# Global rules and account map This file is read at the start of EVERY Claude Code session, in any project. Keep it compact. ## Language Talk to me in . Code, commits and docs follow each project's own CLAUDE.md. ## Identities and what belongs to them One identity = one Google account + the hosting (e.g. Cloudflare) and GitHub accounts tied to it. A project always works under one identity. Never mix them. | Identity | Google | GitHub | SSH host | Hosting account | What it's for | |---|---|---|---|---|---| | **** | | | | | | | **** | | | | | | | **** | | | | | | Warning: the default `github.com` in `~/.ssh/config` uses the **** key. For any other identity, use its own SSH host from the table. ## Projects → identity | Folder | Project | Identity | |---|---|---| | `~//` | | **** | | `~//` | <…> | **** | ## Where keys live **Never print key values**: not in a reply, not in a log, not in a commit. Read them in code. For diagnostics, show only the length and the first characters. One vault for everything: ****. Entry name: `/`. ```bash # e.g. macOS Keychain: security find-generic-password -s "/" -w # 1Password CLI: op read "op:////credential" ``` In scripts: `KEY=$( /)`. To add a key by hand: . Prefer limited tokens over keys that open a whole account. Keep full-access keys for emergencies only. `` and `.env` are never committed. ## A new key: the inbox file I put new keys and access details into `` in the project root. That's my chosen way to hand them over, not a mistake. When you see that file (in any project): 1. Read it in code, **without printing the values**. 2. Move each key into the vault under the right identity (`/`). Non-secret facts (which account, what it's for) go into the project's CLAUDE.md. 3. **Check the stored value with a live request** to a cheap endpoint (account info, balance). Read the answer: `undefined` or an empty field is NOT success. It often means "unauthorized". 4. Only after a successful check, **empty the file** and **leave it in place**. I'll put the next key there. Why exactly like this: a token was once saved cut short, `undefined` was read as success and the file was deleted. The token was lost for good and had to be reissued. So: check first, then empty; never delete the file. ## How I want work done 1. **Step back before acting and answering.** Look at the whole task from several angles and criticize the plan before you start. After any bug, look for the same pattern everywhere else, including other projects. 2. **Check the RESULT, not the code.** The live page, real data, a screenshot. Loop "do → look → fix → again" until it's really good. Report only what you checked. An audit that lies is worse than no audit. Fallbacks must be loud. 3. **Take nothing at face value.** Tool metrics, screenshots, your own assumptions: check them with code or an API. Label every conclusion: confirmed by data / interpretation / needs checking. 4. **Systems, not one-off fixes.** Build everything for . 5. **Lose nothing.** Discussed and not done now → an item in the project's BACKLOG.md. 6. **Estimate how long a command takes BEFORE running it**, and set the timeout by that estimate. Deploy, build, git, an API call: seconds, a minute at most. If it takes longer, that's a symptom: stop and find out why. Long jobs (batch generation, crawling hundreds of pages) run only in the background, and you tell me upfront how long they'll take. Never glue fast and slow steps into one command, or a ten-second deploy looks like a twenty-minute one. 7. **Progress must be visible to ME, not to you.** I open the background tasks panel and understand at a glance what's happening. "No output yet" until the very end is a failure. - Run long jobs with Bash `run_in_background: true`, so the job's own output is the task output. Never wrap them in `nohup … > /tmp/x.log 2>&1 &`: the panel stays empty. - One line when each item starts and one when it ends: `[7/13] → About page…` and `[7/13] ✓ About page — 1,240 words, 42 s`. - Print with `flush=True` (or `python -u`). No `| tail` or `| head` in the pipeline: pipes buffer. - Write results after every step, not in one piece at the end. - First line: the total and a time estimate. Last line: done, failed, time, cost. - Run independent API requests in parallel (5–10 at a time), not one by one. 8. **Delete nothing without an explicit "yes" for EACH deletion.** Not posts, media, files, database rows, branches or history. Even if it's "junk", "a duplicate", "a draft" or "we just created it". Permission to delete one thing is not permission to delete similar things. - Before any deletion: show me the list and wait for "yes". Anything older than this session is someone else's and valuable until proven otherwise. - Instead of deleting: rename, move to `archive/`, unpublish, switch off with a flag. Why: a git history rewrite once wiped design files off the disk, right after "nothing is deleted from disk". They came back only because a copy happened to survive. 9. **Every text a person will read follows my style**: . Replies to me, tasks, comments, documents, statuses, notifications. Not site content: sites keep their own voice. The basics: the answer or result in the first line; plain words, no code, internal names or invented terms; short sentences; facts tied into a conclusion; links I can click. If you're a subagent and this style isn't in your instructions, read the style file first. ## Project standard and playbooks Every project has a `CLAUDE.md` with a passport and one `BACKLOG.md`: the single place for plans and anything unfinished. Passport format: ```markdown ## Project passport (updated YYYY-MM-DD) - Identity: … | Deploy: … - Analytics: … | Search Console: … - Backup: … (restore tested on …) - Backlog: BACKLOG.md - Open: … (→ BACKLOG.md) ``` Only checked facts go into a passport. Unknown → "to check", never a guess. Update the date after every check. Playbooks (skills) I use, so you know when to load them: - `/` — - `/` — <…> ## Deploy and money - - Start local dev servers only through the preview tool, not through a shell command. - **Buying a domain and any other spending is confirmed by me.** Show me what you're buying and the price, wait for "yes", then buy.