How to Use AI for Coaching Session Notes

You finish the last session at 4:30. The notes get written at 9pm, from memory, and they come out as two lines and a feeling.

Here is how to use AI for coaching session notes. You type your rough notes, an agent that follows your rules turns them into a summary, the client's action steps and the focus for next time, and you check every line against what the client said before it goes in the file.

That last part is the job.

A model turns scribbles into tidy paragraphs in seconds. It turns your suggestion into the client's commitment just as fast, and a tidy paragraph hides that better than a messy one.

Twelve clients a week is a second job nobody pays you for. The setup below gets each note down to a few minutes without handing over the one thing a client hired you for, which is your judgment about what mattered.

The notes are yours. The model drafts and you sign.

You will watch the failure happen on an invented client, fix it once in a reusable agent, and then run every client through the same fix.

Tell the client and strip the names before the notes leave your desk

Do this before you paste a single line, and do it once.

Session notes hold the fragile parts of a person's life: the dread before a presentation and the money fight at home. Where those words go matters more than how neat the summary reads.

The International Coaching Federation's Code of Ethics puts the duty in Standard 2.5. You meet your ethical and legal obligations to clients "directly and through any technology systems I may utilize," and the list that follows names artificial intelligence. If you are not ICF credentialed, you have no reason to hold a lower bar.

So tell clients before the first session that touches a model, and put it in your agreement in plain words. Business Insider's profile of three solo coaches describes Katharine Campbell Hirst recording sessions and uploading transcripts to ChatGPT after getting client consent. Consent comes first in that sentence, and it should in your practice too.

Four things belong in the disclosure:

1. That you use an AI workspace to help draft your session notes.
2. What goes in: your own typed notes, not a recording.
3. That you read and correct every draft, and the final note is yours.
4. That the client can opt out, and what you do instead.

Here is a paragraph you can paste into your agreement and change to fit:

I use an AI writing tool to help draft my session notes. After each session I type my own rough notes, and the tool turns them into a summary and next steps. I use only your first name, and I read and correct every draft before it goes in your file. If you would rather I not use it, tell me and I will write your notes by hand.

Then decide what stays out of the prompt. Use first names only. Leave out the employer, the dollar figures, the diagnosis and anything about a third party who has not agreed to be in your notes.

Stripping is not magic, since a first name plus a VP's title can still point at one person. Keep the detail down to what you need next session.

In Magai, the prompts and files you send are never used to train any model, from any provider on the platform, and your conversations stay private to you and the teammates you invite. You can say that to a client. It is one line of the disclosure, not the whole of it.

Start with your rough notes, not a cleaned up version

The temptation is to tidy the notes first. Resist it.

Tidying is the 9pm work you are trying to get rid of, and it strips out the signals a model needs most: the "maybe" and the pause.

Type them the way you already do. Shorthand, arrows, half sentences. Do it within the hour if you can, because after a night's sleep your own memory turns "maybe" into "she'll do it."

One habit pays for itself. Put the client's exact words in quotation marks, and leave your paraphrases bare. The agent you build in a minute reads that difference, and it is the only way it can tell what the client said from what you remember her saying.

You do not need much. Seven lines carry a session:

1. The client's goal for the session, in their words.
2. What shifted.
3. What the client agreed to do, with a day.
4. What you suggested that the client did not accept.
5. What is still open from earlier sessions.
6. What you want to open with next time.
7. Anything for admin, like a renewal or a date to schedule.

Here are session notes for a client called Priya. Priya is invented, and so is everything she says in this post. I wrote the notes the way a coach writes them at 9pm, with five traps in them on purpose.

Priya, session 6, Tue
wanted to talk about Q1 offsite. dreading presenting to VP team
freezes when Lena (VP) interrupts. happened at last review
did the worst-case exercise. she laughed, said worst = awkward 30 sec
doing all the prep at midnight, no wonder
I said maybe send the deck to Lena ahead. she said "maybe"
she'll run through opening 3x before Thurs, out loud, phone recording
check: did she talk to Marcus re headcount? NO. dodged again. 3rd time
she said she wants to be someone who doesn't need to be liked by everyone. big one. pause.
next: offsite is the 14th. start w/ how Thurs went
invoice pkg 2 renewal at session 8

Read them as a model will. Each trap is a place where a tidy summary goes wrong:

1. The deck. A suggestion nobody accepted, with a "maybe" on it.
2. Marcus. A conversation that was dodged, not agreed to.
3. The "liked by everyone" line. It is the turning point of the session, and it is my paraphrase, with no quotation marks.
4. "No wonder." My own aside, a judgment that must never reach the client.
5. The invoice. Admin that has no business in a client's summary.

Real notes carry more traps than that. A date with no month, or a name with no last name.

Open a new chat in Magai for this client, one chat per client, and keep it. The reason shows up at the end. Paste the notes in.

Run the plain prompt first and watch where it goes wrong

The fastest way to see why you need rules is to run without them. Under the notes, type this:

Turn these session notes into a summary, action steps for the client and a focus for next session.

This is the reply, from /Claude:

Summary: Priya's focus this session was her upcoming presentation at the Q1 offsite, where she expects to freeze when Lena, the VP, interrupts. Using a worst-case exercise, Priya recognized the worst outcome is a 30-second awkward pause. She also shared a desire to become someone who doesn't need to be liked by everyone.

Action steps:

1. Send the deck to Lena ahead of the offsite.
2. Rehearse the opening three times out loud before Thursday, recording on her phone.
3. Talk to Marcus about headcount.
4. Stop doing prep at midnight.

Next session focus: Review how Thursday's rehearsal went, then prepare for the offsite on the 14th. Address Priya's pattern of avoiding the headcount conversation with Marcus. Discuss renewal of Package 2 at session 8.

It reads well. That is the problem.

Four things are wrong, and none of them show at a glance:

1. Three of the four action steps are not Priya's. She said "maybe" to the deck, she dodged Marcus, and the midnight prep was my aside. The model promoted all three to commitments.
2. "Priya's pattern of avoiding" is my private hunch, written into the plan as fact. If Priya reads this, she reads a verdict.
3. The renewal sits in a document she might receive.
4. The line she paused on is a clause at the end, with no sign that my notes only paraphrased it.

The same failure shows up outside coaching. Ava Cloghessy, a writing consultant at the College of the Holy Cross, tested ChatGPT on her own session notes and wrote up the result: it "generally cannot distinguish the student's major concerns nor what the consultant believes is most relevant." That is a writing center, not a coaching practice, and the limit is the same. A model reads your words. It was not in the room.

Notice what it did well, too. The summary is accurate about the topic and the structure is clean. You keep all of that. You fix the judgment with rules, and rules belong somewhere that remembers them.

Use AI for coaching session notes with an agent that holds your rules

A plain prompt forgets your rules the moment the chat ends. An agent does not.

In Magai, an agent is a set of reusable instructions. It holds your voice, your rules and your format, it runs on whichever model does the job best, and you can share it with the associates in your practice. The instructions work like a system prompt that travels with the agent, and how to build an AI agent covers the long version. If you do not have an account yet, you can start at app.magai.co.

Make a new agent, name it Session notes, and paste in the instructions below. First, the shape of what it writes. The agent produces six sections, and each one has a reader:

Now the instructions:

You are my session notes assistant. I am a coach. After each session I paste my rough notes. You turn them into a summary, the client's action steps and the focus for the next session.

Rules you never break:
1. Use only what is in my notes. If something is missing or unclear, write [Check: what is missing] and move on. Never fill a gap with a guess.
2. An action step is something the client agreed to do. If my notes show only that I suggested it, or the client said "maybe," "I'll think about it" or nothing, it is not an action step. List it under "Suggested, not agreed."
3. Write action steps in the client's voice, in first person ("I will run my opening three times"), with a day when my notes give one.
4. Words in quotation marks in my notes are the client's exact words. Anything without quotation marks is my paraphrase. Never put my paraphrase in quotation marks.
5. If a line in my notes marks a turning point ("big one," "pause," "she went quiet"), include it in the summary as the client said it and do not interpret it. If it is my paraphrase, add a [Check] asking for the exact words.
6. Keep my observations, patterns and hunches out of anything the client will read. Put them under "Coach only," labeled as my read, in one or two lines.
7. Keep invoices, renewals, scheduling and anything about money between us under "Admin," never in the summary.
8. Never diagnose, label or explain the client's psychology. Report what was said and what was decided.
9. If my notes suggest risk to the client or to anyone else, stop. Write "Risk flag" at the top, repeat my words exactly and do not summarize around them.

Output, in this order:
Summary: three to five sentences, plain, past tense, written so the client could read it.
Action steps: agreed only, numbered.
Suggested, not agreed
Next session focus: one opening question and up to two topics, all taken from my notes.
Coach only
Admin
Checks: every [Check] item in one list.

Voice: warm and brief. No coaching jargon, no praise words like "great" or "powerful." American spelling. Short sentences.

Each rule answers one of the four failures above. Rule 2 is the big one, because it stops a suggestion from turning into a promise. Rules 4 and 5 handle the line she paused on. Rules 6 and 7 build the wall between what the client reads and what you keep. Rule 9 is the one you hope never fires.

The voice line is thin on purpose. It gets thicker when you paste in two or three follow-up emails you have already sent, under a heading that says "My voice." Cloghessy did the same thing with her archive of past notes and saw ChatGPT start borrowing phrases like "going forward" and "best of luck" that she used herself. Your emails do the same job. The agent learns your sentences, not a generic coach's.

Before a real client goes through it, test the agent on a session you wrote up by hand last month. You know what happened, so you can see every place the agent disagrees with you.

The first run will miss something. When it does, change the instructions and not the output. Add one rule for each miss, and stop adding at a dozen or so, because a rulebook you cannot hold in your head stops being read.

Run the same notes through the agent and read what comes back

Open a new chat with the agent selected and paste the same notes. No extra prompt. The instructions already say what to do.

This came back, again from /Claude:

Summary: Priya came to talk about presenting at the Q1 offsite, where she expects to freeze if Lena, the VP, interrupts. It happened at her last review. We ran the worst-case exercise, and Priya laughed and named the worst case as about thirty seconds of awkwardness. She has been doing her prep at midnight. Near the end she said she wants to be someone who doesn't need to be liked by everyone.

Action steps:

1. I will run my opening out loud three times before Thursday, recording it on my phone.

Suggested, not agreed:

• Send the deck to Lena ahead of the offsite. Priya said "maybe."

Next session focus: Open with: "How did the three run-throughs go?" Then the offsite on the 14th, and whether to send Lena the deck.

Coach only: Your read: the headcount conversation with Marcus has been dodged three times. Midnight prep may be feeding the freeze.

Admin: Package 2 renewal at session 8.

Checks: [Check: your notes paraphrase her "liked by everyone" line. Add her exact words if you have them.] [Check: the offsite is "the 14th," with no month.]

Read it the way you read the first one, line by line.

What it kept: the one real commitment, written in Priya's voice. The "maybe" is fenced off where it cannot become a promise. The renewal is out of her copy. My hunch about midnight prep is labeled as mine.

The checks are the part I did not write specific rules for. The month check comes from rule 1, which says to flag what is unclear, and the agent applied it to a date. It followed rule 5 on the paraphrased line and asked for her exact words instead of dressing up mine. That is the reason to tell a model what to do when it does not know.

What I asked for again: one thing slipped. Marcus shows up only as a hunch under Coach only. The open thread itself, a conversation Priya has put off three times, is not carried to next session, and the opener never mentions it. My notes were clear that this matters. The agent had nowhere to put it.

So the agent gets a tenth rule and a new section. Add this to the instructions:

10. If my notes mention a topic from an earlier session that is still open, list it under "Carried forward" with how many times it has come up. Do not move it into action steps or suggestions.

And add "Carried forward" to the output list, after "Suggested, not agreed." Run the same notes again and a new section appears:

Carried forward: Talk to Marcus about headcount. Raised three times, not yet done.

Now the thread has a home, and you will see it at the top of next week's draft whether or not you remember it.

Check the draft against your notes before it goes in the file

The agent drafts. You read the draft against your notes, line by line, with four questions:

1. Said. Did the client say this, or did the model smooth two things into one?
2. Agreed. Did the client choose this step, or did your suggestion get promoted?
3. Safe to read. Would you be comfortable if the client read this sentence tomorrow morning?
4. Missing. What did you underline in your notes that the draft dropped?

Most of the time the first two questions catch it. Questions three and four catch the quiet failures, like a sentence that is accurate but cold, or a pause that did not survive the summary.

Then use the thing a single-model app cannot do. Switch models inside the same thread and ask a second one to audit the first. Type /GPT at the start of your message:

/GPT Here are my rough notes and the draft the agent wrote from them. List every statement in the draft that my notes do not support, and every action step the notes do not show the client agreeing to. Quote the line, then say which part of my notes it came from, or that it came from nowhere.

The thread keeps its memory across the switch, so the second model sees the notes and the draft without a second paste. I am not going to tell you which model catches more, because it changes with your notes and your habits. Run the audit on a week of real notes once, count what each model caught, and keep the one that caught more. Switching models in one workflow covers the mechanics.

Here is the objection you may be holding. If a model writes the note, do you stop thinking about the client?

Some coaches say the writing is how they think, and that is fair for the sessions where it is. Others find the 9pm writing is mostly typing.

The thinking was never in the typing. It lives in the check and in the Coach only section, which repeats your read and never adds one. If you skip the check, you stop thinking about the client. If you keep it, you think about them twice.

How much time does this save? One leadership coach, Kim Surko, told Business Insider, "I've saved almost an hour per client per day by relying on AI to take notes and summarize them for me." That is her figure from her own practice, and yours will differ. Time your first five clients start to finish, from paste to signed note, and trust that number over hers.

Turn the summary into the follow-up the client should have had yesterday

The notes are for you. The follow-up is for Priya, and it is the one thing on your list that clients notice when it is late.

In the same thread, with the agent still on, type:

Draft the follow-up email to Priya from the summary and the agreed action steps only. My voice, under 130 words, plain text. Leave out Coach only, Admin and Carried forward.

The reply:

Subject: Thursday's run-throughs

Hi Priya,

Thanks for today. You named the worst case at the offsite as about thirty seconds of awkward, and you said you want to be someone who doesn't need to be liked by everyone. I'd hold on to both of those.

You're running your opening out loud three times before Thursday and recording it on your phone. When we meet next, I'll start by asking how those went.

The deck and Lena are still open. No decision needed yet.

[Your name]

Judge it as the coach, because the model has no way to. The structure is right. It stays inside the summary, so nothing from Coach only or Admin leaked. It kept Marcus out, which is correct, since that conversation is a thread you hold and not a topic for her inbox.

But "I'd hold on to both of those" is the model putting advice in your mouth. Priya said the line. Whether you underline it for her is your call.

Cut that sentence, then add an eleventh rule so you never edit for it again:

11. When I ask for a follow-up email, use only the summary and the agreed action steps. Never add advice, praise or interpretation that is not in my notes.

Whether to quote her "liked by everyone" line back to her at all is also your decision. For some clients it lands. For others, seeing it in writing three hours later feels like being watched. The agent cannot know which client Priya is. You can.

The same thread does the next job too. Before session 7, type:

Prep me for session 7 with Priya. From this thread, list what is carried forward, what she agreed to do, and the opening question. Do not add anything that is not in my notes.

You get the open threads, her commitments and your opener in one place, drawn from what is already in the thread. Read it on the way into the room instead of rereading six weeks of notes.

Copy the follow-up into your own mail and send it from your own address. Keep the send step manual for the first month, until you trust the drafts.

Know which sessions never go through the model

Some sessions should not go through a model at all, and the agent will not be the one to tell you.

If a client discloses something that suggests risk to themselves or someone else, you do not draft that summary with a tool. You follow what your training, your agreement and the law where you practice require. Rule 9 in the agent is a backstop for the sentence you missed at 4:30. Your own judgment is the plan.

The same goes for sessions that drift toward therapy. Notes about trauma, diagnosis or medication belong in whatever record your scope allows, and a general workspace is a poor place to decide what that is. Do not ask a model whether a client needs a referral. That judgment is yours, and so is the paperwork around it.

Employer-sponsored work needs a second look. If a company pays for the coaching, your agreement with the sponsor may limit what you can store or share about the client, and no model knows what you signed. Check the agreement before the first note goes in.

Then there is the client who says no. Write their notes by hand and leave a plain line in the file that says so. A no is a normal answer and should cost the client nothing, which means you do not run their notes through the model quietly "just for the summary."

I will concede where this approach stops being true. For a hard session, the one where you drove home quiet, writing the note by hand at 9pm is the right call, and the agent can wait. It is built for the Tuesday afternoon sessions, the ones that went well and left twelve things to remember.

Set it up once and the next session takes five minutes

You built three things today: an agent with its rules written down, a consent paragraph and a habit of putting quotation marks on exact words. None of them need to be built again.

For every client, keep one chat. Put those chats in a folder, since Magai's folders and collections exist to keep chats organized, and the notes for Priya, session 7, go in the same thread as session 6. The thread keeps its memory across model switches, which is why the second-model audit above needs no second paste.

Do not lean on that memory alone. The Conference Board's October 2025 research on AI coaching found that general-purpose models "retained context from previous interactions but were inconsistent." So paste last session's Carried forward list above the new notes every time. It takes five seconds, and it means the list is in front of the model whether or not its memory holds.

A rhythm that holds up: type each set of notes while the session is fresh, run them all through the agent after your last session, read each draft against your notes, and send the follow-ups that evening. Once a month, look at the checks that keep coming back. Each one is a rule you have not written yet.

If you work with associate coaches, share the agent with them. Everyone's notes then come out in the same sections, and the rules that protect your clients sit in one place instead of living in your head. More on Magai for coaches lives on the coaches page.

The same pattern works for other client paperwork. Client reporting uses it for a different job: paste the raw material, let the agent shape it, and check it against the source.

Next Tuesday, after the last session, this is all you type:

Session 7 notes for Priya. Last session's Carried forward is above.

[your rough notes]

Type the mess. Read every line. Sign it.