· Early Childhood Teacher
Last week a three-year-old in our pre-school room discovered that his shadow stretches long at sunset. He chased it across the yard for twelve minutes and built his own theory out loud: the sun pulls it long. When I fed my rough notes to an AI assistant, the draft said he “explored shadows during outdoor play.” Accurate. Dead. This post is how I write EYLF learning stories with AI now: the five-block format I keep every single time, and the moment-specific child voice I rewrite by hand before anything goes near a family.
The learning story format I never change
Here’s the part AI is genuinely good at: structure. I hand it the same five blocks for every story, and I never let it reorder them or pad them out.
- Title — short, the child’s name, the moment.
- Narrative — what happened, in plain language.
- Analysis — what the moment shows about this child’s learning.
- EYLF V2.0 outcome link — the outcome, tied to evidence in the narrative.
- Next steps — what I’ll set up next to extend it.
That order is the spine. The model fills it reliably, which saves me the blank-page stall I used to get at 4pm with eleven stories still open.
Why I keep word count to ~150–250
A learning story is not an essay. In my practice the sweet spot is roughly 150 to 250 words. Long enough to hold a real moment, short enough that a parent reads it on their phone at pickup. I tell the model the ceiling up front, because left alone it will happily write 600 words of warm fog (I learned that the slow way, editing one back down at 5pm). Families don’t read fog.
The exact prompt I paste my rough notes into
My notes are never sentences. They’re a dump: “Kai, yard, ~4:40pm, sunset. Shadow. Chased it 12 min. Said sun pulls it long. Tried to step on it, laughed when it moved with him.” Bullets, fragments, the bits I’d forget by the time the room was packed up.
Telling the model the format and the framework
I name the framework explicitly. EYLF V2.0, the Australian one, not US or UK standards. Pin that or the model drifts toward generic “developmental milestones” language that means nothing to an ACECQA assessor. Five blocks, the word ceiling, a strength-based tone. That’s the whole prompt.
Privacy: what I strip before anything touches AI
This is the non-negotiable one. Before a single note goes into a hosted AI tool, I strip the child’s real name for a placeholder, any sibling or family detail, and obviously no photos or faces. Our service treats third-party AI tools as external disclosure. So I check our enrolment privacy collection notice and ask my director before any identifying detail leaves the building. Ask your supervisor where your service draws that line. It varies, and it matters.
Why the first AI draft always sounds wrong
The first draft is structurally perfect and emotionally empty. There’s a tell I now spot in two seconds.
The generic-voice tell
If a sentence could be pasted into any child’s story without changing a word, it’s filler. “Kai is a confident and capable learner who engaged in rich play.” Sure. So is every other three-year-old in the country. That line says nothing about the shadow, the twelve minutes, or his theory about the sun.
Outcome links that are decoration, not evidence
The model loves to staple outcomes on. It claimed this moment showed Outcome 3 (a strong sense of wellbeing) because Kai was “active outdoors.” That’s decoration. There was no wellbeing evidence in what I actually saw. There was a child building a theory. So I cut it.
The voice I rewrite by hand
Here’s the keep-versus-rewrite split. I keep the format. I rewrite three things, every time.
Putting the child’s actual words back in
The AI wrote “explored shadows during outdoor play.” What actually happened: Kai chased his shadow for twelve minutes and announced “the sun pulls it long.” His words are the whole story. I put them back verbatim. A child’s real theory, in their real phrasing, is the single thing no model can generate, because it wasn’t in the room. I was.
Cutting policy-speak back to plain educator language
The draft said Kai “demonstrated sustained engagement and emergent scientific inquiry dispositions.” I write: “Kai stuck with one idea for twelve minutes and tested it. He’s starting to think like a scientist.” Same meaning. A parent can read the second one.
Making the analysis specific to THIS moment
Generic analysis fits any day. Specific analysis names what I noticed: that Kai didn’t just play with his shadow, he formed a hypothesis and tested it by stepping on it. That’s the line that makes the story his.
Checking the EYLF outcome link is real, not pasted
For this moment the honest links were Outcome 4: Children are confident and involved learners, because Kai persisted and theorised, and Outcome 5: Children are effective communicators, because he narrated his thinking out loud. Both sit on something an assessor can actually find in the narrative. So I match every claimed outcome to a specific line. No line, no outcome. I delete it, even the one the model was most confident about. An over-claimed link is worse than a missing one under the EYLF V2.0 framework: it tells a family a story the moment didn’t earn.
My before and after: one real story, draft vs final
Here’s the shadow story both ways, so you can see exactly which lines I had to own.
The AI draft (generic bits marked):
Title: A Fun Afternoon Outdoors Narrative: Kai explored shadows during outdoor play and engaged in active movement around the yard. [generic — fits any child, any day] Analysis: Kai is a confident and capable learner who demonstrated sustained engagement and emergent scientific inquiry dispositions. [policy-speak, no specific evidence] EYLF link: Outcome 3 (wellbeing) and Outcome 4. [Outcome 3 over-claimed: no wellbeing evidence in the notes] Next steps: Provide further opportunities for outdoor exploration.
Read it back. Nothing in there happened. It’s a competent piece of writing about no child in particular.
My final (the lines I rewrote in bold):
Title: Kai and the Shadow That Stretched Narrative: At about 4:40pm Kai noticed his shadow had grown long in the late sun. He chased it across the yard for twelve minutes, tried to step on it, and laughed each time it moved with him. Then he told me his own theory: “the sun pulls it long.” Analysis: Kai didn’t just play — he built a hypothesis and tested it. He stuck with one idea for a full twelve minutes. He’s starting to think like a scientist, and he narrated every step of it out loud. EYLF link: Outcome 4 (confident and involved learners), because he persisted and theorised. Outcome 5 (effective communicators), because he put his thinking into words. Next steps: Set up a torch-and-objects shadow table indoors so Kai can test whether he can make a shadow “pull long” on his own.
The format is identical between the two versions. Every word that carries the actual child, I put back by hand. The model builds the shelf. What goes on it is mine.
My honest opinion: AI is a drafting tool, never the documenter
I’ll take a stance some educators won’t like. AI should never be the one who “knows” the child. And if your workflow lets it write the analysis and the outcome link unedited? That’s exactly what you’ve handed it. The framework is built on the educator’s professional judgement of a specific child in a specific moment. A model wasn’t in the yard at 4:40pm. It didn’t hear the theory. It cannot do the one thing the documentation exists to do: hold this child’s learning, not a child’s. So I let it draft the scaffolding all day. I never let it own the meaning. The twelve minutes are mine to write.
If you want the cross-niche version of this same instinct, the four things I check before any AI-drafted clinical note goes in the record makes the same argument for health records. And for where the ideas come from in the first place, how I picked one curriculum idea from five the AI gave me covers how I treat AI suggestions as raw material, not gospel. I’ve written more on writing child observations with AI without losing your voice too.
TL;DR / Key takeaways
- Keep the format, rewrite the voice. Five fixed blocks: title, narrative, analysis, EYLF V2.0 outcome link, next steps. The model fills them; you own the meaning.
- Strip identifying details first. No names, no faces, no family detail before anything touches a hosted AI tool. Check your service’s privacy notice and ask your director.
- The child’s real words are the story. “The sun pulls it long” beat “explored shadows during outdoor play.” Twelve minutes of chasing a shadow beat “engaged in play.”
- Cut decoration outcomes. Match every EYLF link to evidence in the narrative; delete any the model over-claimed.
- Keep it ~150–250 words. Parents read it on their phone, not as an essay.
Sources
- ACECQA — Belonging, Being and Becoming: The Early Years Learning Framework for Australia V2.0 (2022)
- ACECQA — Approved learning frameworks under the National Quality Framework
Written by Megan, ACECQA-registered Early Childhood Teacher (Sydney). Last reviewed 4 June 2026. This is documentation practice from my own rooms, not advice — your service’s privacy and documentation requirements may differ, so check with your director or nominated supervisor.
Fact-checked 2026-06-04. Last reviewed 2026-06-04.