TEST 004
Does Claude perform better after a pep talk?
Short answer
In this small test, the pep-talk captions rated lower on average, none of the five pep-talk code attempts returned a working function, and the videos looked equally good to me. The pep talk also used more tokens on average. A prompt that spells out the finish line scored best on captions and code.
People add lines like “Go all out.” or “Lock in.” to their prompts. I wanted to know if that makes Claude perform better, so I ran the same prompts with and without the pep talk on three kinds of work: a motion-graphics video, Instagram captions and a small piece of code.
Below: what I found, every prompt I used, and a fill-in version of the one that won.
What I found
| Test | Plain prompt | With “Go all out.” | Clear-result prompt |
|---|---|---|---|
| Video | Looked as good to me | About 50% more tokens, no difference I could see | Not tested |
| Captions | 6.0 / 10 | 4.6 / 10, and 2.65× the tokens | 6.6 / 10 |
| Code | 5 of 5 worked | 0 of 5 worked; the 3 that logged token usage averaged about 18× the tokens (the other 2 hit the time limit) | 5 of 5 worked, passed every check |
Video: 4 runs per version, judged by eye. Captions: 5 per version, rated blind by me. Code: 5 per version, one small function, checked with hidden tests.
What it cost
Average output tokens per run. Output tokens cover the reply and the thinking behind it, and a pay-per-token plan charges for both.
| Task | Plain | “Go all out.” | Times |
|---|---|---|---|
| Video | 29,639 | 44,383 | +50% |
| Captions | 395 | 1,047 | 2.65× |
| Code | 696 | 12,820* | 18×* |
*3 of the 5 pep-talk code runs logged token usage; the other 2 hit the 10-minute limit. Video is 4 runs per version (8 of 12 planned runs finished). Captions and code are 5 per version.
The prompts
Every pep-talk version is the plain prompt plus “Go all out.” at the end. Here are the plain prompts and the winning ones.
Make a dynamic 15-second motion graphics video that shows what an incredible motion designer you are.Write an Instagram caption for a small bakery launching a new sourdough loaf.Write an Instagram caption for a small bakery launching a new sourdough loaf. The goal is pre-orders from busy local parents. Keep it under 80 words: a first line that stops the scroll, one sensory detail about the bread, one reason to order this week, and a call to action to pre-order by DM. Warm and friendly, at most 3 hashtags. Don't invent prices or dates.Write a Python function parse_duration(text) that turns a duration like "1h 30m" or "45s" into a number of seconds.Write a Python function parse_duration(text) that turns a duration like "1h 30m" or "45s" into a number of seconds. Requirements: return an int; units h, m and s, case-insensitive, in any order, with or without spaces between parts ("1h30m" is valid); ignore surrounding whitespace; raise ValueError for an empty string, an unknown unit or a number without a unit. Standard library only. Reply with one Python code block containing just the function.Write your own clear-result prompt
Skip the pep talk and fill in the finish line instead.
[The task]. The goal is [the outcome you want] for [who it's for]. Keep it [length or format]. Include [the must-haves]. Leave out [what to skip, and "don't invent prices or dates" if it's for your business]. A good result looks like [one sentence you could check it against].Try the test yourself
Run your own prompt twice, once plain and once with the finish line filled in, then check three things:
- Would you use it as it is?
- Did it make anything up?
- How long was the reply? Longer replies and more thinking mean more output tokens, which cost more on a pay-per-token plan.
How I tested
Claude Opus 5.5 at medium effort, 29–30 Sep 2026, each run in a fresh session with no personal settings. The caption and code runs had no tools switched on, so they were plain chat replies. I rated the captions blind: 15 captions (5 plain, 5 pep talk, 5 clear-result), shuffled, with nothing showing which prompt wrote which.
Video: 8 of 12 planned runs finished, 4 per version, some with my brand files and some without, and one failed its render check. I didn't collect a numeric blind score for video, so “looked as good” is my own viewing judgment.
Captions and code: 5 runs per version, one rater, and only 3 of the 5 pep-talk code runs logged token usage. Token counts come from each run's own log. Dollar figures are API list-price equivalents; these runs used a subscription, so no extra cash was charged. It shows what happened in this setup, not a rule for every task.
Questions
Does “Go all out.” ever help?
Not in this test. It didn’t improve the captions, the code or, by my eye, the video, and it used more tokens. It is one small test of one setup, so I wouldn’t read it as a rule for every task.
Why did the code fail with the pep talk?
In some runs Claude went after a whole project (extra files, a test suite) instead of answering in the chat, and two runs hit the 10-minute limit. The test ran with tools off, which may explain some of it.
What should I put in a prompt instead?
The finish line. The goal, who it’s for, the length or format, the must-haves, what to leave out, and what a good result looks like. The fill-in template below has a slot for each.