Choose a chat that got recurring work right
Look for a chat where your agent produced something you used, or would have, for work that comes up again. The best ones are chats where you corrected it along the way, such as “put the decisions first” or “don’t guess an owner”. Those corrections are the parts of your process the agent couldn’t have known, and the card will keep them. If none comes to mind, look back for one like these:
- a status update you had it redo until the blockers came first
- meeting notes it turned into actions once you told it to leave unnamed owners blank
- at home, a weekly meal plan you adjusted until it fit your budget and schedule
If you don’t have a chat like this yet, build your card from the blank template in Build Your Reusable Workflow Card, or let your agent find your recurring work in Let Your Agent Interview You for Workflow Cards.
Open the chat. Your tool keeps a list of your past chats, and if you had this one inside a Project or Notebook, such as My Agent, look for it there.
In ChatGPT, open Search in the sidebar and type a word you remember from the chat.
In Claude, your chats are listed in the sidebar. A Project’s chats are on its page, under Recents.
In Microsoft 365 Copilot Chat, select the chat in the Chats list on the left.
In Gemini for Google Workspace, open Menu, then Search. Your Workspace administrator decides whether chats are kept, so if past chats don’t appear, use one of the two guides above.
Ask your agent to write the card
Download the template and save it in My Agent Files, or another folder you can find again.
YOUR RESOURCE
Reusable Workflow Card template
Download fileThe blank Reusable Workflow Card Raymond uses in Session 4. Fill in its six parts, Trigger, Input, Steps, Standard, Output and Review, for one process you repeat, and save it with your other agent files.
Ask in the same chat when you can, because the agent there can read every correction. Attach the template or paste its text, then copy the prompt below, fill in the name of the work, delete its last two lines and send it.
If the old chat is very long or wandered off topic, or you did the job across two chats, start a new chat instead. Attach the template, fill in the prompt and paste the conversation in place of its last line, keeping your requests, your corrections and the result you kept. Check the data rules before you paste.
Turn this chat into a Workflow CardLook back over the conversation in this chat, or the conversation I've pasted at the end of this message, where you and I got a piece of recurring work right. Write the process we arrived at as a Reusable Workflow Card, using the blank template attached or pasted in this chat. If you can't see the template or the conversation, tell me which one and wait for it. The work: [name the recurring job in a few words, such as my weekly status update] How to write the card: - Describe how this work should be done every time it comes up, not what happened in this one run. Leave out this run's names, dates, numbers and other facts, and name the kind of input instead, such as "this week's project notes". - Turn each correction or change I asked for that should apply every time this work comes up into a Step, a Standard line or, if it was only about the shape of the result, Output. Use my words where you can. If I changed my mind, use what I settled on. - Write only what the conversation shows: what I asked for, what I corrected and the result I kept. Don't add steps, standards, sections or checks I didn't give. If a part of the card needs something the conversation doesn't show, write TO CONFIRM and what's needed rather than guessing. - Keep the template's sections, headings and order. Keep the opening section, RULE PRIORITY, the rule lists in each part and HOW I WANT MY AI AGENT TO USE THIS WORKFLOW word for word. Replace each bracketed placeholder with my process, using as many lines as it needs. Delete surplus placeholder lines, but if a section would be left without any of my content, write TO CONFIRM there and what's needed instead. Remove the template's three sample lists and their labels (Examples, Example and Standards to consider). Copy a suggested standard into Standard only if the conversation shows it. - Give me the whole card as a downloadable plain-text file named Workflow-Card-SHORT-NAME.txt, where SHORT-NAME is the workflow name in a few words joined by hyphens. If you can't create a file, put the whole card in one code block, with the filename on the line above it. After the card, write a line that says LEFT OUT. Under it, list each request or correction of mine that you kept out of the card, one line each, with where it belongs: "this run only" for a fact or a one-time request, or the file that holds rules for all my AI work, such as my Communication Style Guide or Safety & Security Rules. Keep this list out of the card. CONVERSATION (only if it isn't in this chat) [paste the conversation here in place of this line, or delete this line and the one above]
Expect the whole card as a file or a block of text, with TO CONFIRM wherever the chat doesn’t show what a part needs. A LEFT OUT list follows, naming any request kept out of the card and where it belongs under Session 4’s one best home: a one-time request in a future run’s Context Sandwich, and a rule for all your AI work in the file that holds it, so the card keeps only what’s true every time.
Check every part against the chat
Writing up a long chat, an agent can tidy a correction into something you never said, or keep a fact from that one run. With the chat open beside the card, check:
- Every lasting correction is there, as a Step, Standard or Output line, and anything you asked for only that once is on the LEFT OUT list. If you changed your mind, the card holds what you settled on.
- Nothing is invented. Each line the agent filled in points back to something you asked for, corrected or kept. Delete a line you can’t trace, unless it’s true of how you really work. The template’s own wording, such as RULE PRIORITY and the rule lists, stays as it is.
- No facts from that run. Names, dates and figures would leak into your next result, so replace each with the kind of input, such as “this month’s program updates”.
- Review says what you check. Chats rarely show it, so it’s often TO CONFIRM. Write in what you actually check, and fill in or delete any other TO CONFIRM line.
Ask for each fix in the same chat, such as “Step 2 isn’t from our chat, so delete it”. If a correction on the LEFT OUT list is really part of this work every time, ask for it in the card. A rule on that list for all your AI work stays out of the card: add it only to the file it names, when you next update that file. Then ask for the whole card again as a file or one code block. If every part already checked out, keep the card as it is.
Save the card and run it on a new example
Save the latest whole card as a plain-text file in My Agent Files under the filename the agent gave, pasting it into a new file if it came as text.
Choose a new example of the same work, such as next week’s notes, prepared the way the data rules and your Safety & Security Rules say, or write a short made-up one. Predict what should come back: your card’s Output, your old corrections followed without being asked, and nothing from the old chat’s run.
Start a new chat outside any Project or Notebook, so the card has to carry the process without help from the old chat or your saved files. Attach your card, your Safety & Security Rules and Agent Profile, and your Communication Style Guide if the result is something you send. If you don’t have some of them yet, attach what you have. Copy the prompt below, fill in your workflow’s name and add the new example under Context. Delete the bracketed Constraints line unless this run needs a one-time change. If the agent asks for something your card’s Input lists, answer in the same chat.
Run a Workflow Card and check it against its StandardTask: Run my [workflow name from your card] workflow, using the Reusable Workflow Card in this chat. Context: [paste or attach this run's material: real work your company allows in this tool, or a fictional or non-sensitive version] Constraints: Use the workflow defaults. [add a one-time change, or delete this line] Output Check: Use the card's Output. After the result, write a line that says STANDARD CHECK. Under it, take each line of the card's Standard in turn. Say whether the result meets it, misses it or you can't tell, and quote the part of the result that shows it. For each miss, name the one part of the card (Trigger, Input, Steps, Standard, Output or Review) that most likely explains it, and suggest the smallest change to that part. Do not rewrite the card.
Fix the smallest part that missed
Below the result, a STANDARD CHECK marks each Standard line as met, missed or can’t tell, quoting the result. Check the quotes yourself, because the agent marking its own work isn’t proof. Then check the prediction you made in step 4: your old corrections followed, and nothing from the old chat’s run.
If something missed, change the one part of the card that explains it:
- A skipped or out-of-order action: Steps.
- A quality you’d reject: Standard.
- Right content, wrong shape: Output.
- Missing material: Input.
- A fact from the old run: rewrite the card line that holds it as the kind of input. If no line holds it, leave the card as it is and take the detail out of this result.
Save the change in your card file, because a chat doesn’t change it, then rerun the same example in a new chat. If every part checked out, there’s nothing to fix.
Next time this work comes up, send the prompt again in a new chat with your files and the new run’s details. Save it in My Agent Files if you like, and edit your copy to fit how you run this work. When another chat gets a recurring job right, turn it into a card the same way.
Confirm your result
Confirm the outcomes below to finish this guide.
Final checks
You have a Workflow Card made from a chat that worked, saved in My Agent Files or your own folder and checked against that chat.
It has run once on a new example, and you’ve checked the result against its Standard and for anything carried over from the old chat.
Provide Feedback
Thanks. That went straight to the people who write these guides.
Week 4 field guides
Keep exploring when you’re ready
Revisit this week’s guides whenever they fit something you want to do.