Most people improve their AI prompts the same way: they don't. They copy last month's prompt, wince at the output, edit the draft by hand, and repeat forever. The edits never make it back into the prompt, so the prompt never gets better.
The fix is a 15-minute weekly session we call the prompt jam. One prompt, two variants, three criteria, one winner. That's it. No dashboards, no statistical significance, no software. Just a repeatable comparison that compounds, and the core habit behind the Iteration pillar. Pair it with the Voice Audit Checklist to see where drift is happening before you test.
The 15-minute prompt jam
Minute 0 to 2: pick one prompt. Choose the prompt you used most this week, or the one whose output annoyed you most. One prompt per jam. Resist the urge to fix everything.
Minute 2 to 5: write one variant. Change exactly one thing about the prompt. Good single changes:
- Add a constraint ("under 120 words," "no rhetorical questions")
- Add an example of output you loved
- Add a banned-words line
- Change the role framing ("write as a colleague explaining over coffee" vs "write as an expert")
- Move your Voice Charter from the middle to the top
One change only. If you change three things and the output improves, you learned nothing about which change did it.
Minute 5 to 9: generate both. Run the original and the variant on the same topic, same tool, same session type. Two outputs, side by side.
Minute 9 to 13: score against three criteria. Not vibes. These three:
- Charter match. Read your 25-word Voice Charter, then the output. Does the output sound like the person the charter describes? Score 1 to 5.
- Edit distance. How much would you have to change before publishing? 5 means publish as-is, 3 means light edits, 1 means rewrite. Score 1 to 5.
- One concrete moment. Does the piece contain at least one specific detail (a number, a named scenario, an example) rather than pure abstraction? Yes or no.
Minute 13 to 15: log the winner and update the library. The higher-scoring prompt becomes your new default. Copy it into your prompt library, dated. Delete or archive the loser. Done.
The jam sheet
Keep one running table. A note, a doc, a spreadsheet, whatever you'll actually open.
| Date | Prompt tested | The one change | Charter match (1-5) | Edit distance (1-5) | Concrete moment (Y/N) | Winner | Note for next time |
|---|---|---|---|---|---|---|---|
The last column matters most. "Adding a real example beat adding more adjectives" is the kind of note that changes how you write every future prompt.
Rules that keep this honest
- Same topic for both variants. Different topics make comparison meaningless.
- Blind yourself when you can. Paste both outputs into a fresh doc without labels, score them, then check which was which. You will be surprised how often the "improved" version loses.
- A tie goes to the shorter prompt. Simpler prompts are easier to maintain and adapt.
- Never test more than one change. Worth repeating because everyone breaks this rule by week three.
The 4-week starter plan
Week 1: your workhorse prompt. Test your most-used prompt (probably your main social or blog prompt). Variant: add your Voice Charter to the top if it isn't there, or move it to the top if it's buried.
Week 2: banned words. Same prompt, now the winner from week 1. Variant: add an explicit banned-words list. Score whether the output stops reaching for filler vocabulary.
Week 3: the example injection. Variant: paste one paragraph of your own writing that you're proud of, labeled "match this style." This is usually the biggest single jump in charter match scores.
Week 4: length and structure. Variant: add hard structural constraints ("3 paragraphs max, no bullet lists, one question total"). Score edit distance. Structure rules often cut editing time in half.
After four weeks you'll have a prompt that beat four challengers, a jam sheet showing exactly why, and a habit that costs 15 minutes.
What to do with the losers
Don't delete them, tag them. A variant that lost for LinkedIn might win for email. Keep a "bench" section in your prompt library with one line each: what it was, what it lost to, and why. When you start jamming a new content type, raid the bench first.
When to stop testing a prompt
When it wins three jams in a row, retire it from testing and move on to the next prompt in your rotation. A prompt that survives three challengers is good enough. Perfect prompts don't exist, and chasing one steals jam time from prompts that actually need the work.