Shell

always use ripgrep instead of grep always use fd instead of find always use git commit --no-gpg-sign when committing always prefer brew over npm for system tools

Slack drafts

Slack drafts can't be edited, and a draft tool's "Draft created" reply is not proof the draft exists. Add -w <workspace> to any slack-drafts command for a workspace other than slackdump's current one.

• Create drafts with the Slack connector's draft tool, never slack-drafts restore: Slack keeps restored drafts on its server, but the app does not show them.
• Before writing, run slack-drafts list <channel_id>. If the destination already holds a draft, delete it with slack-drafts clear <channel_id> [thread_ts], because the connector's draft tool silently writes nothing to a destination that already has one (its reply then carries no draft_id). If clear prints "kept", the User typed or edited that draft in Slack: build the new draft on its text, then remove it with slack-drafts clear --all <channel_id> [thread_ts] (it saves a backup first).
• After writing, run slack-drafts list <channel_id> again, and report the draft only when one starting with the new text is there.
• Never make a link bold: Slack's composer carries the bold into the text that follows it.

Coding

prefer python and use uv

Markdown

adhere to DavidAnson markdownlint

Writing preference

• never use em dash or en dash
• never use "this isn't just..., it's..." trope
• NO METASPEAK: every deliverable (report, email, doc, Jira/Slack post) reads as a standalone artifact by a fresh author. Never reference prior drafts, revisions, review feedback, the writing process, or what changed ("as discussed", "updated to reflect", "per your feedback", "in this revision"). Before presenting any deliverable, self-check for these and strip them.

Public repositories

Before writing a person's name, a machine path, an employer, or an internal host into any file under git, check whether the repository is public. Do it without credentials, so the answer does not depend on which account is logged in or which org owns the repo:

url="$(git remote get-url origin | sed -E 's#^(ssh://)?git@([^:/]+)[:/]#https://\2/#; s#\.git$##')"
curl -s -o /dev/null --max-time 10 -w '%{http_code}\n' "$url"

200 means public. 404 or any 3xx means private or gone. Anything else (000, 5xx, 429) means unknown: ask the User before writing.

If it is public, write every file so a stranger can use it:

• Refer to the person as "User", never by name, and use they/them.
• Use generic paths: ~/..., $HOME, or a configurable variable. Never hardcode /Users/<name>/....
• Never name an employer, an internal host, or an internal repository path.

Research & Verification

Before guessing how a third-party service behaves, verify first. In priority order:

1. Live state (account/instance-specific): query the actual API, run the CLI, read the real config/source. Never assume runtime state.
2. Documented behavior of libraries/SDKs/APIs/CLIs (WorkOS, Slack, Homebrew, etc.): use Context7.
3. Niche/self-hosted tools Context7 won't have (go2rtc, HA integrations): WebSearch/WebFetch, or read the tool's own source.
4. Provide links when posting, e.g. Jira Service Management, so Users can see the data, logs, code, documentation, etc.

Show the source you verified against before proposing a fix.

External memory hub

A private hub repository per identity, built from the external-memory/ pattern in this repo, holds the current state of every live initiative. The rules files the hub ships name its path and ritual: a SessionStart hook pulls it and loads INDEX.md, durable facts are written with memq as they come up: what clearly matters is saved at once without asking (memq add --sure, undone with memq undo), the agent asks in the turn only when unsure, and only changes the user could not answer at the time wait for /memq or the session's exit. The work identity has its own hub and the two never mix.