Week 8 · Synthesis · The whole system

The Whole
System.

You already work this way. Over seven weeks we named the five layers underneath it — governance on the floor, the quality gate on top.

Preston Magouirk · DC CAP Enterprise AI Leadership Pilot · How it all connects.

Seven weeks, one build-up

We built it one layer at a time.

Every week added a piece. In Week 2 we named the four competencies. In Week 3, the things you build. In Weeks 5 and 6, how you check the work. In Week 7 we put them together. Stack the pieces and you have the system this deck walks through — and you have been using it all along.

Always on

Governance

The four data tiers. The floor every other layer sits on. Classify before you upload.

Week 2

The 4 D's

Delegation, Description, Discernment, Diligence. The competencies underneath every move.

Week 2 + 3

The 3 A's

Automation, Augmentation, Agency. How the human and Claude split the work.

Weeks 3, 4, 7

The containers

Prompt, Skill, Project, Agent. What you build, and what becomes reusable.

Week 6

The 8 questions

How you know the output is good. Week 6's verification set — Diligence, made into a checklist.

The whole system on one slide

Five layers. Five questions you answer every time.

Layer 5 · The gate
The 8 quality questions
Diligence as a checklist. Nothing leaves the workspace until it passes — light, standard, or full gate by stakes.
"How do I know it's good?"
Layer 4 · What you build
The 4 containers
Prompt, Skill, Project, Agent. Default to the smallest container that survives the job.
"What am I building, and what becomes reusable?"
Layer 3 · The split
The 3 A's
Automation, Augmentation, Agency. The balance of human judgment and Claude's work.
"How do the human and the AI split the work?"
Layer 2 · The competencies
The 4 D's
Delegation, Description, Discernment, Diligence. The four skills underneath every interaction.
"What am I actually doing?"
Layer 1 · The floor
Governance — 4 data tiers
Classify the data before it touches Claude. When in doubt, treat it as one tier more restrictive.
"Am I allowed to put this in?"
↑ The gate — checked last, before anything shipsThe floor — checked first, every time ↓
Layer 1 · The floor

Governance: can this data go in?

You start here every time. Before you write a prompt or open a Project, you classify the data. You use the four tiers to decide what Claude may see. When you are unsure, you move up one tier and treat the data as more restricted.

Tier 1 · Prohibited

Never in Claude

Student PII, FERPA records, SSNs, financial-aid detail. Stays out, no exceptions.

Tier 2 · Restricted

Role-specific only

Budget data, staff evaluations, partner-confidential, small-cell aggregates under 10.

Tier 3 · Open to licensed users

Most of your work

Strategy docs, grant narratives, program models, de-identified aggregates.

Tier 4 · Unrestricted

Already public

Published materials, public data. No restriction on use.

The governance framework on the hub holds the full tier definitions and the incident-response path. This is the question that comes before every other question in the system.

Layer 2 · The competencies

The 4 D's: what am I actually doing?

Every time you work with Claude, you draw on four skills. Name the ones the task needs, and you know where to put your attention.

Delegation

What to hand off

Decide which part of the work goes to Claude and which stays with you. The judgment about division of labor.

Description

How to ask

Spec the work precisely — the goal, the context, the format, the constraints. The prompt and the Project instructions.

Discernment

Reading the output

Judge the draft in real time. Catch where it is thin, off-voice, or wrong before it goes anywhere.

Diligence

Verifying before it ships

Check the claims, the numbers, the sources. The discipline that becomes the eight questions in Layer 5.

Adapted from the Anthropic 4D framework (Dakan & Feller, 2025). These are the named pieces of fluency — every higher layer is built from them.

Layer 3 · The split

The 3 A's: who does the work?

The three modes are built from the four D's. Each mode is a different balance of human judgment and Claude's output — and each draws on a specific pair of competencies.

Automation
You hand off a defined task and verify the result. Claude does the doing.
Delegation Diligence
HumanClaude
Augmentation
You think alongside Claude — drafting, pushing back, refining together in the moment.
Description Discernment
HumanClaude
Agency
Claude runs a process you authored, across steps, with checks between them. All four D's are in play.
Delegation Description Discernment Diligence
Human · designsClaude · runs

The bar shows the balance of execution. In Agency the human share looks small because the work has moved up front — the judgment lives in the process the designer wrote.

Layer 4 · What you build

The containers: what becomes reusable?

Four containers, one ladder. Each one makes a different thing reusable. Every higher container holds the lower ones, so moving up is cheap. Start at the bottom and promote only when the work demands it.

Stage 1

Prompt

Reuses nothing

One person, one task, one time. The move lives in your head.

H1 · Personal
Week 8 build
Stage 2

Skill

Reuses the move

Same procedure, varied inputs. The team runs your best version every time.

H2 · Team
Week 8 build
Stage 3

Project

Reuses the context

Same body of knowledge, varied asks. Files, instructions, and governance in one workspace.

H2 · Team
Stage 4

Agent

Reuses the trigger

Cadence or event-driven, across roles, with an audit trail. Also called an orchestration — the build that puts you in Agency mode (Layer 3).

H3 · Enterprise

Default to the smallest container that survives the job. Most pilot builds land at Stage 2 or Stage 3 — that is the right call. The two highlighted containers are what each unit makes available this week.

Layer 5 · The gate

The 8 questions: how do I know it's good?

Diligence turns into a checklist. These are Week 6's eight verification questions — the ones every new user asks the first time Claude lands a confident, wrong answer. You met the mechanics behind them in Week 5 and the build-questions in Week 7; this set is the quality gate. Run them on every container before anything leaves the workspace.

1

Why does Claude hallucinate?

2

Why does it sound confident when it's wrong?

3

Where is Claude most likely to be wrong?

4

What is context overload?

5

Why do my earlier rules fade?

6

What manual checks should I always run?

7

How do I ground Claude in real sources?

8

How do I build verification gates into the workflow?

Question 3 names the five danger zones: names, numbers, dates, citations, and your world. Those are the first places to check, every time.

The system, in one task

A partner-update draft, top to bottom.

Take a real task — a quarterly update to a university partner. The same five layers run, in order, whether you write it once or build it for the team.

GovernanceLayer 1
The update uses retention rates and program summaries — Tier 3, open. No individual scholar records go in. Floor cleared.
The 4 D'sLayer 2
You lean on Description to spec the update and Diligence to plan the checks. Discernment reads each draft.
The 3 A'sLayer 3
First pass is Augmentation — you draft with Claude. Once the unit writes it every quarter, it shifts toward Automation.
The containerLayer 4
One update is a Prompt. A repeating quarterly update earns a Skill for the move or a Project for the partner context.
The 8 questionsLayer 5
Verify the partner name, the retention numbers, the dates (the danger zones). Ground every figure in the source data. Standard gate, then send.
The Week 8 move

Build the two containers your team will reuse.

Over these two weeks, each of the six units builds and shares at least one Skill and at least one Project in Claude Enterprise. No one turns anything in. The bar is simpler: a teammate opens it and uses it, ideally anyone in the org.

Stage 2 · Reuses the move

One Skill

Take the move your unit makes again and again and write it down once — the steps, the standards, the voice. The team loads it and gets your best version every time.

In your world: a student-outreach Skill, a partner-communications Skill, a governance-QA pass, a data-cleaning routine.
Stage 3 · Reuses the context

One Project

Bundle your unit's working knowledge into one Claude Project — the files, the instructions, the data-tier rules. Anyone on the team produces a strong first draft against shared context.

In your world: a funder-update Project, an analysis workspace, an onboarding Project for the next person who joins.

These are the two team-horizon containers from Layer 4. One packages your best move. One packages your shared context. Both pass the eight questions and carry a governance line before you share them.

Why these two

A Skill and a Project are the team pair.

The pairing maps straight to the ladder. A Skill and a Project are the two containers that live at the team horizon, and together they cover what a team actually reuses.

01

The move + the context

A Skill carries how your team does something. A Project carries what your team works from. Build one of each and you have covered both halves of repeatable work.

02

It replaces tribal knowledge

The move that lived in one person's head becomes a Skill anyone can run. The context scattered across folders becomes a Project anyone can open. The unit keeps working when someone is out.

03

It is H2, by design

We named this the team horizon back in Week 2. A personal prompt sits at H1. An agent that runs across roles sits at H3. A Skill and a Project are what make the whole team faster — the realistic win for a 60-day pilot.

What carries forward

H1 personal. H2 team. H3 enterprise.

The three-horizons frame from Week 2 closes the loop. Each build lives at one horizon. This week, every unit plants something at H2 — the level where the whole team inherits the work.

H1 · Personal gain

Faster work on your own plate

The Monday-morning prep, the prompt you run every week. Where most Stage 1 work belongs, on purpose.

H2 · Team gain

The Skill and the Project

What every unit makes available this week. Your team loads the move and opens the context. The pilot's center of gravity.

H3 · Enterprise gain

How DC CAP operates

The financial_modeling orchestration, the AI Development Office pipeline. Rare, and a year in the making. The Ship It deck walks one live.

The pilot ends. The system stays.

This is
how we work now.

Governance on the floor. The four D's, the three A's, the four containers, the eight questions stacked on top. You run the whole thing every time you open Claude, and now you can name each layer. Build the two containers your team will use, make them available, and tell me where they break.