Not everything belongs on the internet. Today you build a dashboard that lives on your machine, wire yesterday's loop to keep it fresh — and learn the pattern behind our lab's tools: local software that borrows your agent's intelligence through the subscription you already pay for.
Build a personal dashboard from your own data, connect it to your Day-7 loop so it updates itself, and start your capstone's first working version tonight.
You need: your laptop with both agents · any real data of yours (a CSV export, a notes folder — messy is fine) · your spec.md.
A dashboard is a page that answers your recurring questions at a glance: how much, how many, what's overdue, what's next. Companies pay real money for these. You'll commission yours in about twenty minutes — and because it never leaves your laptop, private data stays private.
Point Codex at your data — let it look first
New folder my-dashboard, copy your data file(s) in, and start with reconnaissance, not building:
Look at the data file(s) in this folder and tell me: what's in them, what's messy or inconsistent, and what three questions this data could answer at a glance. Don't build anything yet.Step done when: you've picked the 3 questions your dashboard will answer (yours, not necessarily its suggestions).
Commission the dashboard
Build dashboard.html that reads my data file and answers these three questions at a glance:
1. [your question]
2. [your question]
3. [your question]
Big readable numbers on top, one simple chart, and a table below. Plain HTML/CSS/JS, everything local — the page must work by double-clicking the file, no internet needed.
Plan in 4 bullets first, wait for my OK.Verify against your three questions — can you answer each in two seconds of looking? Tune what's slow to read (“make the overdue number red and twice as big”).
✦ Software with exactly one user is a luxury only agent-builders can afford. You just afforded it.
Step done when: all three questions answered at a glance, on your real data.
Fill it with synthetic data — see the dashboard under load
Ten real rows can't show you what December will look like. Ask for realistic fake data shaped exactly like yours:
Generate 200 realistic synthetic rows in my data file's exact format: same columns, believable values, spread over the last 12 months. Include realistic mess on purpose: two empty months, one huge outlier, a few near-duplicates.
Save it as a separate demo file and add a small switch to the dashboard: REAL / DEMO data.Flip to DEMO and look hard: do the charts survive an empty month? Does the outlier crush the scale? Do 200 rows make the table unusable? Every weakness you find now is a bug real life won't get to show you later.
Step done when: DEMO mode shows 200 rows and you found at least one thing to fix at scale.
Let your loop feed the dashboard
A dashboard you must update by hand dies in a week. But you own a scheduler now. In Hermes:
New scheduled task "dashboard refresh", daily at 07:00: read [where your new data appears — a folder, an export file], update the data file inside my my-dashboard folder to include anything new, and send me a one-line Telegram report of what changed.
Guardrails: you may modify ONLY the data file inside my-dashboard; if the source is missing, report and change nothing. Dry-run now.Step done when: the dry run updates the data correctly and your dashboard shows the change after a refresh.
The Lens pattern — intelligence on subscription
Open our lab's Prompt Lens experiment. Under the hood it's exactly what you've built this week, one step further: a local web tool that hands work to an agent through your subscription login — no separate paid API key. Upload an image, and the local app asks the agent to describe it as a reusable prompt.
The pattern, in course vocabulary: local tool for the interface · agent for the thinking · subscription for the fuel. Watch it in action once:
Add a small "Insight" box to my dashboard: a button that, when I click it, shows three observations about the current data — trends, oddities, one suggestion.
Generate the observations yourself right now from the data file and store them in the page (we'll regenerate them on schedule later — for now the button just reveals them).Then close the circle with one line to Hermes: “Extend the dashboard-refresh task: after updating the data, also rewrite the three observations in dashboard.html.” Now the intelligence refreshes on schedule too — an agent thinking about your data every morning, billed to the subscription you already pay.
✦ Interface local, thinking agentic, fuel by subscription — you now recognize this pattern everywhere, including in the tools of this course.
Same pattern, other lives: a price list whose product descriptions rewrite themselves overnight; a folder where scanned documents name and file themselves; on Day 10 — a bot that drafts answers to customers. Always the same three roles: local interface, agent thinking, subscription fuel.
Step done when: the Insight box shows observations that are actually about YOUR data — as concrete as the example above.
Your tool stops being a spec: build the first working version, locally, ugly, tonight.
homework complete — mark the day done below
Tick when the homework above is finished — it syncs to your account.
When its only user is you — especially with private data. Local means no hosting, no accounts, no leaks: the file opens in your browser and the data never leaves the laptop. Publish when someone ELSE needs to open it; that's tomorrow.
Local tool = interface · agent = thinking · subscription = fuel. The tool collects input and shows output; the agent does the intelligent part; and it all runs on the plan you already pay for, not on separate API bills.
Because it runs unattended, and unattended writes need the smallest possible territory. If the loop misbehaves, the damage is one regenerable file — not your dashboard, not your project, not your Documents. Fences are sized to the worst case, not the best.