AI Camp: Program Design Across Four Age Bands

Prepared: 15 October 2026 Builds on: AI Camp Starter Courses (Original) Method: Course design grounded in published AI-literacy frameworks, platform age policies, child-safety assessments and learning research, checked on the web on 15 October 2026. Sources are listed at the bottom. Platform age rules change often. Re-check them before each camp season.

1. The idea in one paragraph

Every camper leaves with something that works, built for a part of their own life they chose. The original three courses stay intact: AI Mastery for Living Life (personal, social and work/school agents), AI Mastery for the Work/School World (Tech, Arts and Biz), and AI Mastery for Personal & Social Life. What changes by age is the tooling, how much autonomy the agent gets, and how deep the "how does it actually work" goes. The through-line at every age is the same: you are the boss of your agent, and your agent should push you toward real people and real work, not away from them.

2. The four camps at a glance

Default format for every course: five sessions of three hours (15 contact hours), run as a one-week day camp. Each syllabus also runs as ten 90-minute weekly sessions by splitting each day at its midpoint. Course 1 is the survey course and the right first purchase. Courses 2 and 3 are deep dives that assume Course 1 or equivalent. Course 4, new for ages 12 and up, is the systems course underneath the other three. It works best after Course 1, and pairs naturally with whichever course a camper takes next.

3. Why these frameworks, and how we use them

We did not invent a scope and sequence from scratch. Three public frameworks already map what young people should know about AI, and every syllabus cites the one it draws on.

• UNESCO's AI Competency Framework for Students sets 12 competencies across four dimensions (a human-centred mindset, ethics of AI, AI techniques and applications, and AI system design) and three progression levels: Understand, Apply, Create. ^[1] We use the progression levels to set each band's ceiling: Explorers mostly understand, Builders apply, Makers and Architects create.

• AI4K12's Five Big Ideas (perception, representation and reasoning, learning, natural interaction, societal impact) organise what's under the hood, by grade band. ^[2] Every course touches at least three of the five.

• The OECD and European Commission AI Literacy Framework for primary and secondary education was finalised on 18 June 2026 and feeds the PISA 2029 Media & AI Literacy assessment. ^[3] It is the most likely yardstick parents and schools will hear about next, so the syllabi use its language where it fits.

The US policy context points the same way: Executive Order 14277 (April 2025) set up a White House task force and a Presidential AI Challenge to promote AI education for K–12 students. ^[4]

4. Design principles that hold at every age

1. Projects, peers, passion, play. Campers build things they care about, with others, by tinkering. This is Mitchel Resnick's "Four P's of Creative Learning," the design basis for Scratch. ^[5] It also matches what self-determination theory says sustains motivation: autonomy, competence and relatedness. ^[6] The original course copy already asks kids to pick what matters to them. We keep that choice at the centre.

2. Coach, not crutch. In a field experiment with nearly a thousand high-school math students, unrestricted GPT-4 access improved practice scores, but students did 17% worse once access was removed. A tutor version that gave hints instead of answers largely removed the harm. ^[7] A carefully designed AI tutor at Harvard, by contrast, produced more than double the learning gains of an active-learning class. ^[8] The design lesson is that the same technology helps or hurts depending on whether it does the thinking for you. Every agent campers build has to make them do something, not just do it for them.

3. Brain first, then the bot. An early, not-yet-peer-reviewed MIT Media Lab study found lower brain engagement and weaker recall among adults who wrote essays with ChatGPT than among those who wrote unaided. ^[9] It's preliminary, but it supports a simple camp rhythm: every build starts with a few minutes of unplugged thinking on paper before anyone opens a tool.

4. No companion bots. Common Sense Media rates social AI companions an "unacceptable" risk and recommends nobody under 18 use them. ^[10] In its 2025 survey of teens, it found many already do. ^[11] So the Personal & Social courses never build an AI friend. They build agents that point you toward real people.

5. Real health guidance, not vibes. When a camper builds a fitness or food agent, it is anchored to a public standard. For movement, that's the CDC guideline of 60 minutes or more of moderate-to-vigorous activity a day for ages 6–17. ^[12]

6. The Agent Card. Every age band uses the same five-part scaffold to describe an agent. It scales from crayons at age 8 to code at 17:

   • Goal: what it's for, in one sentence.

   • Tools: what it can look at or do.

   • Memory: what it remembers, and what it must never keep.

   • Rules: what it must never do.

   • Check-in: when it has to stop and ask a human.

   The vocabulary follows Anthropic's distinction between workflows (steps fixed in advance) and agents (systems that choose their own next step and tool). ^[13] Younger campers build workflows and call them agents. Older campers learn the difference, and when not to use an agent.

5. Harnesses and second brains (ages 12 and up)

The three original courses teach campers to build agents. Course 4 teaches the part that decides how good those agents actually are, and it's where AI practice has moved fastest since 2025.

The harness. A language model remembers nothing between conversations. Everything it knows in the moment sits in a limited context window, and the software around the model, the harness, decides what goes in: persona and instruction files, memory, skills, tools, and summaries of earlier work. Andrej Karpathy calls the craft "the delicate art and science of filling the context window with just the right information for the next step." ^[14] Anthropic's engineers describe context as "a finite resource with diminishing marginal returns." ^[15] Independent tests show why: across 18 models, reliability fell as inputs grew, ^[16] and models make better use of information at the start and end of a long input than information buried in the middle. ^[17] So we teach harness setup and tuning: choosing, trimming and ordering context, and measuring the difference.

The second brain. The other half is a personal knowledge vault: plain-text Markdown notes the camper writes, links and owns. The idea runs from Vannevar Bush's 1945 memex, "an enlarged intimate supplement to his memory," ^[18] to Tiago Forte's CODE and PARA methods, ^[19] to Obsidian CEO Steph Ango's "file over app": "files you can control, in formats that are easy to retrieve and read." ^[20] A second brain is good for a learner on its own terms, and it is also the best context an agent can be given. Brain first still applies: campers write their own notes, and agents only read them.

The tools, and their limits.

• Hermes Agent (Nous Research) is open source under the MIT license, works with any model provider, and exposes the harness as editable files: SOUL.md (persona), AGENTS.md (workspace rules), MEMORY.md and USER.md (memory), plus skills that follow the open Agent Skills standard. ^[21] Its memory is deliberately small, about 800 tokens for the agent's own notes and 500 for its profile of the user, which makes "what deserves to be remembered?" a concrete exercise. ^[22] Hermes is self-hosted, so it runs on Colearn-provisioned machines and capped model keys, with command approval on and its messaging connections off.

• Obsidian is the best-known second-brain app, but its terms of service limit use to people who have reached the age of majority. ^[23] Campers under 18 build in Logseq, which is open source (AGPL) and local-first, ^[24] or in any Markdown editor. Instructors demo in Obsidian. Because every vault is plain Markdown, a camper can move theirs into Obsidian when they're old enough.

6. Tools by age: what campers may touch directly

Platform age rules decide a lot of the design. As of October 2026:

• Chat GPT is not meant for children under 13, and 13- to 17-year-olds need parental consent. ^[25] OpenAI added linked parent accounts and an under-18 experience in September 2025. ^[26]

• Claude.ai requires users to be 18 or older. ^[27] Developers who build on Anthropic's API for minors take on extra duties: age verification, content moderation, monitoring and child-privacy compliance. ^[28]

• Gemini Apps can be switched on for under-13s only by a parent, through Google Family Link. ^[29]

• COPPA: if Colearn collects personal information from campers under 13, the FTC's amended rule took effect on 23 June 2025, with full compliance due by 22 April 2026. ^[30]

✅ hands-on · 👀 watch and steer · ◻️ optional · ❌ not used

Mixed-age cohorts follow the youngest camper's column. A 12-year-old in the Builders camp never gets a personal chatbot login.

7. Shared assessment: the Four Lenses

There are no grades for prompts. Every capstone is reviewed through four lenses that map onto UNESCO's four dimensions. ^[1] Each lens is scored as Getting There, Got It or Wow.

Every course ends with a Showcase: a 3-minute demo for families, where the camper (not the agent) explains the Agent Card.

8. Family partnership

• Before camp: families get a one-page tool list for their camper's band, plus consent forms where platforms require them.

• During camp: a daily "what we built" note, with one dinner-table question.

• After camp: campers leave with their Agent Cards and a suggested AAP Family Media Plan session, so new tools land inside rules the family agrees on together. ^[34]

9. Free curriculum we lean on (and credit)

Ready-made lessons save prep time and are well tested: MIT RAISE / Day of AI (free, grade-by-grade K–12 lessons), ^[35] Experience AI from the Raspberry Pi Foundation and Google DeepMind (ages 11–14), ^[36] and the CSTA AI Learning Priorities, whose six-step AI project cycle (question → data → evaluate data → train → evaluate model → improve) is the backbone of every train-a-model lab. ^[37]

10. Open decisions for Colearn

• Delivery: online live, in person, or hybrid? Every syllabus assumes live instruction with a ratio of at most 1:8 for 8–11 and 1:12 for older bands.

• Accounts: whether Colearn provisions supervised accounts, which decides the 12–14 and 15+ tool columns.

• Data: any camper data that leaves the room (photos for a Teachable Machine model, chatbot logs) needs a written retention rule before launch.

• Obsidian for minors: its terms limit use to adults. Confirm with Obsidian whether supervised educational use is allowed before naming it in camp marketing. Until then, Logseq is the camper default.

• Hermes hosting: decide where camper sandboxes run, who holds the model keys, and whether memories, skills and vaults are wiped or exported to families at the end of camp.

• Grading.md: the existing Grading note in the Colearn folder is empty. The Four Lenses above could fill it.

Sources