4.3 -- Iteration: your first version won't be the right one

Context

Your first system prompt will be bad. Not because you're bad — because nobody knows what they want from an agent until they've used it. That's normal. The prompt reveals itself through use, not through thinking about it.

Plan for 2-3 rounds of iteration before you have a prompt that feels right. After that, occasional tweaks when your needs change.

The iteration process

Round 1: the rough version

Write your first prompt (see section 4.1). Use it for 2-3 days. Note every friction point:

• "It responds too long"
• "It forgets I want informal speech"
• "It suggests Windows solutions when I'm on Linux"
• "It asks for context every time"

Round 2: the correction prompt

Use this pattern to fix it:

Here's my current system prompt:
[paste the prompt]

Problems observed:
1. [friction 1]
2. [friction 2]
3. [friction 3]

Rewrite the prompt fixing these issues.
Keep the structure. Don't exceed 250 words.

Review the suggestion. Manually edit whatever doesn't fit. Test for 2-3 more days.

Round 3: condense

After round 2, your prompt probably runs 400-500 words. That's too much. Every token counts.

Condensation prompt:

Here's my system prompt (480 words).
Condense it to 150 words maximum.
Keep ALL the rules and restrictions.
Remove explanations -- the agent doesn't need to understand why.

The rule: 150 words > 500 words. A short, precise prompt beats a long, detailed one. The model doesn't need your justifications — it needs your instructions.

Why condensing works

A long prompt creates noise. The model has to sort what matters from what's filler. The shorter it is, the more weight each word carries.

Comparison:

Before (68 words):

When you suggest technical solutions to me, I'd like you to take into
account the fact that I work primarily on Linux Ubuntu environments, and
that I generally don't need solutions specific to Windows or macOS, unless
you're explicitly asked for them in my question.

After (12 words):

Environment: Linux Ubuntu. No Windows/macOS solutions unless requested.

Same information. 5x fewer tokens.

When to stop iterating

Stop when:

• You're no longer correcting the agent on tone or format.
• Responses are useful on the first try in 80%+ of cases.
• Your corrections are about substance (knowledge) not form (behavior).

If you're still iterating after 5 rounds, the problem probably isn't the prompt. It might be the model, the tool, or your expectations.

Common mistakes

Never iterate. The first draft stays in place for months. The prompt accumulates obsolete instructions and the agent gradually degrades.

Iterate without noting. You change the prompt "from memory" without having written down the friction points. You fix one problem and create another.

Add without removing. Every iteration adds rules but never removes any. After 5 rounds, the prompt is 800 words and contradicts itself.

Steps

1. Use your current prompt for 2-3 days.
2. Note every friction point in a file prompt-feedback.md.
3. Apply the correction prompt (round 2).
4. Test for 2-3 days.
5. Condense to 150 words (round 3).
6. Test. If it holds, you're done.

Checklist

• [ ] The current prompt has been tested in real conditions (not just reread).
• [ ] Friction points are written down, not from memory.
• [ ] The final prompt is under 200 words.
• [ ] No instruction contradicts another.
• [ ] 80%+ of responses are useful without correction.