by James M Dewar
This project is unfunded — depending on usage, rate limits may apply.

Summarize and restart

Long conversations make AI less reliable. Facts you gave early on get buried under, or confused with, everything that came after, especially later material that says something different. This guide shows that happening in a short roleplay, then shows how summarizing the conversation into a new prompt and continuing in a fresh conversation brings the facts back.

Steps

In the workshop, these steps are the exercise. You can also follow them on your own.

  1. In Steps 1–3 · Roleplay, press Start the conversation. You give the AI two facts and ask for them back, and it gets them right.
  2. Press Simulate a long conversation. Many turns go by, then you ask for the same two facts, and this time the AI gets them wrong.
  3. Choose Remind the AI or Summarize and restart and follow it to the end, then press Back to the choice and try the other one.
  4. The next time one of your own conversations runs long, paste the prompt from Step 4 · Summarize and restart prompt at the end of it. Check the new prompt it writes, then paste that into a new conversation.

The roleplay is scripted. Real models forget less predictably than this, and the new prompt can miss things too, so read it before you rely on it.

Steps 1–3 · Roleplay

You play the researcher. The AI’s replies are scripted to show a common problem in long conversations and one way to fix it.

Conversation

How full the AI’s context window is0 turns · 0% full

Illustrative, not to scale. The context window is everything the AI reads before each reply, and answers can get less reliable well before it is full.

The conversation appears here. Press Start the conversation below.

You give the AI two facts, then ask for them back.

Step 4 · Summarize and restart prompt

Run this at the end of a long conversation. The AI rewrites the conversation as one new prompt, which you paste into a new conversation.

When to use it: When a conversation has grown long and the AI is getting less reliable.

Copy this

**ROLE**
You are an expert at distilling a long AI conversation into a single, self-contained prompt.

**TASK**
Review our entire conversation above. Produce ONE new prompt that I can paste into a fresh chat to continue this work with full context but none of the accumulated clutter. The new prompt must stand alone — assume the new model has not seen this conversation.

**CONSTRAINTS**
- Capture the goal, the current state and decisions reached, any constraints or preferences I expressed, and the immediate next step.
- Preserve the specific facts, names, numbers, and verbatim text that the next step depends on.
- Omit dead ends, corrected mistakes, and back-and-forth that no longer matters.
- Write it as instructions addressed to the model, not as a summary addressed to me.
- Output only the new prompt, inside a single code block.

Read the new prompt before you paste it. It is all the new conversation will know, so correct anything it got wrong or left out.

The same prompt is in the Prompting library, as Summarize & start fresh.

Inspect the result

  • Does the new prompt keep the facts, numbers and decisions that the next step depends on?
  • Did it carry over anything you had already corrected or dropped?
  • In the new conversation, does the AI answer from the new prompt rather than guessing?

Record

The new prompt you carried over, and the first answer in the new conversation.

When to use this

When a conversation has grown long, when the AI starts repeating a mistake or contradicting what you told it, or before you change topic.

Continue to Maintaining privacy