Write only what you saw and did, in time order: date, time, place, the child by code, what happened, any visible injury, first aid given, who was told and when, and what changes now. AI can tidy staff notes into that shape if the child is already reduced to a code before anything is pasted.
The member of staff who was there owns the account, and the AI is an editor, not a witness. It can fix grammar, split a long paragraph and flag an opinion word, but it cannot know what happened. Every line it adds that the member of staff did not write is a possible invention, and in an incident file an invented line is worse than a clumsy one.
The facts-only structure, field by field
Most weak incident reports fail in the same place: they mix three different things in one paragraph. What the member of staff saw, what they were told, and what they assume. Splitting those apart is the whole craft, and it is also the part an AI tool handles well once you give it the headings to work with.
| Field | What goes in it | What stays out |
|---|---|---|
| When and where | Date, clock time to the nearest five minutes, room or outdoor area, the activity under way | "Around lunchtime" with no time; "the usual spot" |
| Who was present | Child by code, other children by code, staff by role (Room Leader, Assistant) with the ratio in the room at that moment | Full names, dates of birth, home details |
| What was seen | The sequence in short sentences, written as observed | Reasons ("because she was being silly"), blame, guesses at intent |
| What was heard or reported | Another child's account, marked clearly as reported speech | Treating a three-year-old's explanation as established fact |
| Condition of the child | Visible marks, their size and position, whether the child was crying, how long it took to settle | Medical conclusions such as "sprained" or "concussed" |
| Action taken | First aid given, by whom, at what time; whether a cold pack, plaster or observation followed | Treatment that was not given but "would have been" |
| Who was told | Director informed at 10:40; parent phoned at 10:50, spoke to, agreed to collect or not | "Parent was informed" with no time and no method |
| Follow-up | What was checked later, what changes in the room, who owns the change | Vague promises such as "we will be more careful" |
Give staff this table as a one-page card by the first aid box. They fill the narrative on paper or in your childcare software as five or six short sentences under those headings. Only then does anything reach an AI tool, and only the narrative box.
The initials rule that keeps the child out of the tool
The point of a name-free narrative is not tidiness. A child's name attached to an injury, a behaviour concern or a medical note is exactly the data you do not want sitting in a chat history. Our guide on children's data and AI in a nursery covers the wider risk. For incident reports, a short rule is enough.
- Initials plus a number, not bare initials. Initials alone can identify a child with a rare name in a small room. Use a register code instead: the child's initials plus a two-digit number from your own list, such as "TB-07". The list that links the code to the child lives in your register, never in the AI tool.
- Staff become roles. "Room Leader", "Assistant 1", "Cover Practitioner". Staff names are personal data too, and the file already records who was on shift.
- No date of birth, address, allergy list or medical history in the narrative box. If an allergy is relevant to the incident, write "child with a recorded allergy" and let the file hold the detail.
- No photographs. Photos of an injury go in the child's own record only.
- The paste check. Before pasting, read the box once for any surname, nickname, parent name or room-plus-date combination that points to one child.
Which tool? Use an account your daycare controls, with the option to use chats for training switched off, or a business plan where the vendor states it does not train on your content by default. I would not use a free personal account on a staff member's phone for this, however convenient. If you have not yet drawn up what staff may and may not paste, the page on what a daycare can and cannot put into ChatGPT gives the split.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowOpinion words and how to replace them
An inspector, an insurer or a parent's lawyer reads an incident report looking for the point at which the writer stopped describing and started judging. The usual culprits are a small set of words, and a checker can find them for you. Here are the swaps I would build into your prompt.
| Judging wording | Factual replacement |
|---|---|
| "TB-07 deliberately pushed MK-03" | "TB-07 put both hands on MK-03's back and MK-03 fell forward onto the grass" |
| "She was clumsy and tripped" | "She was walking towards the slide, her foot caught the edge of the mat and she fell on her left knee" |
| "A nasty bump" | "A raised area about the width of a fingertip above the left eyebrow" |
| "He was obviously in pain" | "He cried for about three minutes and held his wrist against his chest" |
| "Parent was fine about it" | "Parent said she would collect at 15:30 and asked for a copy of the form" |
| "It was nobody's fault" | Delete. The report records, the director assesses. |
The fingertip is only an example of a size reference. Whichever one your staff use, make it the same one every time so reports compare like with like.
A prompt that edits and never invents
The prompt below tells the AI exactly what it may do and, more usefully, what it must not. The last instruction matters most: it makes the tool list what it could not place, rather than quietly filling the gap.
You are editing a daycare incident narrative written by a member of staff.
Rules:
1. Use only facts in the text I give you. Do not add times, injuries, names,
causes or feelings that are not in the text.
2. Keep the order of events. Use short sentences in the past tense.
3. Replace opinion or blame words with what was seen, only if the text
supports it. If it does not, leave the sentence and mark it [OPINION?].
4. Keep child codes and role names exactly as written.
5. Put the result under these headings: When and where / Who was present /
What was seen / What was reported by others / Condition of child /
Action taken / Who was told / Follow-up.
6. At the end list anything missing that a licensing inspector would expect
(a clock time, first aid detail, a parent notification time).
Narrative:
[paste narrative box here]Using a Claude or ChatGPT Project for this keeps the instructions in one place, so staff paste only the narrative. Claude Projects are available on every plan including Free (up to five projects), and OpenAI says ChatGPT project sharing works on every plan; a shared project on a business plan is the easier one to control.
Main worked example: a playground bump, start to finish
Take an illustrative case in a daycare with a toddler room of twelve children and three staff. A member of staff writes this raw note at 10:55, straight after the event:
Around 10.30 in the garden TB-07 ran into the climbing frame and was really hurt, poor thing, she was being silly and not looking. She had a big bump on her head and cried loads. I put a cold pack on and told the director. Rang mum, she was fine about it.
Run through the prompt, the tool returns something like the following (illustrative output):
When and where: Garden, approximately 10:30; activity not stated. Who was present: TB-07; staff present and ratio not stated. What was seen: TB-07 ran towards the climbing frame and her head made contact with the frame. [OPINION?] "was being silly and not looking". Condition of child: A raised area on her head and crying; the phrase "big bump" has no size or position. Action taken: Cold pack applied. Who was told: Director informed (time missing). Parent telephoned (time and outcome missing). Follow-up: none stated.
Notice what the output leaves alone. It does not fill in who was supervising or what the session was, even though a garden session in a toddler room almost always has a room leader and a known activity. A version that wrote "Room Leader and two assistants supervising" would be guessing, however likely the guess, and that is exactly the line rule 1 forbids.
This is the useful kind of output, because it is a to-do list for the member of staff. They go back and supply the facts only they hold: the Room Leader and two assistants were supervising twelve children; the bump was above the left eyebrow and about the width of a fingertip; she cried for about three minutes; the cold pack went on at 10:33 and stayed for five minutes, and she was playing again by 10:45; the director was told at 10:40; the parent was phoned at 10:50 and said she would not collect but wanted a call at 14:00 if anything changed. They delete "poor thing", "silly" and "fine about it". The tidy final narrative then runs to six sentences and tells the whole story.
The test of the sum: if a stranger read only that narrative, could they say who did what, when, and what happened to the child afterwards? Here they can. In the raw note they could not.
What the AI got wrong, and how it showed up
In practice the failures are small and easy to miss, which is why the original notes stay on file.
- The helpful insertion. Asked to "make this more professional", a tool can turn "cried loads" into "appeared distressed and in considerable pain". Nobody saw pain; the member of staff saw crying. The fix is rule 1 of the prompt and a human comparison of before and after.
- The tidy timeline. A tool can rearrange two sentences into a clean order that was not the order things happened. Say staff phoned the parent before applying the cold pack, and the edited version reverses it. Compare the sequence, not just the sentences.
- The invented follow-up. A draft might end with "Staff will review supervision levels in the garden." Nobody decided that. If the follow-up is blank, leave it blank and ask the director.
- The softened second child. In a biting incident, a tool may write that the second child "may have been provoked". Remove it; that is blame.
Three more incident types, three different traps
Biting between two children
Each child gets their own report, referencing the other by code only. The parent of the bitten child hears that "another child" was involved, never the name. The narrative states the position of the staff at the time and whether the child had shown signs of frustration earlier, but only if somebody saw it. The common mistake is writing the biter's history into the bitten child's report.
Allergic reaction
The narrative gives the time the food was served, the time symptoms were first seen, what was observed, the medication given, the time of the emergency call if made, and the time of the parent call. It does not say "an allergic reaction" as a conclusion unless a medical professional has said so; write "red marks on the cheeks and lips and repeated coughing" and let the record show the plan that was followed. This is the one report where timings matter minute by minute, so ask the tool to flag any missing clock time.
A child collected by the wrong adult, or nearly
These are near misses, not injuries, and they are the reports daycares most often skip. Write who arrived, what the authorised-pickup list said, what the member of staff did, who was contacted and when the child was released or kept. The follow-up is the point: what changed at the door. Leaving a near miss unwritten is how the same gap opens twice.
The parent notification: wording that stays factual
Telling the parent is a separate act from filing the report, and both are logged. For a minor injury, the phone call comes first, and the written note follows the same day. A draft for the written note, using placeholders, might read:
Dear [parent name], I am writing to confirm our call at 10:50 today. At about 10:30 in the garden, [child's first name] ran into the climbing frame and bumped her head above her left eyebrow. She cried for about three minutes. We applied a cold pack at 10:33 and she was playing again by 10:45. We will keep checking her through the day and call you if anything changes. A copy of the incident form is attached for you to sign and return. [your name], Room Leader.
Two rules for the wording. State what happened and what you did, not what you think it means medically. And never name other children. AI can draft this from the final narrative in seconds, and because the narrative is already name-free you add the real names yourself in your childcare software or email, after the AI step. That ordering is what keeps the child out of the tool.
A head injury note usually needs a line on what to watch for after collection. Take that wording from your own health policy or a first aid trainer, not from the AI, and make it the same text every time.
Licensing, reporting and retention: look it up for your state
Rules differ between states and by type of setting (home-based or centre-based, group size, funding). I will not quote a number of days or years here, because a wrong figure in an incident procedure is a worse problem than none. What you need to find, from your state's child care licensing rules or your licensing officer, is below. Put the answers in a one-page table in your procedure with the date you checked.
| Question | Where to look | Your answer, date checked |
|---|---|---|
| Which incidents are "reportable" to the licensing office? | Licensing rules or the licensing agency's incident guidance | (fill in) |
| How fast must a serious incident be reported? | Same | (fill in) |
| Must parents be told in writing, and when? | Rules plus your enrolment contract | (fill in) |
| How long must reports be kept, and from what date? | Rules, insurer, any subsidy programme you accept | (fill in) |
| Who may see them (inspector, parent, other)? | Rules and your privacy notice | (fill in) |
| Does your insurer want notice of certain injuries? | Your policy wording or broker | (fill in) |
An AI tool can help you read the rules, with a caution. Paste the relevant section into a Project and ask it to list the deadlines, but then check the list against the original text. These are rules where a misread number has consequences. Where the rules are unclear, ask your licensing officer in writing and file the reply.
The director's sign-off, in two minutes
A report that nobody senior reads is a record, not a control. A short sign-off turns it into the second one. Print this at the bottom of every form, or add it as the last fields in your childcare software:
- I have compared the final narrative with the member of staff's original notes. Yes / No
- The narrative contains no opinion, blame or medical conclusion. Yes / No
- Times for first aid, director informed and parent informed are all present. Yes / No
- The incident meets the reporting threshold in our procedure. Yes / No, and if yes, reported on (date and time)
- Follow-up action, owner and due date are recorded.
- Signed, dated, and the parent's copy issued.
The director is the person who catches the AI's helpful insertions, because they have the original beside the final. Make comparing the two a stated step, not an assumption.
A monthly audit of how the routine is working
Once the routine exists, check it. Every month, pull ten recent reports at random and score each from 0 to 6: all times present, no opinion words, sequence matches the original, parent contact logged with method, follow-up has an owner, director signed. A score under five on any report is a retraining note, not a telling-off.
A rough sum shows why it pays. Say a daycare has about six reports a month and each takes a member of staff 15 minutes to write cleanly and the director another 10 to correct. If the structure and AI edit bring that to 10 and 5, the saving is about 10 staff minutes per report, an hour a month. The money is small. The figure worth tracking is how many reports reach the director with a missing time: count them each month, because the AI's list of missing items should catch most of those gaps before a draft ever reaches the director's desk.
When your files are tidy, an inspector's request to see "incident reports for the last quarter" is easy to answer, which links directly to preparing your records for an inspection. If something goes badly wrong, the wider AI incident response plan covers when the problem is the AI use itself rather than the child's injury.
What I would keep human, always
Deciding whether an incident is reportable. Deciding whether a pattern across several reports means a room needs changing. Talking to a parent whose child was hurt. And any serious injury, where the first account should be written by the person who was there, with the director, before anything else touches it. AI is useful for the tidy-up that follows, and not for the judgement.
If you would rather set the routine up with someone looking at your own forms and software, that is what my AI implementation consultation covers.
Questions staff ask before using AI on an incident report
Can I paste the whole incident form into ChatGPT?
Not as written. The form carries names, dates of birth and sometimes medical details. Replace the child and any other child with codes, swap staff names for roles, remove the date of birth, and use an account your daycare controls with training switched off. Then paste only the narrative box, never the header.
Should the AI-polished version replace the original staff notes?
No. Keep the member of staff's own first account, written at the time, alongside the final report. If anyone ever asks how the report was produced, the original notes show what was seen, and the edited version shows what was tidied.
How soon should a report be written after an incident?
The same shift, while memory is fresh. Quick notes at the time, then the narrative before the person leaves for the day. Licensing rules in some states set a deadline for serious incidents, so look up yours and put the number in your procedure.
Does a parent get the same report the file holds?
Usually the parent receives a notification or a copy of the incident form for their own child, with other children's names removed. What you must give, and how fast, comes from your state's rules and your enrolment contract, so check both rather than guessing.
Further reads
- AI for Daycares: Nine Admin Jobs From Enquiry to Pickup Message — Eight other admin jobs where AI helps and where it stays out.
- How Nurseries Use AI to Write Daily Reports for Parents — Daily parent reports use the same name-free habit.
- Nursery Staff Training: A One-Hour AI Briefing on Children's Data — A one-hour briefing that teaches staff the initials habit.
- Early Years AI Policy Template for Nurseries and Childminders — Put the rules for AI use into a written policy.
- Daycare Late Pickup Policy: Can AI Word the Fees Fairly? — Another policy where plain, neutral wording protects you.
- Nursery Data Protection Policy: Draft and Update It With AI — Keep children's records covered by a current data policy.
- Can AI Draft a Martial Arts School Waiver for Children? — A children's martial arts waiver drafted with AI in half an hour, then the consent, injury, photo and data checks a solicitor must make.
- AI for Care Homes: A Registered Manager's 90-Day Implementation Plan — A month-by-month AI rollout for a registered manager, with owners, measures, budget and an approved-tool list for a 40-bed home.
- Adult Day Care Software for a Small Centre: Billing and AI Notes Checklist — A grouped checklist for a small adult day centre choosing software for attendance, transport, billing and care notes, with AI drafting kept under review.
- After School Club Booking System: Bookings and AI Reminders — Compare free, generic and dedicated booking tools for an after school club, then use AI to draft reminders without exposing children's details.
- Brightwheel vs Procare: Which Daycare Software Has Real AI? — Brightwheel's AI assistant against Procare's RoomRunner and Camber, with quote-only pricing questions and a scored demo for a small daycare.
- AI Receptionist for a Daycare: Tour Bookings, Calls and Pricing — What an AI phone receptionist costs a daycare per minute, per call or per caller, which calls it should take, and a test script to run first.
- AI for Disability Services: Service Agreements and Incident Logs — Draft service agreements from your own template and structure incident logs with AI, while keeping participant details out of the prompt.
- Can AI Write a Care Home Incident Report? Template and Checks — A factual care home incident report template, prompts that tidy staff notes without adding facts, and the manager checks AI must never replace.
- Can a Nursery Use AI for Accident Reports? What to Keep Out — A facts-only workflow for turning keyword notes into accident report wording: a child code rule, parent message wording and a sign-off and reporting check.
- Child Care Subsidy Billing: Can AI Match Payments to Attendance? — How a daycare can match subsidy remittances to attendance with AI-built formulas, find shortfalls and write top-up notices without names.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Claude help-centre article on projects and OpenAI help-centre article on projects in ChatGPT (plan availability and sharing, checked November 2026); general childcare licensing practice; your own state's child care licensing rules and your insurer's guidance should be checked for incident definitions, reporting deadlines and retention periods.