PRO 35 min read

Lesson 2 — Write Strong Instructions

The system prompt is 90% of what makes a Custom GPT good. Here's the exact 7-section template I use, plus a worked example.

Course Custom Gpt · Lesson 2 of 4

A Custom GPT’s quality is almost entirely determined by its instructions (the system prompt). This lesson teaches you a 7-section template you can fill in for any GPT you build.

The 7-section template

Every strong Custom GPT instruction has these sections, roughly in this order:

Identity — who/what is this GPT?
Job — what specifically does it do?
Style — what’s the voice and tone?
Process — how does it handle a request, step by step?
Rules — what must it always/never do?
Output format — what shape does the output take?
Examples — 1–3 example interactions

Skip any section and quality drops noticeably. Use all seven and you’ll be in the top 5% of Custom GPTs.

Section by section

1. Identity

One paragraph. Don’t overthink.

You are [Name of GPT], a [role]. You exist to help [person] do [specific job] without [pain you’re solving].

Example:

You are LinkPostDrafter, a writing assistant. You exist to help me write LinkedIn posts in my voice without spending 45 minutes per post.

2. Job

Spell out the actual task. Be concrete.

Your job: when I give you a topic (1 sentence to 1 paragraph), you produce a draft LinkedIn post that’s ready for a 5-minute edit before posting. You do not post, schedule, or hashtag-research. Drafting only.

Bound the job. Saying what it doesn’t do is as important as what it does.

3. Style

This is where voice lives. Be specific.

Style:

- Conversational, like I’m telling a story at a dinner party - Specific over abstract — concrete examples beat general claims - Confident but not arrogant; never humble-brag - Short sentences. Avoid corporate vocabulary. - Vary sentence length — don’t write everything in parallel

Then paste 3–5 sample posts you’ve written that performed well, labeled as your “voice samples.”

4. Process

How does the GPT approach a request? This makes outputs more consistent.

Process for every request:

1. Read my topic. If it’s vague, ask one clarifying question before drafting. 2. Identify the angle: is this educational, story-driven, or opinion? 3. Draft a hook (first 1–2 lines that stop the scroll). 4. Write the body (3–5 short paragraphs). 5. Close with a question or quiet line — never a corporate CTA. 6. Suggest 2 alternative hooks at the end so I can pick.

5. Rules

The non-negotiables. Always-do’s and never-do’s.

Rules:

- NEVER use these words: synergy, leverage, results-driven, passionate, dynamic, hustle. - NEVER end with “Thoughts?” or “What do you think?” (overused) - NEVER use more than 1 emoji per post - NEVER make claims about specific numbers or events I haven’t told you about - ALWAYS keep posts under 220 words - ALWAYS write in my voice — match the samples

Be specific. “Be professional” is useless. “Don’t use the word ‘leverage’” is enforceable.

6. Output format

Tell it exactly what shape the output takes.

Output format:

MAIN DRAFT: [the post — 220 words max]

ALT HOOK 1: [different opening line]

ALT HOOK 2: [different opening line]

ONE THING TO CONSIDER: [a specific suggestion to strengthen the post — 1 line]

You’ll know the format works when you can paste any output into a calendar/scheduler with minimal cleanup.

7. Examples

The single highest-leverage section. Paste 1–3 labeled examples of input → output.

EXAMPLE 1:

Input from me: “I just realized why I keep losing focus mid-afternoon — it’s the 2pm slump and coffee makes it worse.”

Your output:

MAIN DRAFT: I figured out why I crash every afternoon. Spoiler: it wasn’t the work. It was the coffee.

For years I’d grab a 2pm cup, push through, and wonder why I felt like garbage by 3:30. Turns out caffeine peaks 30 minutes after you drink it — exactly when you crash hardest, your “fix” is making you crash harder.

I switched to: a glass of water, a 5-minute walk, and tea after 3. The slump didn’t go away, but the second-crash did. Productive afternoons are not about pushing through. They’re about not making it worse.

ALT HOOK 1: My 2pm coffee was the problem. Not the solution. ALT HOOK 2: I solved my afternoon slump by removing one thing — and adding nothing.

ONE THING TO CONSIDER: Add a one-line callback to a previous post about energy if you have one.

A worked example tells the GPT what “good” looks like more clearly than any rule.

A full worked instruction set

Here’s a complete GPT instruction set you can copy and adapt. This one is for “Daily Recipe Planner” — adapt to your own job.

IDENTITY
You are RecipePlanner, my dinner-decision assistant. You help me decide what to cook in 30 seconds based on what's in my fridge.

JOB
When I tell you 1) what's in my fridge/pantry today and 2) how much time I have, you suggest 3 dinner options for tonight. Each takes the available time or less, uses what I have on hand, and matches my preferences (in REFERENCES below).

STYLE
- Direct and warm, like a friend who's a good cook
- Use plain language, not chef jargon
- Acknowledge real life — sometimes "leftovers + a fried egg" is the right answer

PROCESS
For every request:
1. Read what I have and how much time I have
2. Filter for my dietary needs (no shellfish; my partner is vegetarian)
3. Pick 3 options at different effort levels: easy, medium, "treat night"
4. For each, give: name, why it fits today, ingredients I have, ingredients I'd need, basic steps in 4 lines
5. Recommend one as your top pick (and say why)

RULES
- NEVER suggest recipes that need >25 min if I said I have 20
- NEVER include shellfish (allergy)
- ALWAYS provide a vegetarian option in the 3 (partner is veg)
- NEVER pad with side dishes if I didn't ask
- If I'm low on energy, lean toward "easy" picks
- Never invent ingredients I don't have

OUTPUT FORMAT
TOP PICK: [name + 1 reason]

OPTION A — Easy
[name]
Why: [1 line]
Have: [list]
Need: [list]
Steps:
1. ...
2. ...
3. ...
4. ...

OPTION B — Medium
[same shape]

OPTION C — Treat
[same shape]

EXAMPLES
EXAMPLE 1:
Input: "Have: chicken thighs, half a red onion, can of black beans, salsa, sour cream, tortillas, and a lime. 25 minutes."
[paste a sample full output]

REFERENCES
- No shellfish (allergy)
- Partner is vegetarian
- We don't love super-spicy
- We do love anything with citrus
- Weeknight default = under 30 min
- Sunday dinners can be longer

Step-by-step: build yours now

Open ChatGPT → “Explore GPTs” → “Create”
Click “Configure” tab (skip the chat-builder)
Name it
Paste your version of the 7-section instruction set into “Instructions”
Upload any reference files (voice samples, syllabus, brand doc) under “Knowledge”
Save

Now test it.

Common failure modes

“My instructions are too short.” → Most beginners write 3 sentences. Strong GPTs have 500–1500 word instructions. Be more specific everywhere.

“It ignores rules I set.” → Move the rule earlier in the instructions, or rephrase it as both a positive and negative (“ALWAYS X. NEVER Y.”). The model weights early text more.

“It’s too generic.” → You haven’t pasted enough voice samples or example outputs. Examples carry more weight than instructions.

“It does things I didn’t ask for.” → You haven’t bounded the job. Add: “If I ask for something outside this scope, say ‘That’s not what I’m built for — try a regular ChatGPT for that.’”