Your bookmarks

Loading your bookmarks

Latest updates

New courses, skills, book chapters, and site improvements will appear here.

Create free accountSign upDashboardDashboard

Prompt stuffing

  • Context
  • Practitioner slang

Prompt stuffing is cramming every rule, example, and document into one prompt without choosing what the current task needs or keeping the additions up to date.

It grows one patch at a time. Each time the AI gets something wrong, someone adds a paragraph, and nobody removes the old ones. The model weighs everything it's given, so stale rules still steer answers, contradictions leave it choosing between them unpredictably, and the one rule a task needs competes with dozens it doesn't.

Where you’ll meet it

The custom instructions for the assistant you use to plan family trips have grown to three pages over a year. One paragraph says “always book the cheapest flight,” another says “never book a layover,” and its suggestions flip between the two from one chat to the next.

What to do

  1. Decide which flight rule wins, and delete the other.
  2. For each remaining paragraph, ask what mistake it prevents, and delete the ones nobody can answer.
  3. Ask again for two trips you've planned before, and compare the suggestions with the old ones.

Easy to mix up

Context engineering is a skill: choosing what an AI is given for the task in front of it. Prompt stuffing is the habit that skill replaces, piling everything in without choosing.

Check yourself

Your team's assistant runs on 40 pages of instructions covering refunds, shipping, and billing, all current and none in conflict. Its answers often skip a rule that applies, a different one each time. Which change is likelier to help?

Go deeperExtra text slows the model and can lower its accuracy

Prompt stuffing is practitioner slang, used as a warning, with no formal research definition. The effects it warns about show up in measurement. The model processes every word on every request, which adds time and cost, and Chroma's 2025 context-rot report found models doing worse on the same task as unrelated text was added to the input.

Moving details out of the prompt has its own tradeoff, since a reference file only helps when the assistant knows to open it. Rules it needs on nearly every request stay in the prompt, and tables or details it needs now and then go in files or a search tool it can call.

Practice Modern AI

Or take the learning quiz on a mix of every set