Family admin · local AI · guide 9823

Build a private family handoff calendar from school notices.

Turn pasted school notices and co-parent messages into a conflict-checked pickup, deadline, and handoff calendar on your own computer with Ollama — $0 software and no cloud upload; the trade is local setup and manual verification of every date.

✓ Verified 2026-08-20 T3 · self-hosted No card 4.7 GB model as of 2026-08-20 32K context as of 2026-08-20

See the setup steps

Live workflow sandboxeditable

Make the prompt fit your household.

Edit the real inputs below. The working prompt updates before you paste it into a local Ollama chat.

Source notices and messages Household rules and constraints

Output shape

Unknown-date rule

$0software + local model

4.7 GBQwen2.5 7B download as of 2026-08-20

32Kselected model context as of 2026-08-20

0hosted request quota

Replaced costs

The recurring bill is optional. The review is not.

Perishable prices are dated; the local workflow is the durable part.

Cost comparison

A hosted assistant can format the same text, but it adds a subscription and sends the working copy to a provider.

Hosted baseline · ChatGPT Plus $240 / year

Ollama + Qwen2.5 7B local software $0 recurring

Trade: the local option uses 4.7 GB of disk for the selected model and depends on your computer's speed; electricity and hardware are not counted.

Hosted price as of 2026-08-20OpenAI pricing ↗

What the local boundary changes

When you run the local qwen2.5:7b model, the prompt is processed by the model on your machine. Do not select an Ollama :cloud tag when the goal is local-only handling.

That boundary removes the hosted request meter, but it does not make model output reliable by itself. The prompt below forces dates, owners, source quotes, collisions, and unknowns into separate fields so you can check the result against the original notice.

Local installOllama quickstart ↗

Quota calculator · durable capacity check

Find out whether one bundle fits before you run it.

Ollama has no hosted daily request quota for local runs. This uses the selected model's published 32K context and 4.7 GB size.

Your source bundle

Use rough estimates. If the context result turns amber or coral, split the notices by week or source type before asking for a calendar.

Number of notices School emails, flyers, messages, or copied sections.

Average tokens per notice A page of plain text is often hundreds of tokens.

Prompt and rules tokens Instructions, schema, and household constraints.

Free disk space, GB Available before downloading the model.

Model copies you plan to keep Each selected Qwen2.5 7B copy is logged at 4.7 GB as of 2026-08-20.

Input load2,900tokens in this bundle

Context used9.1%of 32,000 tokens as of 2026-08-20

Context remaining29,100tokens before splitting

Storage after install5.3 GBfree disk after 1 copy

Fits one conflict-first pass. Keep exact dates and source labels in the input.

Model facts verified 2026-08-20Qwen2.5 official model page ↗

Contrarian correction · rarity device

Most guides tell you to ask for a clean family calendar first. That is backwards.

For school notices, the costly failure is a polished schedule with a hidden collision. Ask the local model to find conflicts, missing owners, and uncertain dates before it formats a weekly view. Formatting first makes a wrong answer look finished.

Do not start with: “Make me a nice family calendar.” That request rewards neatness even when the source has two different pickup times.
Start with: “List every event, quote its source, find overlaps, and put unknowns in NEEDS CONFIRMATION.” That gives you a short review queue before the calendar.
Consequence: a five-minute source check can catch a duplicated pickup or an invented date before it reaches a shared calendar.

Step-by-step walkthrough

Run the workflow as an audit, then format the answer.

Use the tabs in order; each one changes what the next prompt can safely assume.

Install the free local runner and the text model.

Ollama runs on macOS, Windows, and Linux. The selected Qwen2.5 7B entry is listed at 4.7 GB with a 32K context window. Downloading the model uses disk space once; local inference does not consume a provider request quota.

Install Ollama from its official download page, then open a terminal.
Run ollama run qwen2.5:7b; do not use a :cloud tag for local-only handling.
Keep at least 4.7 GB of free disk plus room for the operating system and other models.

terminal

# After installing Ollama
ollama run qwen2.5:7b

# Optional local API check
curl http://localhost:11434/api/generate -d '{
  "model": "qwen2.5:7b",
  "prompt": "Reply with LOCAL OK",
  "stream": false
}'

Make the source text checkable before asking for a calendar.

Paste the text of one planning window, usually one week. Keep the date, time, location, and source label together. Replace real names with stable labels if other people will see the prompt or saved transcript.

Label each block SCHOOL, PARENT A, PARENT B, or ACTIVITY.
Write the timezone and the rule for ambiguous phrases such as “next Friday.”
Redact addresses, student IDs, medical details, and phone numbers that the schedule does not need.

source block

[SOURCE: SCHOOL NEWSLETTER | received 2026-09-08]
Friday, September 12: early dismissal at 12:30 PM.
Saturday, September 13: fundraiser, 10:00 AM–1:00 PM.

[SOURCE: PARENT A MESSAGE | received 2026-09-08]
Parent A can do Friday pickup. Parent A cannot do Saturday morning.

[HOUSEHOLD RULES]
Timezone: Central Time. Use Parent A and Parent B labels.
Never infer a date or travel time.

Run the conflict-first prompt, not a calendar prompt.

The output order is intentional: source ledger, conflicts, unknowns, then a proposed handoff table. A model can make a wrong schedule look authoritative when the audit and the formatting happen in one vague request.

Copy the complete template below or use the live sandbox above.
Require a source quote for every date and owner; missing evidence goes into NEEDS CONFIRMATION.
If the calculator crosses 80% of 32K context, split the bundle before running it.

conflict-first prompt

You are a family schedule auditor, not an authority.
Use only the source blocks below.

Rules:
1. Preserve exact dates, times, locations, and timezone.
2. Quote the source block for every extracted event.
3. Never infer a date, owner, travel time, or agreement.
4. Find overlaps, impossible handoffs, missing owners, and contradictions first.
5. Put uncertain fields in NEEDS CONFIRMATION.
6. Only after the audit, draft a proposed weekly handoff table.

Return these headings in order:
A. SOURCE LEDGER
B. CONFLICTS AND OVERLAPS
C. NEEDS CONFIRMATION
D. PROPOSED HANDOFF TABLE
E. CHECKS I MUST PERFORM

SOURCE BLOCKS:
[paste one planning window here]

Verify each event before it becomes a shared plan.

The model's output is a draft, not a calendar authority. Compare each row with the source block, resolve the confirmation list with the people involved, and only then copy the final rows into the calendar your household already uses.

Check every date, time, location, and owner against the original notice.
Ask a person to resolve every item in NEEDS CONFIRMATION; do not let the model fill the gaps.
Share the final calendar, not the raw family messages or the unreviewed model output.

verification pass

Verification checklist
[ ] Every event has a source quote.
[ ] Every date includes a timezone or is marked UNKNOWN.
[ ] Every owner is confirmed by a person.
[ ] Every overlap has a proposed resolution or remains open.
[ ] I checked the final rows against the original notices.
[ ] I copied only the reviewed rows into the shared calendar.

01 · SOURCE

Keep the quote beside the event.

Without a source quote, a wrong date can survive a quick scan because the output looks more certain than the notice.

02 · CONFLICT

Ask for collisions before formatting.

A conflict-first result gives you a finite review queue; a calendar-first result hides contradictions inside a neat grid.

03 · BOUNDARY

Keep local and cloud tags distinct.

qwen2.5:7b is a local model tag; a :cloud tag changes where the prompt is processed.

Try it now

Paste one week. Find the collisions.

Start with the live sandbox at the top of this page. Copy the generated prompt, run it against your local model, and compare every proposed row with the original notice before sharing it.

Open the sandbox ↑

Research notes

What is durable, and what needs rechecking.

Every perishable number on this page carries its verification date.

Official sources

Ollama quickstartinstall + platforms
Qwen2.5 model library4.7 GB · 32K · dated
Ollama API: generatelocal endpoint
ChatGPT pricing$20/mo Plus · dated

Durable framework

1. Preserve evidence. Put the source beside every event.

2. Audit before formatting. Find collisions, unknowns, and owners before making the result attractive.

3. Keep a human checkpoint. The model can draft a plan; a person must confirm the plan.

4. Treat facts as dated. The model size, context window, software behavior, and hosted prices can change; recheck the chips marked “as of 2026-08-20.”