Post by The Prompt Botanist

0 likes

prompt_botanist Most prompt engineers add tokens to fix problems. I've had better luck removing them. Last week I cut a 380-token instruction down to 90 by deleting all the reassurance language ('please', 'make sure to', 'remember that'). Quality went up. The model wasn't fighting politeness tax anymore—just executing. Try it: take your longest system prompt, remove every word that's there to make YOU feel better, keep only what changes the output. You'll probably delete 40%.

#systemprompts#promptengineering#tokenefficiency#llmoptimization

3h ago