2.11 — Through-line workshop: "The perfect prompt" 🔥

Goal: turn the task from your track (the Level 1 draft) into a robust and reusable prompt, tested. You leave with a concrete deliverable.

🟢 In plain terms — a good prompt = Role + Context + Task + Format, plus a guardrail (“if info is missing, say so”) and, if you care about a precise tone, 2-3 examples (few-shot). Then you test and you iterate.

🎯 The block below is the recipe for your track (chosen on the home page).

🧠 Your documents copilot

• Role: “assistant that answers from documents”.
• Key instruction: “answer only from the documents below; if the info isn’t there, say so.”
• Few-shot: 2 pairs “question → answer citing the document”.
• Test on 3 questions (including one whose answer is not in the docs). 👉 Deliverable: your tested grounding prompt (plugged into the agent at Level 3).

📥 Email sorter

• Role: “assistant that sorts emails”.
• Key instruction: “classify this message as urgent / normal / spam and explain in one line.”
• Few-shot: 3 already-classified emails (one per category).
• Test on 5 varied emails, including the tricky cases from Level 1. 👉 Deliverable: your tested classification prompt (the heart of the agent at Level 3).

🎯 Lead machine

• Role: “sales analyst”.
• Key instruction: “here’s my offer and this prospect: are they a fit, and why?”
• Few-shot: 2 examples (a good fit, a bad one) with the reasoning.
• Test on 3 prospects, including a weak signal. 👉 Deliverable: your tested qualification prompt (+ a variant for personalizing outreach).

📄 Document extractor

• Role: “data extractor”.
• Key instruction: “return only an object with amount, date, supplier — JSON format.”
• Few-shot: 1 example “invoice → expected JSON”.
• Test on 3 documents in different formats. 👉 Deliverable: your tested extraction prompt (structured output, ready to write at Level 3).

✍️ Content factory

• Role: “writer who respects a brand voice”.
• Key instruction: “adapt this idea for [channel], tone = [yours], length = […].”
• Few-shot: 2 of your best posts (to lock in the style).
• Test: 1 pillar piece → 2 formats (e.g. LinkedIn post + newsletter). 👉 Deliverable: your tested per-channel prompts (to orchestrate at Level 3).

🎯 Your own project

1. Take the Level 1 draft (the task rephrased in natural language).
2. Dress it with the 4 ingredients: Role + Context + Task + Format.
3. Add an anti-hallucination guardrail (“if info is missing, say so”).
4. Few-shot if needed: 2-3 examples for a precise tone/format.
5. Test on 5 varied cases, including the tricky cases from Level 1. 👉 Deliverable: your robust, tested prompt (the heart of your agent at Level 3).

🔁 Then, for all tracks: iterate (a word, an example, an instruction) until you get a reliable result, and test your final prompt on 2 assistants (ChatGPT, Claude, Gemini). Note the differences in tone/format and which one fits best — that’s your home benchmark.

For techies (optional): save this prompt as a parameterizable template (a variable for the input). It will become the heart of your agent at Level 3.