How we run Claude Code as the daily development tool for a multi-repo web product. Adapt it, do not install it.
Explanations are in docs/, templates in templates/.
Turkish-language documentation index: tr/README.md
Four repos under one root:
project/
├── legacy/ in-house PHP MVC framework (multi-host)
├── api/ PHP REST API (FastRoute + Eloquent)
├── web/ Next.js frontend (App Router, 7 locales)
└── cms/ headless WordPress (content source)
Work arrives as Jira tickets. Finished tickets go to a QA tester for UAT and come back if something is wrong.
flowchart LR
J[("Jira")] -->|ticket| WS["ws CLI"]
WS -->|create| T["Task workspace<br/>worktrees + own port + own host"]
C["Claude Code session"] -->|edits| T
C -->|runs| WS
T -->|serves| P["Rendered page"]
V["page-logic-verifier<br/>+ ws browse"] -->|reads| P
C -->|delegates| V
G{"ws uat<br/>gate"} -->|pass| U[("UAT")]
WS --> G
U -->|returned| C
| Problem | Answer | Doc |
|---|---|---|
| Several tickets open at once: branch switching, server restarts, mixed-up changes | A worktree, port and hostname per ticket, from one command | 02 |
| About a third of UAT hand-overs came back, mostly for preventable reasons | A checklist from the ticket, a logical check of the rendered page instead of unit tests, a mandatory UAT gate | 05 |
| Claude rediscovered four repos every session | Layered CLAUDE.md, path-scoped rules, a code graph | 04, 06 |
| Claude's comments, commits, UI copy and translations read like AI text | A slop checker on every edit, commit message and reply | 07 |
| Requests buried in Slack and Jira mentions | A menu bar inbox where Claude keeps only mentions that need action, each with a one-line summary | 10 |
| Pages break after hand-over; the machine slows down without saying why | Daily page check with a diff against yesterday, daily clean-up, a read-only machine doctor, all reporting to one Slack DM | 11 |
| Deploys meant clicking through Bitbucket; estimates were guesses | ws deploy presses Run from the terminal; ws estimate estimates from your own history, with a backtest |
12 |
| # | File | Topic |
|---|---|---|
| 1 | 01-architecture | Parts, principles, shared vs per-task resources |
| 2 | 02-task-workspaces | Worktrees, ports, DNS, HTTPS, nginx; the ws CLI |
| 3 | 03-workflow | A ticket from start task to stop task |
| 4 | 04-claude-configuration | CLAUDE.md layers, rules, commands, agent, skill, hooks, memory, permissions |
| 5 | 05-verification-and-uat-gate | Reading pages instead of tests, the shared browser, the UAT gate |
| 6 | 06-code-graph | Code graph with graphify |
| 7 | 07-slop-check | The slop checker |
| 8 | 08-plugins-and-tools | Plugins, skills, MCP servers |
| 9 | 09-lessons | Lessons, pitfalls, where to start |
| 10 | 10-notification-inbox | Slack and Jira mentions filtered and summarised by Claude |
| 11 | 11-automation | Daily page check, clean-up, machine doctor, Slack DM |
| 12 | 12-deploy-and-estimate | Deploys from the terminal through a signed browser bridge; estimates from history |
templates/
├── CLAUDE.md root instructions skeleton
├── .claude/
│ ├── settings.json hooks and read-only permissions
│ ├── commands/start-task.md start a task
│ ├── commands/stop-task.md stop a task
│ ├── agents/page-logic-verifier.md page verification sub-agent
│ ├── rules/web.md path-scoped rule example
│ └── slop/ slop_check.py + rules.txt (7 languages)
├── skills/status/SKILL.md "where do things stand"
├── nginx/task-vhost.conf.example per-task vhost
└── git-hooks/post-commit refresh the code graph on commit
The slop checker runs as is. The rest assume our ws CLI and layout; swap in your commands and paths.
ws itself is not included (it is tied to our Jira and environment); docs 02
and 05 describe it well enough to write your own.
Diagrams are Mermaid and render on GitHub.
MIT. See LICENSE.