AI in the Scale ERA

Part 3 — Scale

8. AI Harness — the agent scaffolding
9. MiniApp → Enterprise
10. SME → PCL maturity ladder
11. ESG & Reporting (tied to Week 4)

Part 4 — Take it home

12. Durable vs Disposable — what's worth investing in
13. Starter Pack — your Monday-morning checklist
14. FAQ — security, cost, governance, rollout

Intent-first

Describe the problem and the result you want — not the implementation. "I want an app that does X" rather than "write a for-loop that…"

Iterate by feel

Run it. If something's off, refine the prompt and try again — you don't debug the code, you debug the description.

AI is the developer · you're the PM

Roles flip — you judge "right or not yet"; Claude writes, runs, tests, and fixes. It's like briefing a senior developer on requirements.

Describe (round 1)

"Build a React app where users upload a CSV (columns: id, message, date). Show it as a table filterable by sentiment. Use Tailwind. No backend — browser-only."

Refine (rounds 2-3)

"Add a category dropdown filter. Put summary cards on top showing counts per sentiment. Switch to dark mode. If the CSV is malformed, show a clear error message."

From chat to agent

Claude no longer simply answers — it plans, calls tools, iterates, and verifies independently. The unit of work has shifted from "a reply" to "a completed task."

From prompt to skill

Repeatable expertise is packaged as Skills, Plugins, and MCPs. Build once, reuse across the organisation — well-crafted prompts become institutional assets.

From sandbox to system

Integrated with files, calendar, CRM, and ERP via MCP connectors. AI moves from an auxiliary tool to an embedded component of core operations.

Chain-of-thought

"Think step-by-step, then answer" — meaningfully better accuracy on multi-step reasoning tasks.

Few-shot examples

Show 2–3 input/output pairs in the prompt — Claude learns the desired pattern instantly.

Output schema

"Reply in JSON matching this schema: …" — guarantees machine-parseable output.

Role framing

"You are a senior auditor reviewing…" — shifts register and rigor of the answer immediately.

Self-critique loop

"Draft → critique → improve" — now native in Opus 4.7, still useful for Sonnet/Haiku.

Negative constraints

"Don't do X" sometimes works better than "Do Y" — especially for preventing known failure modes.

Prompt Engineering

Shape the model's behavior without touching weights — instructions plus context are enough for most use cases · 0–10 examples · iterate in minutes.

All three demos today live on this rung.

RAG (Retrieval-Augmented Generation)

Connect Claude to your own knowledge base — fetch relevant chunks into the prompt on every call · ideal for internal docs, policies, FAQs the model has never seen.

Demo #2 (policy markdown) is the simplest version of RAG.

Finetuning

Continue training on your own data — changes model weights · hundreds to tens of thousands of examples · improves long-term latency and cost, but takes real time and resources.

Consider only after prompt + RAG can't reach your quality bar.

Refusal

Symptom: "I can't help with that…" · Cause: ambiguous or risk-flavoured request · Fix: add business context — "I'm a compliance manager at X reviewing…"

Hallucination

Symptom: numbers/names/dates not in the source · Cause: no grounding · Fix: "answer only from attached docs · if missing, say 'not specified'"

Schema break

Symptom: output isn't valid JSON · Cause: schema under-specified · Fix: give the full JSON schema + "if any field is uncertain, set null · do not fabricate"

Length drift

Symptom: too long or too short · Cause: scope under-specified · Fix: "5 bullets, each ≤ 20 words" · if over-limit, chunk input first

Tone drift

Symptom: casual voice in formal context · Cause: no role framing · Fix: "You are a SOX auditor · write as it would appear in the annual report"

Prompt caching

Long system prompts + reusable reference docs — cached for 5 min · cache-hit input tokens cost only 10% of normal · essential for any skill that runs frequently

Batch API

Anything that doesn't need real-time (reports · scheduled briefs · content batches) — send via batch API · half price · results within 24h

● Event

An occurrence — a form is submitted, an email arrives, or a timer fires.

▭ Task

Work performed — by a person, by Claude, or by a system.

◇ Gateway

A decision point — "if the amount exceeds 5K, route to the Finance Director."

→ Flow

Direction of work — defining who passes what to whom.

1. It scopes the prompt

Once the process is drawn, the boundary becomes clear: which tasks Claude owns, and which remain with humans.

2. It surfaces the data

Every arrow represents a payload. Labelling the arrows identifies exactly which CSV, document, or API call is required.

3. It becomes the audit trail

During audit, the BPMN diagram combined with Claude's run log forms a complete compliance evidence package.

v0.dev Vercel

Specialised for Next.js + shadcn/ui · strong on polished UI · one-click deploy to Vercel · best for landing pages, marketing sites, and dashboards.

Bolt.new StackBlitz

Full-stack in-browser · Node + frontend in one place · clear file tree · best for MVPs that need backend logic.

Cursor IDE

A VS Code fork with AI baked in · best if you already live in an IDE · works on your local repo · Cmd+K rewrite plus side-panel chat.

Policy ≡ UI

The markdown in the textarea IS the system. Edit it → the behavior changes. Audit trail via the policy-version counter.

Humans see only judgment calls

The app auto-approves clear cases; humans review only the flagged exceptions.

Exportable decision pack

Decision CSV + .eml emails + policy snapshot — all zipped, ready for audit and handover.

Reads 4 sources

Sales, stock, tickets, and customer master — synthesised into insight, not merely aggregated.

Sub-agents

Each check runs as an independent sub-task; the lead agent composes the final output.

Schedulable

This prompt becomes a daily 06:00 scheduled task — the mini-app is operational.

Output looks right but isn't what you wanted

Symptom: prompt left gaps, AI guessed — sometimes right, sometimes not.

Recovery prompt:

"Show me 3 concrete input → output examples of what you're about to build. I'll confirm before you write any code."

Output truncates or runs too long

Symptom: response cuts off mid-file, or Claude keeps asking you to "continue."

Recovery prompt:

"Show only the changed section in diff format. Don't rewrite the whole file."

Fixing one bug creates another — infinite loop

Symptom: 4-5 rounds in, AI fixes one spot but breaks another.

Recovery prompt:

"Stop. Start fresh — rebuild this from scratch, knowing everything we've discussed. Write me the prompt you should have received from the start."

Claude says "I can't do that" or asks for more info

Symptom: refusal based on perceived limits, or endless clarifying questions.

Recovery prompt:

"Take your best shot. Use sensible assumptions for anything unclear, and list those assumptions at the top. I'll correct them later if they're wrong."

Output uses libraries/APIs that don't exist

Symptom: runs fail with "module not found" or calls to nonexistent functions.

Recovery prompt:

"Use only browser-standard APIs. If you need to install anything, name it + version + install command first. Error: <paste error>"

App works but doesn't look polished

Symptom: UI looks like a skeleton — functional, but real users won't want to engage with it.

Recovery prompt:

"Make it look like a Stripe / Linear / Vercel product — more spacing, clearer typography, a neutral palette with one accent color. Add hover states and micro-interactions."

Orchestration Loop

The "brain" — decides the next step, runs a task, comes back to think again until done.

Tools

What it can actually do — read/write files, call APIs, run code, browse the web.

Memory

What it remembers across turns — short-term in context, long-term in files/DBs.

Context Management

What gets into the context window — compress, prune, summarise before discarding.

State Persistence

Long jobs (hours/days) survive across sessions and pick up where they left off.

Guardrails + Observability

Stops things going wrong + lets you see when they do — evals, logs, kill switches, schema checks.

Persist externally

Keep context outside the window — file, database, or memory store · fetch back on demand.

Seen in Demo #2: the policy markdown file

Retrieve relevant

Pull only the relevant slice via RAG / lookup · don't load everything into context.

Seen in ESG Demo: grid-factor lookup by region + year

Summarize

Condense long context before passing on · fewer tokens, better signal-to-noise.

Seen in Demo #3: lead agent merging sub-agent outputs

Separate contexts

Give each sub-agent its own context window · they only see what they need.

Seen in Demo #3: sub-agents A/B/C/D each read their own CSV

Boundaries

Who may run it, on which data, and within what limits.

Memory

What is retained across runs, and what must be discarded.

Observability

How you confirm it is working, and how quickly you detect when it is not.

Human in the loop is mandatory

AI suggests; a person decides and clicks send every time · maps to Stages 1–2 on our ladder.

AI interacts with internal users

AI talks to your own employees directly — with audit, eval, and rollback in place · maps to Stages 3–4.

deploy customer-facing chatbots to production

deploy external recommendation algorithms to production

Environmental disclosure

PCL requirement: Scope 1/2/3 emissions, climate-related risk per TCFD

Where Claude helps: extract emission data from invoices and utility bills, normalise units, draft the narrative sections

Use case: quarterly Scope 2 calculation from electricity bills — from 15 hours to 30 minutes

AI's own footprint: select smaller models (e.g. Haiku) for low-stakes work — reduces both cost and carbon emissions

Social disclosure

PCL requirement: human capital, diversity, supplier code of conduct

Where Claude helps: synthesise employee survey data, screen supplier attestations, draft DEI sections

Use case: process 1,000 supplier compliance attestations in a day — compared to approximately 50/day manually

AI's own social impact: bias auditing, output accessibility, human-in-loop for decisions affecting individuals (hiring, lending)

Governance

PCL requirement: board oversight of all material risks (including AI risk), audit committee, whistleblower channel

Where Claude helps: synthesise board reports, draft policy documents, monitor compliance continuously

Use case: continuous monitoring of expense policy violations — flagging on detection reduces detection time from months to hours

AI's own governance: model cards, eval logs, RBAC — maps directly to the 6 components of AI Harness

AI prepares the groundwork

Use AI to assemble initial ESG data, draft narrative content, and understand the frameworks that will apply at higher stages.

AI runs continuous ESG

Connect to utility providers, HR systems, and supplier portals via MCP — capture data continuously rather than collating it quarterly.

Same shape as Demo #2

Multi-source data plus a policy reference (here, the grid factor table) plus a decision table — the same shape as the PR approval workflow.

Uses 4 of the 6 components

Orchestration loop, tools (CSV reader), memory between steps, context management — only state persistence and guardrails are added when this runs monthly on a schedule.

BPMN + run log = evidence

When the auditor asks "why is Q1 2026 this number?", the run log contains every calculation step with source citation.

Study in depth

BPMN · the 6 harness components · SME → PCL ladder · the process / skill / connector / model separation

Be aware; do not memorise

Framework tables · benchmark figures · specific model names · current prices — review for context, not for retention

For next week

When selecting a vendor, ask: "if we replaced this vendor tomorrow, what would we lose?" The correct answer: only disposable elements

Technology

Once: whoever trained the best model won · Today: the best foundation models are available to everyone via API · technology has become a commodity — virtually every team gets to use the same Claude / GPT / Gemini.

Warning: if your strategy is "we use Claude" — that's not a moat.

Distribution

The ability to put a product in front of many users · belongs to Google, Microsoft, Apple, Meta — companies already inside users' devices, browsers, and OSes · SMEs can't out-distribute these giants head-on.

Stress test: "Could this product become a feature inside Google Docs within 12 months?"

If we don't, AI-enabled competitors will make us obsolete

AI poses an existential threat to the business · most visible in document processing, financial analysis, insurance, advertising · the cautionary tales: Kodak, Blockbuster, BlackBerry — companies that responded too late.

7% of executives in Gartner 2023 fall here · invest fully · highest priority

We're not sure where AI will fit yet — but don't want to be left behind

Chasing every hype train is bad — but many companies died from waiting too long · if you can afford it, investing to understand how AI will impact your business is the right call.

Fits R&D budget · low investment · high tolerance for failure

This deck

Reopen it as a reference whenever a new colleague is onboarded.

Sample data set

8 files in sample_data/ — replay the demos independently.

One BPMN sketch

Of your first process. Bring it to the next session for review.

Claude Projects

Context that persists across conversations · upload your skill repo + reference docs once, reuse every session · ideal for onboarding company knowledge into Claude.

Claude Agent SDK

Custom harness from 20-50 lines up · the standard path from scripted skill → mini-app → service when a workflow grows bigger than a single prompt.

Computer Use

Claude controls a browser directly — automation across apps that have no API (legacy dashboards, vendor portals) · turns "can't be done" into "can be done, with oversight."

Eval set + observability

When a workflow has an SLA or a regulator will inspect it — a 20-50 question eval set that runs on every deploy, plus run logs + alerts · this is the "durable" layer that survives model upgrades.