jsahani/already-told-you

Already Told You: automate one job

Work through one job at a time, in plain English, and confirm each step with the user. Nothing outside ~/already-told-you/ is created or changed without a yes.

1. Pick the job. Read the newest ~/already-told-you/<YYYY-MM-DD>/findings.json (Glob). None → "Run /already-told-you:run first" and stop. Show section D as a numbered list: title, hours/week now → minutes after, suggested how. Use $ARGUMENTS if it names one; otherwise ask which.
2. Agree the shape. Restate the job as: trigger (when / who starts it) → what Claude does → what comes out → where the human approves. Confirm the automation type using the options below; change it if the user prefers. Ask only what is missing (which folder, which app, what time, where the output should go).
3. Make the skill. Every automation runs through a skill. If section C drafted one for this job, start from <report>/skills/<name>/SKILL.md; otherwise write a new draft there with the same rules (frontmatter name + description quoting the user's words; numbered steps; last step = approval). Show it, refine it with the user, then install it with the bundled installer, never by writing into a skills folder yourself: ask where (~/.claude/skills = all projects, ${CLAUDE_PROJECT_DIR}/.claude/skills = this project), run python3 ${CLAUDE_PLUGIN_ROOT}/scripts/install_skill.py --check --src <report>/skills/<name> --dest <dest>, and after a yes python3 ${CLAUDE_PLUGIN_ROOT}/scripts/install_skill.py --install --src <report>/skills/<name> --dest <dest> (it asks for permission and never replaces an existing skill unless the user agrees to --force). The approval step: if the job sends, publishes, posts, pays or deletes, the skill prepares the result and stops. Scheduled runs have nobody to ask, so in a scheduled run the skill only writes a draft to a folder (or a draft in the app, e.g. an email draft) and the user approves it later. Never let a scheduled run send.
4. Connect / schedule it - only what the chosen type needs:
   • skill + MCP: name the app connection the job needs. If it is not connected, tell the user how (claude mcp add ... from the app's docs, /mcp to check, or a claude.ai connector). Never paste secrets into files or chat; API keys belong in the user's own secret store or environment.
   • schedule (desktop): give step-by-step clicks for the Claude Code Desktop app → Routines → new local task, with the exact prompt to paste (/<skill-name> plus any detail), the schedule, the project folder, and the permission choice. Remind: it runs only while the app is open and the computer awake.
   • schedule (routine): give the exact line to type, e.g. /schedule every Monday at 9am, run /<skill-name> ..., and say it runs in the cloud on a fresh copy of a GitHub repo with connectors - no local files.
   • schedule (cron) on macOS: fill in the launchd template below. Find the claude path with command -v claude (ask before running). Write the filled plist to ~/already-told-you/automation/<label>.plist first and show it. After a yes, copy it to ~/Library/LaunchAgents/ and load it with launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/<label>.plist (each command asks for permission). --allowedTools in it must list only what the job needs - never --dangerously-skip-permissions. On Linux, show the equivalent crontab line for the user to add with crontab -e.
5. Try it once if the user agrees: run the skill now by hand and show the draft it produced.
6. Finish. Summarise what was set up, where the approval happens, the time it should save (hours/week now → minutes after, from the report), and exactly how to switch it off.

---

Automation options (checked against code.claude.com/docs, October 2026 - use only these)

Pick the lightest option that fully does the job. The how field takes one of: skill, skill + MCP, schedule (desktop), schedule (routine), schedule (cron).

1. skill - the job is the same recipe every time, started by the user. A skill in ~/.claude/skills/<name>/ (all projects) or <project>/.claude/skills/<name>/ (one project) turns a 15-minute explain-and-steer into one /<name> command. Usually also drafted in section C.
2. skill + MCP - the recipe needs another app (email, calendar, Notion, Slack, a database...). The skill calls that app through an MCP server. The user connects it with claude mcp add ... or /mcp, or uses a claude.ai connector. Only name a specific app if the history shows they use it.
3. schedule (desktop) - it should happen on a rhythm and needs local files. Desktop scheduled tasks in the Claude Code Desktop app (Routines page → new local task) start a new session at the chosen time and frequency. Runs on the user's machine, only while the app is open and the computer is awake; permissions are set per task.
4. schedule (routine) - it should happen on a rhythm and does not need local files (it works on a GitHub repo or through connectors). Routines run on Anthropic's cloud even when the laptop is closed; create one with /schedule in Claude Code (e.g. /schedule every Monday at 9am, ...) or at claude.ai/code/routines. Research preview, Pro/Max/Team/Enterprise plans, minimum interval one hour.
5. schedule (cron) - fallback without the Desktop app: macOS launchd (or cron) runs claude -p "/<skill-name>" headless at a set time, with --allowedTools listing only what the job needs.

Not an option for a recurring job: /loop and session cron tasks only run while a Claude Code session is open, and expire after seven days.

Approval rule for every automation: anything that sends, publishes, pays or deletes stops for one human approval step (the run prepares a draft or a preview; the user says yes before it goes out).

---

launchd template (schedule (cron) on macOS)

Fill in every {{...}} and save it as <label>.plist:

• {{label}} - e.g. com.already-told-you.weekly-client-update
• {{claude}} - absolute path printed by command -v claude
• {{skill}} - the installed skill name
• {{tools}} - only what the job needs, e.g. Read,Write(~/drafts/**),Bash(git log *)
• {{project}} - absolute project folder the job works in
• {{weekday}} / {{hour}} / {{minute}} - 0 = Sunday; hour 0-23
• {{home}} - the user's home folder written out in full (launchd does not expand ~)

The first line of the file must be the XML declaration, followed by the Apple plist DOCTYPE line (PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd"), then:

<plist version="1.0">
<dict>
  <key>Label</key><string>{{label}}</string>
  <key>WorkingDirectory</key><string>{{project}}</string>
  <key>ProgramArguments</key>
  <array>
    <string>{{claude}}</string>
    <string>-p</string>
    <string>/{{skill}} (scheduled run: prepare drafts only, never send, publish, pay or delete)</string>
    <string>--allowedTools</string>
    <string>{{tools}}</string>
  </array>
  <key>StartCalendarInterval</key>
  <dict>
    <key>Weekday</key><integer>{{weekday}}</integer>
    <key>Hour</key><integer>{{hour}}</integer>
    <key>Minute</key><integer>{{minute}}</integer>
  </dict>
  <key>StandardOutPath</key><string>{{home}}/already-told-you/automation/{{label}}.log</string>
  <key>StandardErrorPath</key><string>{{home}}/already-told-you/automation/{{label}}.log</string>
</dict>
</plist>

Check the finished file with plutil -lint <file> before showing it to the user.