Ask an LLM the same thing twice. You get two answers.

In pandas, a seed makes random results repeat. Seed is a small add-on for Claude Code that does the same for an LLM's answer. Press both buttons a few times.

Without a seed

Write a one-line slogan for an AI analytics team.

With seed 1

/seed 1 Write a one-line slogan for an AI analytics team.

These are recorded answers from Claude Haiku on 5 October 2026, played back in your browser in the order they were given. Nothing here calls a model.

Step 1 of 5

How much an answer changes depends on the question

We asked three questions 30 times each, with nothing held fixed. Each square below is one ask. Squares that touch gave exactly the same text.

30 asks of each question, grouped by identical answer

Nothing fixed: 30 separate asksOne seed, asked 30 times
Show these numbers as a table

The creative question gave 22 different slogans in 30 asks. The factual one gave the same fact in 2 wordings. And the one question that asked for randomness got none.

"Pick a random number between 1 and 100"

42, every time

All 30 asks returned the same number. Each grey cell is a number the model never picked. An LLM is not random where you expect it to be, and it is random where you do not.

The slogans vary, but less than they seem to

29 of the 30 slogans opened with the same five words. Only the ending changed.

"Turning data into decisions, …

    Step 2 of 5

    A seed fixes one answer to one question

    Type /seed, a seed and your question. The first time, the model answers as usual. From then on, that seed and that question return that answer, word for word. A different seed gives you a fresh one.

    The same slogan question, asked 30 more times with seed 1

    All 30 replies carried the same fingerprint, b29369701c51. A fingerprint is the start of a SHA-256 hash of the answer text, so one fingerprint means one text.

    /seed 1 Write a one-line slogan for an AI analytics team.

"Turning Data Into Decisions, Insights Into Impact."

— seed 1 · replayed, identical to the first draw · model haiku · fingerprint b29369701c51
    Step 3 of 5

    What happens to a question

    The mod does not change how the model picks its words. Chat tools give no control over that. It works one level up: the first answer drawn for a seed is saved, and the seed finds it again.

    The path of one /seed line

    You type a seed and a question/seed 42 Write a one-line slogan…
    The mod builds a keyFrom three things: the model's name, the seed, and the exact question.
    Is there a saved answer under that key?It looks in the seeds folder of the project.
    No
    Ask the model onceSave the answer with the time and a fingerprint. Show it as a first draw.
    Yes
    Return the saved answerCheck its fingerprint first. The model is not asked. Show it as replayed.

    Each saved answer is one small file that a person can read:

    {
  "seed": "42",
  "question": "Write a one-line slogan for an AI analytics team.",
  "model": "haiku",
  "answer": "\"Turning data into decisions, one insight at a time.\"",
  "drawnAt": "2026-10-05T16:53:28.855Z",
  "sha256": "f3b787d2e728…"
}

    If someone edits that file later, the fingerprint no longer matches the text and the mod refuses to show it.

    Step 4 of 5

    A colleague gets the same words

    A seed is not tied to a chat window. The saved answer is a file in the project, so it goes wherever the project goes.

    How one answer travels

    This project already carries two saved answers, so anyone can check it:

    git clone https://github.com/eshwarpotturi/llm-seed
cd llm-seed
claude --plugin-dir .

/seed 42 Write a one-line slogan for an AI analytics team.

    This needs a Claude Code version that supports mods. It has been run on one computer so far, not yet on a second.

    Step 5 of 5

    What it does not do

    • The model is still random.Only the result for a seed is fixed. Ask without /seed and answers vary as before.
    • A saved answer is one draw.It shows what the model said once, not what it usually says.
    • The question must match exactly.One changed word is a new question and a new draw.
    • It is evidence, not proof.The fingerprint catches a later edit, not someone who rewrites the answer and the fingerprint together.
    • Saved answers are plain text.Anyone who can see the project can read the questions and answers.
    • Text in, text out.No files, images, web search or earlier chat messages go with the question.

    Common questions

    So is it a cache?

    Yes, with three differences that matter. You choose the seed, so you can ask for a new draw on purpose. The saved answers are files that travel with the project. And each one records the time it was drawn and a fingerprint.

    Why not use a real seed?

    Where a model's developer interface offers a seed setting, it is best effort: the same seed usually gives the same output, not always, and it can change when the provider updates the model. Chat tools do not offer it at all. Saving the answer is the only way to get an exact match every time.

    What does the seed number mean?

    Nothing by itself. It is a label. 42, 7 and demo-day all work.

    How do I get a new answer to the same question?

    Use a different seed, or delete that answer's file in the seeds folder.

    Does a replay cost anything?

    No. A replay reads a file. Only a first draw calls the model.

    Does something like this already exist?

    In parts. Developer tools cache model calls, and some model interfaces have a best-effort seed setting. This puts the idea where the work happens, as one command, with answers shared through the project.