with NO edits.
Return: inline comments, risk ranking (High/Med/Low), and a “merge/no-merge” call.

Claude will auto-route to these when your prompt matches. If it doesn’t, call the agent explicitly.

The 35-minute build loop

Run this loop until you ship. Don’t stretch it. The constraint keeps quality high.

Minute 0–5 — Frame

  • Edit /docs/01-scope.md (what we’re doing next, acceptance).
  • Update /CONTEXT.md if structure changed.

Minute 5–20 — Build

  • Prompt Claude with one atomic task (see templates below).
  • Ask for patch diff, not prose.
  • If you get noise, ask for the same patch but smaller (“touch max 3 files”).

Minute 20–30— Test

  • Request a failing test first, then a fix.
  • Run locally, keep logs, trim scope if needed.

Minute 30–35— Commit & compress

  • Commit: feat: add invoice export (csv)
  • Append one line to /docs/02-decisions.md if a decision happened.
  • Rewrite /CONTEXT.md deltas (keep it under 200 lines).

Repeat.

3) Prompt templates that actually work

A. New feature (atomic)

Goal: [one sentence]
Constraints: [stack, libraries, patterns that must remain]
Touch budget: max 3 files
Return: unified PATCH DIFF only + brief rationale (≤3 bullets)
Use CONTEXT.md to preserve architecture.

B. Styling pass (no logic changes)

Apply design tokens (font scale, spacing, radius).
No JS behaviour edits.
Return patch diff for .tsx/.css only.
If tokens missing, create tokens.ts and refactor to use it.

C. Bug fix with safety rails

Reproduce bug: [steps]
Write failing test first in /tests/[name].spec.ts.
Then propose smallest fix.
Return: test diff + code diff.
Explain root cause in 2 sentences.

D. “Too big” response recovery

Your patch is too large. Split into sequential patches of ≤50 lines each.
Return only PATCH 1 now. I will apply and ask for PATCH 2.

4) When you get stuck, use the Stuck Ladder (in order)

  1. Narrow the ask (one file, one function).
  2. Add a concrete example (input/output JSON).
  3. Duplicate the page and reframe the prompt (fresh context often routes the right agent).
  4. Switch the persona (call modular-architect to reshape before coding).
  5. Cut the surface area (feature flag, ship a smaller slice).

If you’re still stuck after 20 minutes, you’re solving the wrong problem. Go back to /docs/01-scope.md.

5) Scaling to large codebases

  • Compress first. Keep /CONTEXT.md up to date; ask Claude to regenerate it when modules shift:
  • Summarise repo by modules, public interfaces, data contracts. ≤ 180 lines. Highlight coupling hotspots. Output to CONTEXT.md (overwrite).
  • Work in slices. “Only touch /src/billing/*.” Claude respects fences when you declare them.
  • Batch refactors. Lock logic, run the ui-stylist or linting changes as separate patches to avoid noisy diffs.
  • Index hotspots. Maintain a short list at the top of /CONTEXT.md: “Here be dragons” with file paths.

6) Ship ritual

  1. Reviewer-readonly agent must say “merge” (no edits).
  2. Tests run locally (even if minimal).
  3. Docs bumped (one line in decisions + any API shape changes).
  4. Release note (3 bullets: what changed, risk, rollback).

7) Example: feature from nothing

Brief: “Export invoices to CSV with filters.”

Start

  • Scope: add “Export CSV” button, filter by date/status, job finishes.
  • modular-architect: adds /src/export module with interface.
  • Feature prompt (atomic): returns patch for service + controller + button stub.
  • bug-fixer: writes failing test for timezone edge case, then patch.
  • Commit, decisions line: “CSV uses UTC; display converts to local.”

8) FAQ in one-liners

  • Why patch diffs? They anchor Claude to change less, so you get safer edits. Here is an example. This is also why a tool like v0 or ChatGPT is slower for coding with AI as they always regenerate the entire file.

1250×956 38.7 KB

  • Why CONTEXT.md? It’s the guardrail. Claude reads it; you maintain it.
  • Why 35 minutes? Long enough to finish a slice, short enough to avoid scope creep. If you can’t ship features at a regular cadence your scope is too big.
  • Why five agents? More agents does not equal more speed. These five cover 95% of work. I found that if you have too many it wasn’t routing to the correct one.

9) Copy-paste starter pack

Add these to the top of every Claude session:

House Rules:

  • Return patch diffs, not prose.
  • Respect /CONTEXT.md constraints.
  • If unsure, propose 2 options with trade-offs (≤80 words).
  • Keep changes surgical: max 3 files unless I expand scope.
  • If more than 3 files tell me why and what

First message in a new project:

Read CONTEXT.md and 01-scope.md.
Propose a clear MVP plan with a JSON backlog.
List what we will NOT build.
Then wait.

I like saying list what we will not build.

This is the whole thing. Keep the docs small, keep the tasks atomic, demand patch diffs, and let the agents do the jobs you hired them for.

Before you go

  • I now have a Substack and it’s a great way to connect. I post one unique piece a week there every Sunday New Zealand Time. Come join!

Realworld AI Use Cases
](https://medium.com/realworld-ai-use-cases?source=post_page---post_publication_info--6265009df76d---------------------------------------)

[

Published in Realworld AI Use Cases

](https://medium.com/realworld-ai-use-cases?source=post_page---post_publication_info--6265009df76d---------------------------------------)

3K followers

Last published 1 day ago

I wanted to have a publication that helped share tips and tricks from my consultancy where we implement ai for companies like the Olympic team, the Stock Exchange and other B2C companies.


Chris Dunlop

[

Written by Chris Dunlop

](https://medium.com/@chrisdunlop_37984?source=post_page---post_author_info--6265009df76d---------------------------------------)

6.6K followers

39 following

I’ll help you code with AI. Cursor tips & Business Strategy. I run a company doing AI for the All Blacks, Olympic Team & the Stock Exchange www.cubdigital.co.nz

2

too much words, where’s the repository?

3

He never provided a repo. I’ll ask him when I am back to Santiago.