Edward Harker

How I use Claude Code

Recently I was chatting with a friend about how we both use agentic coding. I thought it would be fun to get Claude to summarise how I use it based on my session history, so I could share concrete details instead of just talking about it abstractly. So here’s how I’m currently (May 2026) using Claude Code to drive software development.

Parallelism with worktrees and cmux

I use cmux to manage many Claude sessions side-by-side, each in its own git worktree. Several features and bugs in flight at once instead of one at a time — all visible and switchable from one place.

/start-task — the front door

A custom skill that takes a ticket ID, a description, or nothing, and walks the session from “I want to do X” to a fully formed plan, ready to code:

  1. Asks for context and whether to auto-/ship at the end (99% of the time, yes)
  2. Pulls or creates a beads ticket
  3. Claims it, marks in_progress
  4. Cuts a claude/<topic> branch
  5. Drops into plan mode — Explore agents map the code, a Plan agent drafts the approach, I confirm
  6. Optionally chains into /ship

/ship — the back door

One command runs the entire close-out. Strict 10-step flow, no stopping between steps:

  1. /simplify — clean the diff, drop dead code
  2. Quality gates — detects what changed and runs only those tests, lints, and builds
  3. Terms/Privacy check — surgical edits to legal pages if the diff touches data collection, sub-processors, retention, etc.
  4. /update-kb — sync the knowledge-base docs
  5. Bug postmortems — reads docs/kb/bug-postmortems.md, scans the diff against past failure patterns to catch regressions before the PR opens, and if this PR is itself a bug fix, appends a new postmortem entry (Symptom / Root cause / Fix / Rule of thumb)
  6. Commit + push
  7. gh pr create with a templated summary and test plan
  8. /review-pr on the PR it just opened — auto-applies high-confidence findings, asks me about subjective ones
  9. Closes the beads ticket
  10. Commits the ticket-state change and pushes again

Typical loop: /start-task <id> → confirm plan → it codes → /ship → PR open, reviewed, ticket closed, branch clean. I supervise rather than drive. The points where I input are writing the ticket, approving the plan and reviewing the PR.

/insights

Generates a report analysing my Claude Code sessions — useful for spotting which skills I lean on, where time goes, and where the workflow could tighten up. Run roughly once per week.

Customization

A few things I’ve layered on top of the defaults:

  • Beads for ticket tracking.
  • A knowledge base of markdown files with deeper context on different aspects of the project.
  • RTK — a CLI proxy that cuts bash token usage by 60–90%.