Save decisions and instructions from an AI conversation

Find what was agreed, check it, then save it

1. 01

   Give your assistant the conversation

   To separate an agreed decision from a passing suggestion, the assistant needs the messages themselves. A remembered summary can drop who said something or whether you accepted it.

   Open the conversation you want to preserve. If the assistant cannot read the messages, paste them into the chat and ask it to list any missing parts. A data export (the account download you request in Settings, sent by email) also works, but it holds your whole history.

   Check: The assistant can point to the messages it is using.

   Remove private details the task does not need before pasting the messages.

   

2. 02

   Ask for decisions and their supporting quotes

   You will check every record before anything is saved, so ask for a quote beside each one. The quote lets you confirm a record's meaning without rereading the whole conversation.

   Paste the extraction prompt into the chat that holds the conversation. It asks the assistant to separate what you agreed to from what it only suggested.

   Extract the useful decisions with your assistant

   Extract useful records and keep their evidence attached.

   Problem: Useful decisions and instructions are buried in an AI conversation I want to preserve. Intent: I want records with supporting quotes that I can check before saving. Use the conversation I identify or the transcript I provide. If you cannot read the messages you need, ask me to paste them. Do not assume that your chat memory or a summary contains the whole conversation. Extract decisions and reasons, next steps, standards, workflows, fixes, open questions, and clearly stated preferences. Skip empty categories. For each proposed record, provide: - Its type and a statement that makes sense on its own - A short direct quote and who said it, with a message or section reference - The conversation title or identifier, source link, and date when available; mark missing values unknown without inventing them - Scope: where the record applies, and any stated time limit - Status: supported, uncertain, rejected, or needs my confirmation Record a decision only when I make or accept it. Label plans separately from finished work, and one-time requests separately from standing instructions (instructions meant for every future task). Mark personal preferences, values, standing instructions, and statements about how I write or sound as needs my confirmation. Do not infer them from tone. Check the proposed records against the transcript again. Repair or reject any record whose quote does not support the whole claim. List any part of the conversation you could not read. Show the records for review before saving them.

   Check: Each proposed record contains one clear point and a quote that supports it.

3. 03

   Check what was agreed

   Some proposed records can claim more than the conversation said. A quote must support the whole record: “We could try weekly reports” does not mean “We decided to send weekly reports.” Another agent may follow preferences and standing instructions (instructions meant for every future task) in your name, so confirm those records yourself.

   Compare each record with its quote, and search the conversation for any quote you do not remember. Correct overstatements, reject records whose quote you cannot find or that claim more than it says, and confirm personal preferences before saving them.

   A one-time report request does not establish a lasting writing preference

   Use this pattern with quotes from your conversation. Do not save this example.

   Source excerpt
   User (message 4): "Use the support-ticket export as the source. Draft the summary in weekly-update.md, and keep unresolved refund cases in a separate list."
   
   Supported one-time instruction
   - Record: For this summary, use the support-ticket export, draft in weekly-update.md, and list unresolved refund cases separately.
   - Evidence: "Use the support-ticket export as the source. Draft the summary in weekly-update.md, and keep unresolved refund cases in a separate list." (User, message 4)
   - Scope: This support-summary request
   - Status: Supported
   
   Rejected record
   - Claim: The user prefers short reports.
   - Reason: The source says nothing about report length.

   Check: You approve the records to save, and unanswered questions remain open.

   Keep a decision’s reason when the conversation gives one. It helps you understand the choice later.

   

4. 04

   Choose where the records belong

   The checked records are easier to reuse in your notes folder, next to the files you already consult for this project. A new, unrelated folder is easy for the next agent to miss.

   Open your notes folder in Codex (in the ChatGPT desktop app), Claude Code, or Claude Cowork (both in the Claude desktop app on a paid plan; in the newer combined Claude layout, any conversation can work in a folder). Tell the agent it may file the checked records there, and name any files it must leave alone.

   Check: The agent identifies the existing files that could hold the records.

   Open only the folder this task needs, and review the app's permissions. An agent can change or delete files in that folder, and some apps can also reach your browser or connected services.

   

5. 05

   Ask the agent to file the checked records

   Each approved record now needs a home in your files. The agent picks a file for each one, keeps its supporting quote, and checks that the record is not already saved.

   Give the agent the checked records and the filing prompt. Ask it to reuse your existing organization and report where it saved each record.

   Save the checked records with your agent

   Run this after opening the folder that should receive the records.

   Problem: I have reviewed records from a conversation and want to save them in my project files. Intent: I want the approved records in useful existing files, with their sources preserved. Use the checked records and destination folder I provide. If either is missing, ask for it before editing. Confirm which folder this task may change. If you cannot edit files here, ask me to open that folder in an agent with file access and provide these same directions there. Check how the existing files are organized, then file supported records in the most relevant files. Create a file only when no existing file is suitable. - Do not save uncertain or rejected records as facts - Preserve available source dates, conversation identifiers, quotes, scope, and status; leave missing source details marked unknown rather than inventing them - If a missing detail prevents matching a record to its source, ask before saving that record - Save personal preferences, values, standing instructions, or statements about how I write or sound only when I have confirmed them - Reuse existing names and structure; do not create a folder system for this batch - Ask before a placement that widens who can read a record, puts it in someone else's file, or changes its meaning - Do not copy private details into a more widely shared file or edit the app’s built-in memory (for example, Codex memories) Compare what you saved with the approved records and quotes. Repair mismatches supported by those sources. Then list changed files, added records, skipped records, and decisions still needed.

   Check: The approved records appear in named files with their sources attached.

6. 06

   Read the saved records once more

   If the agent drops a record's reason or its “uncertain” status while filing, the record reads as more settled than it was. Check that the saved files hold the records you approved and sit in your notes folder.

   Open the changed files and compare the additions with the checked records. Ask the agent to fix missing quotes, duplicate records, or a suggestion presented as a decision.

   Check: The saved records preserve the decisions, reasons, and open questions from the conversation.

   Do not ask the agent to edit an app’s built-in memory, the details Claude, ChatGPT, or Codex saves automatically when memory is turned on. These records belong in your files.