∆ Admin · S09

Wiring the Backend — real data, tools & agents behind your prototype

s09 · Betterworks · 100m · Full Cohort · draft

s09 → 6 blocks, 55 slides, 51 trainer notes, 5 exercises, 1 reflects, 0 quizzes, 0 asides

Pre-read

- Have a **recent Claude chat open** you can run a prompt against — ideally your capstone work
- Bring **one real Betterworks data source** your capstone would need — Aha!, Jira, a sheet, an export
- Optional — try to name, in one sentence each, what a "skill," a "connector," and an "agent" are. Note where you get fuzzy. That fuzziness is today.

No slides to read. Show up with a real chat to point a prompt at, and one real data source in mind.

Session goal

By end-of-session each participant can:

1. **Draw the stack** — Prompt → Skill → Tool → Connector → Agent — and place their own work on it
2. **Run "where am I really"** on any chat — and tell a Skill from a prompt from an Agent by evidence, not vibes
3. **Call a real API** — hit a live public API from Claude and ground an answer in what comes back
4. **Explain a Connector (MCP) and RAG** — why plugging a model into live data beats copy-paste, and what "AI database" means
5. **See a real agent run** — where agents actually run, and turn one thing they do by hand into one

The arc: see where the capstone stands → name the *stack* and the skills you built → run it on your own work → reach *real data* (API → MCP → RAG) → run a real *agent* → keep it *safe*.

B1 00–08

Capstone status quo · and today's map

8 items

We start with where the work actually is — then name the pieces that take it further.

▤ Slide · s9-B1-1 pb-cover B1 L25

Wiring the Backend

Real data, tools & agents — the system behind the screen.

▤ Slide · s9-B1-2 pb-section B1 L32

Capstone — Status Quo

▤ Slide · s9-B1-3 list B1 L42

Where we left off · from S08

Two things from yesterday that carry straight into wiring.

AI made generating cheap — so a clear question or hypothesis is what gives a prototype its purpose. Fidelity matches the question, not your ambition.
Structured prompting (S01) matters more, not less — the lever for scoping, critique, and prioritisation, and the guardrail against hallucination and drifting off-brief.

▤ Slide · s9-B1-4 list B1 L54

Where today goes

We name the pieces, then run them on your own work.

The stack, named — and the skills you've already built
Where am I really — run the stack on one of your own chats
Reaching real data — call a live API, then MCP & RAG
Agents — where they run, and a real one fetching the weather to your phone
… 1 more

B2 08–24

The stack, named · and the skills you've already built

13 items

The words have been flying around without a map. We lay the vocabulary down first — and point at the skills you already shipped.

▤ Slide · s9-B2-1 statement B2 L73

A prompt asks. A system acts.

▤ Slide · s9-B2-2 custom B2 L82

The stack — how the pieces fit

▤ Slide · s9-B2-3 list B2 L121

What a Skill actually is

You built one in S03. Here's the pattern under it.

A packaged, reusable capability — instructions the model loads on demand
It carries its own when-to-use, its steps, and the tools it may call
Write it once; invoke it across projects, teammates, and data
Your M1 cleanup chain is a skill — you just didn't call it that

▤ Slide · s9-B2-4 split B2 L135

You already built one · S03

The ticket-dedup skill — forked from M1, made yours.

col 1 · What it was

In S03 you took the M1 cleanup chain and forked it into your own dedup skill — s

col 2 · Why it counts

That's the pattern under everything today: encode a judgment once, invoke it for

▤ Slide · s9-B2-5 split B2 L151

And a second · S05

Your design-system extraction — pulling tokens into a reusable skill.

col 1 · What it was

In the S05 Design AI Clinic you extracted your own DS — colour, type, spacing to

col 2 · Why it counts

PM and design both left with a skill. Coverage isn't theory — it's two working s

▤ Slide · s9-B2-6 list B2 L167

Tools & function calling — the model acts

A tool turns "tell me" into "do it."

A tool is an action the model can invoke — search, calculate, write a file, call an API
The model returns a structured call (name + arguments); your code runs it and hands back the result
This is how a chat becomes a doer — the model decides what, your tools do the how
Skills bundle the tools they need — that's what makes them act, not just answer

B3 24–33

Where am I really · run it on your own chat

10 items

The stack becomes a diagnostic you point at any conversation — not a Pace example, *your* chat.

▤ Slide · s9-B3-1 split B3 L185

Where Am I Really?

A prompt you paste into any chat to see what's actually happening under the hood.

col 1 · What it does

Reads the conversation you're in and tells you the highest layer it actually rea

col 2 · What it surfaces

What you keep asking for that could become a skill. Whether an agent would help

▤ Slide · s9-B3-2 activation B3 L203

Run it on your own chat

Open a recent Claude conversation, paste "Where Am I Really?", and let it diagnose what's happening there.

▤ Slide · s9-B3-3 code B3 L221

Where Am I Really?

▤ Slide · s9-B3-4 code B3 L242

…the honest verdict

B4 33–63

Reaching real data · API → MCP → RAG

47 items

One escalating story: call a real API (that's a tool) → MCP standardizes the plug (that's a connector) → RAG is the database flavor. This is the block the sponsor asked for — how the backend actually connects.

▤ Slide · s9-B4-1 pb-section B4 L270

Reaching real data

▤ Slide · s9-B4-2 reframe B4 L276

(no title)

Copy-paste is the bottleneck you've been living with.

▤ Slide · s9-B4-3 list B4 L286

What's an API — really?

How software asks other software for things. Two verbs cover most of it.

An API is a defined way to ask a system for data or to change it — a request in, a structured response out
GET — read something. Safe, repeatable. “list the goals for owner 42.”
POST — create/change something. Has consequences. “create a new goal.”
Every request carries auth (a token that says who you are) and gets back JSON

▤ Slide · s9-B4-4 activation B4 L301

Call a real API as a tool

Let's have Claude reach a live, public API — no key, no setup — and ground its answer in what comes back.

▤ Slide · s9-B4-5 list B4 L311

An API you can call right now

Open-Meteo — real weather, no key, no signup.

The endpoint: api.open-meteo.com/v1/forecast?latitude=52.52&longitude=13.41¤t=temperature_2m,wind_speed_10m
It's a GET — you're just reading. Swap the lat/long for your city.
No token, no auth — which is exactly why we can all hit it live

▤ Slide · s9-B4-6 custom B4 L330

…the raw JSON that comes back

▤ Slide · s9-B4-7 split B4 L357

Your own backend needs auth

Public APIs you just call. Yours, you introduce — with a token, and care.

col 1 · The move

Hand Claude your API docs as context (a backend-context.md), and it configures t

col 2 · The rule

Reads are safe; writes need confirmation. Never put a token on a shared screen.

▤ Slide · s9-B4-8 list B4 L373

What a Connector is — and MCP

A standard plug between a model and a real system.

A connector gives the model live access to a system — Aha!, Jira, a sheet, a database
MCP (Model Context Protocol) is the standard for those plugs — one shape, many sources
The API you just called by hand? MCP is that, made a reusable socket — build the plug once, every skill and agent uses it
Think client ↔ server: your model is the client; each data source runs a small MCP server

▤ Slide · s9-B4-9 list B4 L388

RAG — an "AI database"

How a model answers from *your* documents instead of its training.

You have a pile of docs — specs, tickets, past decisions — too big to paste into one prompt
Store them in a special database that indexes by meaning, not keywords (embeddings)
Retrieve only the few chunks relevant to the question — and feed those to the model
So answers cite your source, stay current, and don't rely on the model's memory

▤ Slide · s9-B4-10 code B4 L403

Add an MCP — no admin needed

▤ Slide · s9-B4-11 code B4 L421

…deepwiki connected

▤ Slide · s9-B4-12 gallery B4 L439

No terminal? Add it in Claude Desktop · 1 of 3

Settings → Connectors → "Add custom connector" → name it "DeepWiki" and paste the URL `https://mcp.deepwiki.com/mcp`. No OAuth needed → Add.

▤ Slide · s9-B4-13 gallery B4 L449

Add it in Claude Desktop · 2 of 3

The connector appears as "not connected yet" — one click on **Connect** and Claude reaches out to the server.

▤ Slide · s9-B4-14 gallery B4 L459

Add it in Claude Desktop · 3 of 3

Connected. You can see the three tools it exposes — **Ask question · Read wiki contents · Read wiki structure** — and choose when Claude may

▤ Slide · s9-B4-15 activation B4 L469

Run a prompt against a real MCP

Ask DeepWiki about a real repo — and watch the logs show it connect and call tools.

▤ Slide · s9-B4-16 code B4 L483

Ask DeepWiki

▤ Slide · s9-B4-17 code B4 L495

…it connects and calls the MCP's tools

▤ Slide · s9-B4-18 pb-section B4 L515

You reference — Claude wires

▤ Slide · s9-B4-19 reframe B4 L521

(no title)

You've spent today naming the pieces. The payoff — you point, Claude wires.

▤ Slide · s9-B4-20 split B4 L531

1 · Invoke it in a prompt

Name the API and what you want — Claude makes the call.

col 1 · The prompt

"Call the Open-Meteo forecast API for ⟨city⟩ and tell me if I'll need a jacket t

col 2 · What you're referencing

The tool (the API) by name, plus your intent. You don't write the HTTP request —

▤ Slide · s9-B4-21 split B4 L547

2 · Wire it into a screen

Map the response fields to the UI you're building.

col 1 · The prompt

"Build a weather card. From the Open-Meteo response, map current.temperature_2m

col 2 · What you're referencing

The exact response fields → the components they feed. This is how a real API bec

▤ Slide · s9-B4-22 split B4 L563

3 · Feed the request from real data

Where do the API's inputs come from? Point at the source.

col 1 · The prompt

"For the lat/long in the API call, use the user's saved city from their profile.

col 2 · What you're referencing

The parameters flowing IN — from a user profile, a screen, or another data sourc

▤ Slide · s9-B4-23 split B4 L579

4 · Chain a skill and a connector

Reference more than one connection in a single ask.

col 1 · The prompt

"Using my ticket-dedup skill and the Jira MCP, pull this week's new tickets and

col 2 · What you're referencing

A skill (the how) + a connector (the live data) in one prompt. Same move as the

B5 63–92

Agents · the harness, the judgment, and a real one running

27 items

The top of the stack — and the most over-used word in AI. Define it honestly, name where it actually runs, then run a real one.

▤ Slide · s9-B5-1 list B5 L599

What an agent actually is

A model running a loop toward a goal, with tools and some autonomy.

Gather context → act → verify → repeat — the agent loop, until the goal is met
It chooses which tools to call and when — you don't script every step
It uses connectors to sense the world and tools to change it
Autonomy is a dial, not a switch — more freedom, more that can go wrong

▤ Slide · s9-B5-2 split B5 L614

Workflow vs. agent

The most important design call — and the default is "workflow."

col 1 · Workflow

You fix the steps; the model fills each in. Predictable, cheap, testable. Use it

col 2 · Agent

The model decides the steps. Use it only when the path can't be known in advance

▤ Slide · s9-B5-3 list B5 L632

Where an agent actually runs

Not the chat box. A harness that can loop, use tools, and be scheduled.

Claude Code — the primary harness: runs the loop, bash, MCP, files; schedule it with cron. This is where we'll run today's agent.
Cowork (Claude Desktop tab) — no-code: multi-step and scheduled tasks, from Pro up
Agent SDK (Python / TS) — the same harness, in your own code, to ship into a product
claude.ai chat is NOT an agent harness — it uses tools inside one conversation, but doesn't run standing, scheduled work

▤ Slide · s9-B5-4 activation B5 L646

A real agent, running now

PG builds it live — check the weather, message it to Telegram, on demand.

▤ Slide · s9-B5-5 list B5 L660

Run it yourself

Four steps — and it works even with nothing set up.

Download weather-agent.zip from today's session-material folder, and unzip it
Open a terminal and cd in — type cd then drag the folder into the terminal and press Enter (Windows: drag into Windows Terminal, or Shift-right-click → Open in Terminal)
Run it: claude -p "$(cat AGENT.md)" — or just start claude and type /weather
No Telegram? Still works — it prints the message to your terminal instead of failing

▤ Slide · s9-B5-6 custom B5 L674

(no title)

▤ Slide · s9-B5-7 custom B5 L685

How the pieces talk

▤ Slide · s9-B5-8 code B5 L706

Set it and forget it

▤ Slide · s9-B5-9 custom B5 L722

The same agent in Cowork — and its limits

▤ Slide · s9-B5-10 activation B5 L740

Turn one thing you do into an agent

You just saw mine — now take a task YOU do by hand and make it a runnable agent.

▤ Slide · s9-B5-11 custom B5 L758

Worked example — your dedup skill as an agent

▤ Slide · s9-B5-12 list B5 L778

Make it trustworthy — grounding & evals

Autonomy without a check is just confident guessing.

Grounding — feed real data (retrieval / RAG) so answers cite the source, not vibes
Evals — a repeatable test set: does it still do the job after you change the prompt?
Guardrails — constrain what it can touch; keep a human on the risky moves
Taste is the part the rubric can't score — that's your layer

B6 92–100

Data safety · close

13 items

The layer that decides whether any of this can touch production — and the reason your real wiring doesn't happen in a Zoom.

▤ Slide · s9-B6-1 list B6 L796

Real data, real rules

The moment you connect live data, safety stops being optional.

Know what you're sending — customer PII, secrets, and internal data don't belong in every prompt
Prefer least access — connect to the narrowest slice the job needs, read-only where you can
A live demo in a shared Zoom is not where you wire production data — that's why real wiring is yours to do safely
When unsure, ask — the cost of a leaked field dwarfs the cost of a question

▤ Slide · s9-B6-2 split B6 L810

Data-Safety Check

Prompt move · evaluative — the go/no-go template, your connector dropped in.

col 1 · The template

Connector I'm about to wire: ⟨source, fields I plan to read, where the model run

col 2 · Filled for Pace

SOURCE — Google Calendar (event titles, attendees, start/end) + Apple Health sle

▤ Slide · s9-B6-3 code B6 L827

…the fields that must never leave

▤ Slide · s9-B6-4 pb-section B6 L843

Questions?

▤ Slide · s9-B6-5 statement B6 L853

A prompt asks. A skill repeats. A tool acts. A connector reaches live. An agent

A prompt asks. A skill repeats. A tool acts. A connector reaches live. An agent decides. You judge.

▤ Slide · s9-B6-6 pb-thankyou B6 L866

(no title)

After this session

- **Add one layer** — take your capstone up exactly one rung of the stack this week (skill → connector, or connector → workflow/agent)
- **S10 (Wed) · Capstone Review & Final Q&A** — bring what you wired; leadership joins for the final showcase
- **Safety first** — if you connect live BW data (Pluto included), do it on your own machine with least-access; ask before wiring anything with PII

## Open items (for polish pass / PG review)

0. **Round-2 delivery-prep revisions (2026-07-22)** — "tonight" → "today" throughout; added an S08 recap slide and two *cohort-built* skill recaps (S03 dedup, S05 DS-extraction — S05 has no repo file, grounded on session + PG Google Doc, `[?CONFIRM]`); made the API segment a **real public-API hands-on** (Open-Meteo) with a pre-tested Berlin response slide; consolidated "reaching real data" so API comes before MCP; added an **agent-harness** slide (Claude Code primary; claude.ai chat is not a harness); added **Add-an-MCP** (`claude mcp add`) + a DeepWiki run slide that highlights the connection/tool-call logs; added the **weather→Telegram live climax** (PG-demo, Claude Code). New prompts: `call-a-real-api`, `weather-telegram-agent` (rehearsal recipe). Runs ~100 min with live-trim markers.
1. **Cohort tool for the live API call** — Claude Code curls it; claude.ai fetches via web tool. `[?CONFIRM]` which the cohort uses so the run doesn't stall.
2. **S05 skill specifics** — confirm exact name/output of the design-tokens skill (no packaged file in `prompts/`; content in PG's Google Doc).
3. **Weather→Telegram rehearsal** — dry-run end-to-end (real message arrives) before delivery; keep the bot token off-screen.
4. **Status** — left at `draft`; not cohort-visible until you release S09.

Add ?token=… to the URL.