Claude Code for Teachers (September 2026): How the Usage Pool Really Works, the 18+ Wall, and One Workflow an IT Coordinator Can Run

A week before term starts, a colleague who teaches Form 5 ICT and doubles as IT coordinator turns his laptop round in the staff room. "The academic committee just handed me the timetable export. Twelve hundred rows, sixty columns, and I need sixty individual staff timetables out of it. Last year that took me three evenings." Then he adds the part that has actually stopped him: "People say Claude Code would do this, but I don't know if it'll burn through some quota I can't see, and I don't know whether my students are allowed to touch it."

Both questions have firm answers, and neither depends on how good the tool is: the allowance is a rolling window shared with your ordinary web chats, and the student side is closed entirely. The full set of nine tools is in the nine AI tools Hong Kong teachers keep asking about.

50|## What agentic means

An ordinary assistant answers a question and leaves you to copy the result somewhere. Claude Code is a different shape: it reads the whole project directory, edits across files, runs commands to see what happened, corrects itself from the result, and works directly with git — staging changes, writing commit messages, opening branches and pull requests. The difference is not that it is cleverer. It is that it acts, so it can also break things, and that governs how you should use it.

It runs on five surfaces — terminal CLI, VS Code extension, JetBrains plugin, desktop app and web — and a session moves between them. A CLAUDE.md file in the project root is its long-term memory, loaded automatically at the start of every session. claude -p accepts piped input, so a log file or a git diff goes straight in. Skills package repeated workflows, and hooks run your own commands before or after agent actions. Parallel subagents split a large task and merge the results. Scheduled tasks run in the cloud with your machine off, or locally with access to your own files. MCP connectors attach Google Drive, Jira, Slack or something your school built.

One thing to state without hedging: this is a developer tool, not a classroom product. No classes, no pupils, no curriculum alignment, and Anthropic does not pretend otherwise. What it built for teachers is a separate product, Claude for Teachers, positioned as a lesson-preparation assistant and available only in the United States.

How the usage pool works

60|This section matters more than the feature list, because it decides whether a colleague suddenly hits a wall on a Wednesday afternoon.

Every plan resets usage on a rolling five-hour window, and paid plans add weekly limits on top. The important part: web, desktop, mobile and Claude Code all draw from one pool. Concretely, running a job in the terminal that splits a 1,200-row timetable can leave you unable to use the web app to draft a parent letter the same afternoon. There is no published message count, and how far the allowance goes depends on conversation length, model choice and which features you enable, so it cannot be estimated in advance.

Three commands are worth building into habit. /compact condenses the current conversation into a summary and continues, which pays off most in the middle of a long job. /clear empties the session once a task is done, so old context stops being re-billed. /cost shows what the current session has actually spent, though only meaningfully if you are on an API key.

When the allowance runs out you can enable usage credits and continue at standard API rates, but that converts a fixed cost into a variable one, and a school should budget the two separately.

The usage pool is designed around one person, not around a school supporting a hundred colleagues at once.

70|## One real workflow

Back to that timetable. It is a good first job precisely because it is repetitive, mechanical, unavoidable every August, and contains no student personal data — only staff names, classes, rooms and periods.

Rules before tasks

Install first. On Windows PowerShell:

irm https://claude.ai/install.ps1 | iex

80|On macOS, Linux or WSL use `curl -fsSL https://claude.ai/install.sh | bash`; Windows also supports `winget install Anthropic.ClaudeCode`. Then start it in the project:

cd timetable-split
claude

The first run asks you to sign in. Next, put the panel's conventions in a file instead of repeating them every session:

This project splits the academic committee's timetable export into individual staff timetables.

    90|Rules:
- Python standard library only; school machines have no extra packages.
- data/master.csv is read-only and must never be modified or overwritten.
- Write output to out/, one HTML file per teacher, named by staff number rather than name.
- All output text, error messages and comments in Traditional Chinese (Hong Kong usage).
- Where the source is incomplete — missing room, two classes in the same slot — do not guess. Write the row to out/issues.csv for a human to resolve.

That last rule is the most important line in the file. The machine moves data; a person resolves contradictions. Writing "when unsure, report it" into the rules works far better than checking afterwards.

Then the task

100|```text Read data/master.csv (1,200 rows, 60 columns, first row is headers) and do the following:

1. First just print the column names and the first three rows so I can confirm the mapping. Do not write code yet.
2. Once confirmed, create split_timetable.py that groups by teacher and writes one HTML timetable each: Monday to Friday across, periods 1 to 8 down.
3. Where two records share a slot, or the room column is blank, do not merge or guess. Copy the row into out/issues.csv and mark that cell "to confirm".
4. Finally print to the terminal: files created, row count in issues.csv, and the three teachers with the most and fewest periods.

Step 1 deliberately holds it back. A 60-column export usually has two or three headers that mean nearly the same thing — `Room` and `Venue`, say — and having it report before it builds costs less than letting it guess and then unpicking the result.

   110|If the run fails, do not retype the error:

```bash
python split_timetable.py 2>&1 | claude -p "This script is failing. Identify the cause and propose the smallest fix. Do not rewrite the whole file."

Once it is stable, /schedule turns it into a routine that runs each August, and those few lines in out/issues.csv become the list you take to the academic committee. The first time through, three evenings become a break plus one free period — but somebody still has to read that issues file, and no improvement in the tool changes that.

The student side

120|This is the part a Hong Kong school needs settled first, and it has nothing to do with price.

Anthropic operates an 18-and-over policy; its own guidance states that users under 18 are prohibited from using Claude, enforced with age assurance services. This is a product policy, not an administrator setting. Even Claude for Teachers, launched on 14 July 2026, states explicitly that there are no student accounts — and that programme is restricted to serving educators at US K-12 schools and districts, verified through a third-party service, users aged 18 or over, with at least a year free for those signing up by 30 June 2027. Hong Kong teachers do not qualify and must buy Pro, Max or Team at full price.

In higher education Anthropic offers a university-wide plan priced on enquiry, and researchers at academic and nonprofit institutions can start at no cost. There is currently no equivalent offer for Hong Kong primary or secondary schools.

So the division of labour in a school AI rollout is already decided for you: staff and IT can use this; the student side needs a different system with gating, supervision and records. For how that gating is built, see student AI safety: four-layer gating, PII protection and teacher monitoring.

The shape of the bill

130|Figures below are from the official pricing page and help centre as of September 2026, in US dollars, exclusive of tax. Claude Code is included with every paid plan and is not available on Free.

Pro is US$20 a month monthly, or US$17 a month billed annually at US$200 up front. Max 5x is US$100 a month for five times Pro's usage per five-hour session; Max 20x is US$200 a month for twenty times. Team standard seats are US$25 per user monthly or US$20 annually, with a minimum of 2 members and a ceiling of 150 seats; Team Premium seats are US$125 monthly or US$100 annually for 6.25 times Pro's per-session usage. Enterprise is US$20 per seat per month plus usage at API rates, billed annually, with SCIM, audit logs, custom retention and role-based access.

If you go through the API instead of a subscription, per million tokens: Fable 5.1 is US$10 in and US$50 out, Opus 5 is US$5 and US$25, Sonnet 5 is US$2 and US$10, Haiku 4.5 is US$1 and US$5.

The seat multiplier

Putting all 100 staff on Team standard seats at the annual rate is US$2,000 a month, US$24,000 a year, roughly HK$187,000. That money buys a development and writing assistant — no classes, no curriculum alignment, no pupils. The rational move is to buy for the two to five colleagues who will genuinely use it, usually the ICT teachers and the IT coordinator.

140|Enterprise pricing deserves a second look too: the seat fee is only US$20 because the real cost is consumption at API rates. Any three-year financial plan built on "US$20 per person per month" will come in low.

Who to buy it for

Everything above collapses into one procurement decision: this is a tool for two to five people, not for a whole school.

Its home ground is not in dispute. Claude Code writes code, refactors and understands large repositories far better than a school platform does. ICT testing pipelines, an IT coordinator's data conversion and system maintenance, weekly or annual data chores — handing those over saves real hours. Its Chinese writing is among the best in this class of tool, so the reports and comments it produces read well to colleagues. CLAUDE.md and skills let a panel capture conventions in a file rather than passing them on verbally, and deep git integration finally gives internal tools a maintenance history.

What it does not do is equally clear: no student accounts, no teacher monitoring, no curriculum alignment, no Traditional Chinese interface, a usage pool shared with chat, and no embedding or speech models from Anthropic — so no school can run an entire platform on this one provider. Those gaps are what a school-based system covers. Edor.ai is a per-school annual fee for roughly 800 students and 100 staff with AI usage included, student conversations pass through four-layer gating and teacher monitoring, schemes of work and rubrics sit in the knowledge base, and the data belongs to the school with full export on demand. On accountability, the "智" 啟學教 funding wants evidence of impact, and the PIE module produces it from platform data — see how to use the EDB HK$500,000 AI grant. 150| These are not substitutes for each other. A school running both is the normal case.

Common misreadings

• The claim: a paid plan throws Claude Code in as a bonus, so usage does not really matter. What actually happens: it is genuinely included with every paid plan, but its usage shares one rolling five-hour pool with web, desktop and mobile, and paid plans add weekly caps. A big terminal job really does spend the allowance you wanted for that afternoon's parent letter.
• The claim: Anthropic has a free teacher programme, so a Hong Kong school can apply. What actually happens: Claude for Teachers is limited to serving educators at US K-12 schools and districts, verified by a third party, and has no student accounts. Hong Kong teachers pay full price and do not get the K-12 data processing terms either.
• The claim: the agent verifies its own results, so the output is ready to use. 160| What actually happens: it verifies that the program runs, not that the result matches reality. Sixty timetable files appearing in out/ does not mean none of them puts two classes in the same room, and that check is still a person's job.

Three things to try on Monday

1. Check your own machine for an ANTHROPIC_API_KEY environment variable. If it is there, you think you are spending a subscription allowance while you are actually paying per token — the most common wasted money in this tool.
2. Pick one mechanical job your panel repeats every year that contains no student personal data — timetables, duty rosters, equipment inventories, checking old links on the school site — and give it that one job. Do not start with marking.
3. Add a line to the school AI use policy stating that developer tools are for staff only and the student side runs on the school-approved platform. Write it before somebody asks a whole class to install something.