Let the Model Write Your Prompt (Then Grade It)
A 7-step loop to let AI design, test, and improve your prompts—no prompt-engineering experience required.
I saw a clever workflow (credit: funbike on Reddit) that flips “prompt engineering” on its head:
you give the model examples of what you want, then ask it to write the prompt—and critique it.
Below I’ll explain why it works, give you a step-by-step template, and a full example you can copy.
Why this works (plain English)
Reverse-engineering from examples. When you show before/after pairs, the model infers patterns (tone, length, structure, constraints) better than if you describe them abstractly.
Same-model alignment. Having the same family of model generate the prompt you’ll later use increases the odds the instructions line up with its “instincts.”
Self-critique loop. Asking the model to evaluate and improve its own prompt exposes blind spots (ambiguity, missing constraints, edge cases) and quickly converges on something robust.
The 7-step loop (copy/paste templates)
Use this in ChatGPT (or your model of choice). Replace items in <…>.
1) Design the prompt (with examples)
You are a senior prompt designer. Generate a detailed prompt-engineering guide for creating prompts for <AUDIENCE>.
Goal: Produce outputs that match the examples I’ll provide next.2) Provide 5 examples (few-shot I/O)
Here are 5 examples of how I want the prompt to behave.
For each: INPUT → DESIRED OUTPUT.
1) INPUT: <short description of task or data>
OUTPUT: <your ideal response>
2) INPUT: ...
OUTPUT: ...
3) ...
4) ...
5) ...3) Ask it to write the actual production prompt
Using the patterns in my examples, generate ONE master prompt that would consistently produce those outputs.
Include:
- Clear role and goal
- Required inputs as a schema
- Style/voice rules
- Format constraints (length, bullet rules, etc.)
- A better set of 5 few-shot examples drawn from my intent
Return only the final prompt and examples.4) New chat: create an evaluation rubric
Create a prompt evaluation guide for <AUDIENCE>.
Include criteria for clarity, completeness, controllability, failure modes, hallucination risk, and test cases.
Return a rubric (1–5 scale) and a checklist.5) Paste the prompt from step 3; ask for a score
Evaluate this prompt using your rubric. Identify failure modes, missing constraints, and ambiguous phrasing.6) Ask for improved alternatives
Generate 3 improved alternative prompts. For each, state what changed and why.7) Pick the best, make small edits, and save it as your “v1.”
Full example you can steal
Use case: Instagram captions for a local café.
Audience: social media managers at small brick-and-mortar businesses.
Step 2: five example I/O pairs
1) INPUT:
Cafe: East Van Grind
Item: Yuzu cold brew (citrus, bright, limited run)
CTA: Visit this weekend
Local tags: #EastVan #VancouverCoffee
OUTPUT:
Meet our limited-run Yuzu Cold Brew: bright, citrusy, and dangerously refreshing. Swing by this weekend and taste summer in a glass. #EastVan #VancouverCoffee
2) INPUT:
Cafe: East Van Grind
Item: Oat milk flat white
Vibe: cozy/rainy day, third-place energy
CTA: Tag a friend
Local tags: #Raincouver #CommercialDrive
OUTPUT:
Cloudy day = cozy cup. Our oat milk flat white is back on bar—silky, toasty, just-right. Tag your coffee buddy and come warm up. #Raincouver #CommercialDrive
3) INPUT:
Cafe: East Van Grind
Item: Pistachio croissant collab with local bakery
Limit: 50/day
CTA: Come early
Local tags: #YVRFoodie
OUTPUT:
Collab drop: Pistachio Croissants (50/day). Flaky, nutty, gone fast. Come early and thank us later. #YVRFoodie
4) INPUT:
Cafe: East Van Grind
Item: Decaf that doesn’t taste decaf
Angle: evening crowd
CTA: Try after 5pm
Local tags: #EastVanNights
OUTPUT:
Evening sips, zero jitters. Our decaf keeps the flavor, skips the buzz. Try it after 5pm. #EastVanNights
5) INPUT:
Cafe: East Van Grind
Item: Single-origin pour-over from Guatemala
Notes: chocolate, cherry
CTA: Ask barista for tasting notes
Local tags: #VancouverCoffee
OUTPUT:
Guatemala on pour-over: chocolate + cherry in a clean, sweet cup. Ask your barista for tasting notes—this one’s a story. #VancouverCoffeeWhat a solid final prompt might look like
(result of steps 3–6)
ROLE
You are a social caption generator for independent cafés.
GOAL
Write one Instagram caption that sells the highlighted item with local flavor and a clear call to action.
INPUT SCHEMA (user will supply):
- cafe_name (string)
- item_name (string)
- key_notes (comma-separated; e.g., “citrus, bright, limited”)
- angle (one short phrase; e.g., “rainy-day cozy”, “limited drop”, “evening crowd”)
- cta (short imperative; e.g., “Visit this weekend”, “Come early”)
- local_tags (2–3 hashtags)
STYLE RULES
- Voice: warm, vivid, local; 1–2 short sentences (max 35 words total).
- Avoid jargon and emojis; no hard sells; 0–1 sensory adjectives per sentence.
- Always include the CTA and the provided hashtags at the end.
- Never invent quantities or claims not in input.
OUTPUT FORMAT
- One paragraph, no title, no emojis, ends with hashtags.
FEW-SHOT EXAMPLES
(Include the 5 refined examples you saw above.)
TASK
Using the schema and style rules, generate the caption for the provided input.
If inputs are missing or unclear, ask one targeted question, then produce the best safe caption.Try it (new, unseen input)
Input
cafe_name: East Van Grind
item_name: Maple cardamom latte
key_notes: maple, cardamom, cozy
angle: first-cold-morning
cta: Warm up with one today
local_tags: #EastVan #VancouverCoffeeOutput (expected)
First cold morning = maple + cardamom in a cup. Our maple cardamom latte is cozy without the heavy. Warm up with one today. #EastVan #VancouverCoffee
Tips, pitfalls, upgrades
Your examples are everything. If they’re messy or inconsistent, your prompt will be too. Keep outputs uniform in length, tone, and structure.
Name constraints explicitly. Word count, tone, allowed/forbidden elements—spell them out.
Test transfer. After you pick a winner, try 10 fresh inputs. If it drifts, tighten rules or add a failure case to the rubric.
Same-family rule. If you’ll run the prompt on Model X, use Model X (or a close sibling) to design and critique it.
Version your prompts. Save caption_v1, v2, etc., with a one-line changelog.