Quick Takeaways
- Chain two small, focused prompts rather than one giant 2,000-word prompt that tries to extract, analyze, and translate simultaneously.
- Keep a changelog of your prompt templates in git or a shared doc so your team uses the latest version.
Task Objective & Prerequisites
To automate repetitive tasks reliably: (1) pick a recurring task with predictable inputs and low failure risk, (2) document the manual checklist first, (3) lock a strict JSON or Markdown output template with 2–3 examples, (4) test on 10 real historical inputs, and (5) place a human review checkpoint before any external email, database write, or publication.
What You Need Before Starting
- •5 to 10 real historical examples of the task input and desired output.
- •An AI workspace with saved projects/templates (such as ChatGPT Custom GPTs, Claude Projects, or Gemini).
- •A clear verification checklist for reviewing generated outputs.
Step-by-Step Implementation Guide
Audit Your Weekly Tasks and Score by Frequency vs. Risk
List tasks you repeat at least 3 times a week. Select tasks with high time cost but low external risk (for example, internal note formatting, changelog drafting, or data cleanup) rather than unsupervised customer billing.
Good candidate: Converting raw git commit logs into structured release notes. Poor candidate: Auto-sending unreviewed legal replies.
Write Down the Explicit Decision Rules
Before writing any prompt, write down the exact rules a human follows to complete the task—including how to handle missing fields or ambiguous inputs.
Rule 1: Every bug ticket must be tagged as UI, API, or Data. Rule 2: If no reproduction steps are present, set status to "Needs Info".
Create a Fixed Output Schema with Few-Shot Examples
Save your instructions, output template (Markdown table or JSON), and 2 real input-to-output examples inside a Claude Project or Custom GPT.
Return strictly a Markdown table with columns: [Ticket ID | Category | Severity | Missing Info | Draft Reply]. Follow the two examples below.
Benchmark Against 10 Historical Inputs
Run 10 past inputs through your saved prompt template. Compare the AI outputs against how you handled those tasks manually and refine any rule where the model guessed.
Refinement: Add "Never infer a deadline if the user did not state a date; write Unknown."
Add a 30-Second Human-in-the-Loop Review Gate
Use the AI workflow to prepare 90% of the draft or table, then perform a quick human spot-check on numbers, names, and links before sending or publishing.
Review gate checklist: (1) Names & dates match source? (2) Any unknown fields flagged? (3) Tone appropriate?
Common Mistakes & How to Fix Them
Mistake: Output format changes slightly on every run
Fix: Include a literal template block and 2 complete input/output examples in your system instructions.
Mistake: Model invents missing fields to complete a table
Fix: Explicitly instruct: "If any field is missing from the source text, output N/A—never guess or extrapolate."
Frequently Asked Questions
When should I write a Python script instead of using an AI prompt?
If the input is already structured (such as renaming files, calculating spreadsheet totals, or converting JSON to CSV), a deterministic script is faster, cheaper, and 100% repeatable. Use AI prompts when the input is unstructured text.
Can I use AI to help write the Python automation script?
Yes—using Claude, ChatGPT, or Cursor to write a deterministic local script is often the most reliable form of automation.