Automate my work: find the repeat tasks and hand them to Claude

Use this first, with last week's calendar or task list in front of you.

Help me find the work I could automate. Here is what my last week looked like: <week> [PASTE YOUR CALENDAR, TASK LIST OR A ROUGH DIARY] </week> My job: [WHAT YOU DO]. Tools I use every day: [FOR EXAMPLE GMAIL, GOOGLE SHEETS, NOTION, SHOPIFY]. Process: 1. Ask me one question at a time, up to eight, about the tasks that look repetitive: how often, how long, what the input is, what the output is, and whether the steps are the same each time. 2. Then list every repeat task in a table with columns: Task, How often, Minutes each time, Hours a week, Input, Output, Same steps each time (yes, mostly, no). Include small tasks too. Five minutes ten times a day adds up to over four hours in a five-day week.

Swap the highlighted parts for your own words.

Tip

• If you have no calendar to paste, keep a quick note of what you do for one normal week first. Claude can only find tasks that are on your list.

Use this straight after the audit.

Using the task table above, rank the tasks for automation. Score each task from 1 to 5 on: - Time saved per week - How clear the rules are (5 means the same steps every time) - Low harm if it goes wrong (5 means a mistake is easy to spot and fix) - Ease of building with Claude Code and the tools I use Output a table sorted by total score, with a column for the estimated hours to build it and the weeks until it pays back that time. Then: 1. Pick the top three and say why. 2. Name any task that should stay manual because it needs judgement, a personal touch, or carries real risk, such as sending money or messaging clients. 3. Name any task that would disappear if I changed the process, with no automation needed. Be honest when a task is not worth automating.

Use this on the top task, before you build anything.

Design an automation for this task: [THE TASK FROM THE RANKING]. How I do it by hand today, step by step: [YOUR STEPS]. Write a design with these parts: 1. Trigger: what starts it (I run a command, a time of day, a new file appears, an email arrives). 2. Inputs: where the data comes from, and what access that needs. 3. Steps: numbered, with what Claude does and what a plain script does at each step. 4. Output: what it produces and where it puts it. 5. Human checkpoint: the point where I review before anything leaves my computer or changes something shared. 6. Failure: what happens when an input is missing or a step fails, and how I find out. 7. Where it runs: a skill I run myself, `/loop` while a session is open, a Desktop scheduled task for local files, or a cloud routine for when my computer is off. Recommend one and say why. 8. Cost: rough usage per run and anything paid. Keep the first version as small as possible. List extras for later separately.

Swap the highlighted parts for your own words.

Tip

• Save the design as DESIGN.md in your automations folder. The build prompts read it.

Use this when the automation is something you will run yourself by typing a command.

Read DESIGN.md. Build this automation as a personal Claude Code skill. 1. Create the file `~/.claude/skills/[SKILL-NAME]/SKILL.md`. The folder name becomes the command, so I will run it with /[SKILL-NAME]. 2. Start the file with YAML frontmatter between --- lines, containing: - name: [SKILL-NAME] - description: one sentence on what it does and when to use it - disable-model-invocation: true, so it only runs when I type the command 3. Under the frontmatter, write the steps from DESIGN.md as clear numbered instructions to Claude, including the exact input locations, the output format and the human checkpoint. 4. If a step is plain data handling, put it in a small script inside the same folder and have the skill run it. 5. Tell me how to test it, then stop. Do not run it on real data yet. Keep it as short as it can be while still being exact.

Swap the highlighted parts for your own words.

Tips

• Skills in ~/.claude/skills/ work in every project on your computer. For one project only, put the skill in that project's .claude/skills/ folder.
• Type / to see the list of commands. If /[SKILL-NAME] is missing and ~/.claude/skills/ did not exist when you started Claude Code, type /reload-skills.

Use this when the automation should run from a script, a shortcut or a scheduler, without you typing in Claude Code.

Read DESIGN.md. Build this as a shell script, a small file of terminal commands, that calls Claude with `claude -p`, which runs Claude Code once without opening a chat. Requirements: 1. Name it `[NAME].sh` and put it in this folder. 2. Gather the inputs with plain commands first, then pass them to `claude -p` with a clear prompt. 3. Allow only the tools it needs with `--allowedTools`, such as "Read" or "Read,Edit". 4. Add a `--dry-run` option that prints what it would do and changes nothing. 5. Write a log of each run, with the date, to `logs/[NAME].log`. 6. Stop with a clear error message if an input is missing. 7. Save outputs as drafts for me to review. Nothing is sent or published by the script. Then run it with `--dry-run` and show me the output.

Swap the highlighted parts for your own words.

Tips

• Run chmod +x [NAME].sh once so your computer lets you run it, then start it with ./[NAME].sh.
• Each run uses your Claude plan's usage, the same as a chat. Check how much one run takes before you schedule it often.

Use this after building, before the automation touches real work.

Test the automation in this folder before I use it for real. 1. Make a test folder with sample inputs: one normal case, one with missing data, one with a very large input, and one with odd formatting. 2. Run the automation on each, in dry run or draft mode. 3. For each run, show me the input, the output and whether it was correct. 4. Check that nothing was sent, deleted or changed outside this folder. 5. List every failure and fix them one at a time, running the tests again after each fix. Finish with a table: Test case, Result (pass or fail), Notes. Tell me plainly if you think it is not ready.

Use this for any automation that sends, posts, pays, deletes or changes something other people see.

This automation does something other people will see or that is hard to undo: [WHAT IT DOES, FOR EXAMPLE SENDS EMAILS TO CUSTOMERS]. Change it so it never takes that action on its own: 1. It prepares everything and saves it to a review folder, one clear file per item. 2. It writes a short summary at the top: how many items, anything unusual, anything it was unsure about. 3. It takes the action only when I run a separate approve step, and only for the items I approve. 4. It records what was approved and sent, with the date, in the log. Show me the new flow and run it once in dry run mode.

Swap the highlighted parts for your own words.

Tip

• Keep the checkpoint until the automation has run correctly for at least a few weeks. Remove it only for low risk steps.

Use this when the automation works and you want it to run on a timetable.

I want [THE AUTOMATION] to run [HOW OFTEN, FOR EXAMPLE EVERY WEEKDAY AT 8AM]. It needs: [WHAT IT NEEDS, FOR EXAMPLE FILES ON MY COMPUTER, A GITHUB REPOSITORY, GMAIL]. Help me choose between: 1. `/loop`, which repeats a prompt while this Claude Code session stays open, and stops after seven days 2. A Desktop scheduled task, which runs on my computer and can reach my local files, but only while the Claude Desktop app is open and the computer is awake 3. A routine, set up with `/schedule`, which runs in the cloud even when my computer is off, but works from a fresh copy of a GitHub repository and cannot see my local files For each, say whether it can reach what my automation needs, and what would stop it working. Note two limits: cloud routines cannot read skills that exist only in ~/.claude/skills/ on my computer, and a scheduled `/loop` run cannot call a skill marked `disable-model-invocation: true`. Recommend one, then walk me through setting it up step by step. Ask me before you create anything.

Swap the highlighted parts for your own words.

Tips

• Routines need a Pro, Max, Team or Enterprise plan and run at most once an hour. Each run counts against your plan's usage limits.
• Pick a start time a few minutes past the hour, such as 8:07. Runs set exactly on the hour can start several minutes late.

Use this when a run gave the wrong result or stopped.

My automation [NAME] failed. What I expected: [EXPECTED RESULT]. What happened: [WHAT HAPPENED]. 1. Read the log in `logs/` for the failed run and the automation's files. 2. Find the step where it went wrong, and show me the log lines that prove it. 3. Explain the cause in two or three sentences. 4. Fix it, then add a check so this kind of failure stops the run with a clear message next time. 5. Run it again in dry run mode with the input that failed, and show me it now works. If the cause is outside the automation, such as an expired login or a changed website, tell me what I need to do.

Swap the highlighted parts for your own words.

Use this once a week, or set it as a skill, to keep your automations healthy.

Review every automation I have running. 1. List my personal skills in `~/.claude/skills/` and the scripts and logs in [YOUR AUTOMATIONS FOLDER]. Skip `~/.claude/skills/synced/`, which holds copies of skills from my claude.ai account. 2. For each one, read its log and find: the last run, how many runs this week, any failures, and any warnings. 3. Ask me for anything you cannot see, such as routines on claude.ai, and wait for my answer. 4. Build a table: Automation, Runs this week, Failures, Time saved (my estimate is [MINUTES PER RUN] per run), Verdict (keep, fix, retire). 5. For anything marked fix, say what is wrong. For anything marked retire, say why. Keep the review short. List only what needs my attention under the table.

Swap the highlighted parts for your own words.

Tips

• Turn this prompt into a skill with the skill prompt above, then run it every Friday.
• Retire automations you no longer use, so there is less to check and less that can break.