# Case Study Template

Use this to document a build so you can see whether it actually earned its place — and so someone else can rebuild it in ten minutes.

Fill it in a week *after* the build goes live, not the day you make it. The first day is always optimistic.

---

## The template

```markdown
# [Build name]

**Built:** [date]  ·  **Reviewed:** [date, one week later]
**Bots involved:** [names and roles]

## The job it took over
[What a human used to do, in one paragraph. Be specific about who did it and when.]

## Role and description
- **Role:** [the one thing this bot owns]
- **Its description (the job ad):** [paste the one-liner]

## Connections used
- [ ] Gmail / mail
- [ ] Calendar
- [ ] Project board — [which]
- [ ] Slack
- [ ] Zapier MCP — [which apps]
- [ ] Firecrawl
- [ ] Browser login via take-over — [which site]

## Routine settings
- **Type:** schedule / trigger
- **When:** [e.g. weekdays 7:00am, or: Slack message in #inbound]
- **Instruction:** [paste it verbatim]

## Before → after

| | Before | After |
|---|---|---|
| Who did it | | |
| How long it took | | |
| How often | | |
| What usually went wrong | | |

**Time saved per week:** [hours]
**What I still do by hand:** [be honest — this is the useful part]

## What broke in week one
[Every build gets something wrong at first. Write it down; it's the fix list.]

## Would I rebuild it?
[Yes / no / yes but differently — and why]
```

---

## Three worked examples

These are written from the five builds in the course, as **example builds** — a template of what a filled-in write-up looks like, not reports from named people.

---

### Example 1 — Agency reporting bot

**Bots involved:** Reporting bot (solo, hands off to the chief of staff for delivery)

**The job it took over.** Every Friday afternoon an account manager opened four dashboards and the project board, copied numbers into a template per client, chased two people for blocker updates, and sent the reports out. It reliably ate the last three hours of the week, and it was the first thing to slip when the week got busy.

**Role and description.** Owns weekly client reporting.
*Job ad:* "Builds each client's weekly report from the board and connected dashboards every Friday, and tells me what's missing."

**Connections.** Project board, two analytics dashboards, Gmail (drafts only).

**Routine.** Schedule — Fridays 4:00pm.
*Instruction:* "Build this week's client report: what was delivered, what moved, what's blocked, and what's next week. Pull numbers from the connected dashboards rather than guessing. Put the draft on the board and tell me what's missing."

**Before → after**

| | Before | After |
|---|---|---|
| Who did it | Account manager | Reporting bot, reviewed by the AM |
| How long | ~3 hours | ~25 min of review |
| How often | Weekly, sometimes skipped | Weekly, never skipped |
| What went wrong | Numbers copied wrong; blockers stale | Occasionally an unavailable metric |

**Time saved per week:** ~2.5 hours.
**Still by hand:** the judgement calls — which blocker to escalate, and how to word bad news to a nervous client.

**What broke in week one.** It reported "on track" for a project with one slipped item because the board card hadn't been moved. Fix: the instruction now says never write "on track" unless every item is on track, and it names any card that hasn't moved in seven days.

**Would I rebuild it?** Yes. This is the highest-leverage single build for an agency, because reporting is both mandatory and universally hated.

---

### Example 2 — Freelancer billing flow

**Bots involved:** Inbox bot + Invoice bot, working in a group chat

**The job it took over.** Invoicing happened in a Sunday-night batch — scrolling back through email for anything owed, building each invoice by hand from the last one, writing the covering email, and updating a spreadsheet. Things got missed, and chasing was inconsistent.

**Role and description.**
*Inbox bot job ad:* "Triages my inbox each morning into tasks, replies-needed, and noise."
*Invoice bot job ad:* "Builds invoices in my house style from email requests, drafts the reply, and logs status — drafts only."

**Connections.** Gmail, project board. One example invoice uploaded as a PDF.

**Routine.** Schedule — daily 9:00am.
*Instruction:* "Check my email daily, find anything invoice related, generate the invoice in my style, draft the reply email with it attached, and log the status on my board."

**The bit worth noticing.** The first time it ran, the invoice bot messaged the inbox bot on its own: *please hand me anything invoice or payment related from your triage.* Nobody set that handoff up — the bots arranged it between themselves after reading each other's descriptions.

**Before → after**

| | Before | After |
|---|---|---|
| Who did it | Me, Sunday night | Two bots, daily |
| How long | ~90 min weekly, in one dreaded block | ~10 min of clicking send |
| How often | Weekly, batched | Same day as the request |
| What went wrong | Missed invoices; inconsistent chasing | Nothing yet; drafts occasionally need a line reworded |

**Time saved per week:** ~1.5 hours — but the real win is that invoices go out the day they're requested instead of up to six days later.

**What broke in week one.** It drafted an invoice from a thread that was actually a quote request. Fix: the instruction now distinguishes "asking what it costs" from "asking to be billed", and flags ambiguous ones instead of drafting.

**Would I rebuild it?** Yes, and it'd be the first thing I'd build next time. Send stays manual on purpose — nothing goes out without a click.

---

### Example 3 — Creator content engine

**Bots involved:** Research bot → Content bot

**The job it took over.** Finding topics was a weekly hour of scrolling competitor channels, guessing what was working, and starting drafts from a blank page — with a permanent low-grade fear of running dry.

**Role and description.**
*Research bot job ad:* "Scans competitors daily on the public web and hands the five strongest topic ideas to the content bot."
*Content bot job ad:* "Owns my content — learns my titles and style and drafts from the research bot's ideas in my voice."

**Connections.** Firecrawl plugin, project board. Content bot pointed at the existing channel to learn what already works.

**Routine.** Schedule — daily 8:00am (research). Content bot has no schedule; it's woken by the research bot's message.

**Before → after**

| | Before | After |
|---|---|---|
| Who did it | Me, when I remembered | Research bot, daily |
| How long | ~1 hour weekly, plus blank-page time | ~15 min picking from five |
| How often | Sporadic | Every morning |
| What went wrong | Chasing big numbers, not momentum | Occasional repeat of a covered topic |

**Time saved per week:** ~1 hour, and the idea drought is gone — you wake up to ideas with drafts started.

**What broke in week one.** It kept surfacing huge videos from huge channels as "trending." Fix: the scan skill now ranks on momentum *relative to that publisher's own baseline*, not raw view count.

**Would I rebuild it?** Yes. The handoff is what makes it work — research finding topics is useful, research *waking the drafter* is a system.

---

*Part of the Grok Bot AI Full Course — <https://agentos.guide/grok-bot-course>*