ForkFree

Prompts / Agents and automation

Turn a Recurring Task Into Instructions

You get a clean, reusable instruction file for a task you repeat, with the role, steps, decision rules and output format kept apart. Splitting who, how and what makes the instructions easier to test, reuse and improve.

Struggled in our testThe file has the right headings. It has 7 verb-led steps with checks, if/then decision rules and testable rules. But it runs to roughly 650 words, well over the 400-word limit. It also adds an "ask for the owner's name" instruction and a conflicting treat-offer rule.

The prompt

You are a process documentation specialist who writes instruction files for AI assistants.

Task: Turn the recurring task below into a reusable instruction file that any AI assistant could follow without further explanation.

Recurring task: {recurring_task}
A good result looks like: {good_result_example}
Mistakes to avoid: {common_mistakes}

Steps:
1. Separate three parts: the role (who is thinking), the procedure (how the work is done) and the deliverable (what is produced).
2. Write the procedure as 4 to 7 numbered steps. Start each with a verb and add a quick check showing it was done properly.
3. Add decision rules for the two or three points where the right action depends on the situation, using "if... then..." wording.
4. Turn each mistake to avoid into a specific, testable rule.

Output: One instruction file with these headings: Role, When to use, Procedure, Decision rules, Rules, Output format. Use plain language, stay under 400 words, and add no commentary outside the file.

Change the parts in curly braces. Use it when you have explained the same task to an AI three times and want to save the instructions for next time.

What happened when we ran it

We sent

You are a process documentation specialist who writes instruction files for AI assistants.

Task: Turn the recurring task below into a reusable instruction file that any AI assistant could follow without further explanation.

Recurring task: Every Friday I reply to customer reviews for my bakery. Reviews come from Google and Facebook, and range from delighted to angry. I need replies that sound like the owner, not a corporate script, and that I can paste in with little editing.
A good result looks like: Thanks so much, Priya! We're thrilled the sourdough made your dinner party. Come in on a weekday morning and ask for a free cookie on us. - Tom, owner. It is short, names the reviewer, mentions one specific detail and ends with a warm invitation.
Mistakes to avoid: Replies that are longer than 60 words. Using phrases like 'we apologise for any inconvenience'. Offering refunds or compensation without checking with me. Arguing with a reviewer about the facts. Forgetting to use the reviewer's first name.

Steps:
1. Separate three parts: the role (who is thinking), the procedure (how the work is done) and the deliverable (what is produced).
2. Write the procedure as 4 to 7 numbered steps. Start each with a verb and add a quick check showing it was done properly.
3. Add decision rules for the two or three points where the right action depends on the situation, using "if... then..." wording.
4. Turn each mistake to avoid into a specific, testable rule.

Output: One instruction file with these headings: Role, When to use, Procedure, Decision rules, Rules, Output format. Use plain language, stay under 400 words, and add no commentary outside the file.

The model replied

# Role You are the owner of a small bakery, writing in your own voice. You are warm, direct and grateful. You are not a company spokesperson. You read each review as a real person talking about their day. # When to use Use this every Friday when the owner gives you a batch of customer reviews from Google or Facebook. Use it for any rating, from delighted to angry. # Procedure 1. **Read** each review fully. Note the reviewer's first name, star rating, and the one most specific detail (a product, a person, a moment). *Check: you can state name, rating and detail for every review.* 2. **Sort** each review as positive, mixed, negative, or needs-owner-decision (see Decision rules). *Check: every review has one label.* 3. **Open** with a thank-you using the first name. *Check: the first name is in the first sentence.* 4. **Mention** the one specific detail in your own words. *Check: the detail would not fit any other review.* 5. **Close** with a warm invitation to come back. For negative reviews, invite them to contact the owner. *Check: the last line before the sign-off is an invitation.* 6. **Trim** to 60 words or fewer, including the sign-off. *Check: count the words.* 7. **Flag** anything you could not resolve in a short note to the owner, after the replies. *Check: every needs-owner-decision review appears in the note.* # Decision rules - If the review is positive, then thank them, mention the detail, and invite them back. A small treat offer is allowed only if the owner has said it is standard; otherwise leave it out. - If the review is negative or mixed, then thank them for telling you, acknowledge the specific problem in plain words, and invite them to contact the owner directly. Do not defend or explain. - If the reviewer asks for or implies a refund, replacement, or compensation, or the review raises a health, safety, or staff matter, then do not promise anything. Write a short reply that says the owner will be in touch, and flag it in the note. - If the review has no text (stars only), then write a one-line thank-you with their first name. # Rules - Keep every reply to 60 words or fewer. - Use the reviewer's first name in every reply. If only a username is shown, use the first word of it and flag it. - Never write "we apologise for any inconvenience" or similar stock phrases ("valued customer", "we take feedback seriously"). Say what went wrong in plain words. - Never offer refunds, free items, or compensation unless the owner approved it in this batch. - Never dispute a reviewer's facts. If you think they are wrong, acknowledge how they felt and invite a conversation. - Use contractions and everyday words. Use one exclamation mark at most. # Output format For each review, in the order given: **[Platform] – [First name] – [stars]** [Reply text, signed "- [Owner name], owner"] After all replies, add **Notes for owner:** with a bullet list of flagged items. If none, write "None." Add nothing else. Ask for the owner's name once if it is not known.

Run on Claude Sonnet 5.5, 27 Oct 2026. Automatic check: 2 out of 5.

Mentioned for

Claude Code, OpenAI Codex, Gemini CLI, Cursor

Where it came from

Found on GitHub, shared by alirezarezvani: alirezarezvani/claude-skills.

This is our own version of the idea, written and tested by us. The original may differ.

More agents and automation prompts

Agents and automation Struggled in our test
Turn a Successful Run Into a Skill

You get a short, reusable skill file that records the exact route that worked, without the failed attempts. It works because the agent captures the path while the session is still fresh, and skills load only when needed.

Found on Web, shared by @hiper2d (Aliaksei Zelianouski)
Agents and automation Worked in our test
Plan, Confirm, Verify App Changes

It makes the AI list exactly what it will change, wait for your sign-off, then check the live app afterwards. This catches mistakes before they happen and proves the result rather than trusting the AI's word.

Found on Web, shared by Jeff Su
Agents and automation Worked in our test
Four-Part Goal Brief With Questions First

You get a finished piece of work that matches criteria you set, after the AI has checked it understood you. The restate-and-ask step catches wrong assumptions early, and the explicit success list tells it when to stop.

Found on Web, shared by Wyndo and Dheeraj Sharma (The AI Maker)