The deep version: one production build, taken apart. DC CAP's financial_modeling tool as a Stage 4 orchestration — three data streams, eleven agents, four verification layers. Read it to see where the ladder ends.
Preston Magouirk · DC CAP Enterprise AI Leadership Pilot · The advanced example.
Each phase built on the prior. Foundation → Delegation → Description → Discernment → Identify & Design → Diligence → Build Your Own → Ship It. The phases were strict containers; promotion happened when the work demanded it.
Onboarding. The 4D framework. Automation / Augmentation / Agency. The kickoff that named what each modality is for.
Prompt patterns. The three-horizons frame. The build filter (five criteria). Three real builds you could steal.
How the model actually works. Where it gets confidently wrong. The verification stack — five danger zones, three gates, the source-loaded rewrite.
Designer to agent. The four-stage ladder. The eight build-questions. The Skill and Project you make available.
Your assignment is a Skill and a Project — the team-level containers. This deck walks the far end of the ladder so you can see what Stage 4 looks like and why almost no first build needs to go there. Read it for the pattern, at a larger scale than your own work.
DC CAP's financial_modeling tool. One surface that combines three live data streams into a board-grade view of the org's finances through FY35.
The work crossed several roles and needed an audit trail. That is the signal to orchestrate. Most work never crosses that line, and should not.
A Skill and a Project. The same five layers, at team scale. Stage 4 is here so you can see the whole ladder, not so you climb it.
A Stage 4 orchestration started life as a smaller container that outgrew its bounds. The financial_modeling tool is exactly that arc. DC CAP needed one surface that combined three live data streams into a single audit-trail-grade view of the organization's financial position now and projected through FY35.
The orchestration is a process the designer authored. The agents have no expertise. The designer has the expertise.
Three refresh cadences, three source systems, three sets of business rules. Combining them into one audit-trail-grade surface is what crossed every threshold on the four-stage ladder.
Goals and expected close estimates across individual, corporate, and philanthropic sources. Sourced from the AI Development Office pipeline and Eric's active funder conversations. Refreshes weekly.
Historical actuals (FY12–FY26), current-year tracking, and forward projections. Sourced from the canonical Finance + Development merged Budget Preview workbook. Refreshes on each Finance cycle.
Best-estimate market trajectories from Pathstone, DC CAP's investment partner. Drives endowment performance modeling and the implied draw path. Refreshes quarterly.
Stage 4 means the work is too big for one role. The dispatch protocol splits it into eleven, each with one job and an explicit handoff to the next.
Skills live at three layers — plugin (shared across DC CAP), BRAIN-shared (the named expertise patterns), and references (project-local canonical sources). The dispatch protocol controls which layer loads on which agent.
finance:* · data:*Shared organizational plugin skills. finance:reconciliation + data:validate-data in fpa-data-engineer. data:statistical-analysis in fpa-quant. Each agent declares exactly what it loads.
DC CAP-wide skills loaded by name from inside the agents that need them. fpa-storyteller loads four (preston-writing, dc-cap-org-intelligence, executive-summary-formatter, funder-framing). dccap-brand loads in both fpa-frontend and fpa-viz — same SKILL.md, two lanes.
Damodaran, Tufte, Knaflic, Hyndman, McKinsey, Bridgespan, Few, NACUBO NCSE, AFP (Tier 1) plus BoardSource, CFA IPS, NFF, NACUBO Spending Policy (Tier 2). Each carries YAML frontmatter, verbatim quotes with inline attribution, methodology summary, routing notes.
The four verification layers run automatically. Nothing ships to a Board surface unless all four pass. Latest build: 53/53 checks green.
Every value tags to {sheet, cell, value, last_seen}. Untagged values are dropped.
Refuses to ingest if anchors, sheet names, or row labels have drifted from contract. Caught at intake.
Computed totals cross-foot against the workbook's own totals inside fpa-quant. Mismatches halt the lane.
Default assumptions reconcile against the IC-approved set before any surface renders. Lead consolidates the L4 result.
A request lands at fpa-lead. Lead dispatches lanes in parallel where dependencies allow, sequentially where they do not. Each lane writes to its own state file. Lead consolidates. The Output Verification protocol runs before anything ships.
fpa-data-engineer reconciles the three streams at intake (Development + Finance + Pathstone). Once L1 + L2 pass, fpa-quant runs the math with L3 cross-foots — the endowment draw recomputes from Pathstone bands, the operating gap recomputes from finance + development. In parallel, fpa-insight-hunter runs the Pareto. fpa-storyteller writes the Big Idea. fpa-frontend publishes after L4. Lead writes the consolidated audit trail.
The agents have no expertise.
The designer has the expertise.
The orchestration is not "more AI." It is a process the designer authored. The failure mode lives in the spec the designer wrote. Most of us will stop at Stage 2 or Stage 3. That is the right call. Knowing when to stop is the discipline this pilot trains.
The pilot measurement model carries forward. Three tiers track Engagement, Proficiency, and Impact. Three gates decide what scales, what pauses, what pivots. The cadence is what we use to measure how the cohort's builds hold up across the next twelve months.
Office-hours attendance. Hub-page visits. Briefing reads. Surface metrics that tell us the cohort kept showing up.
Build artifacts shipped. The eight-question answers on file. Post-pilot assessment delta against your Week 1 baseline.
The build's effect on the work. Hours saved on the recurring task. Voice consistency across the team. The pattern that became how the unit operates.
Scale / Pause / Pivot gates decide what becomes a DC CAP-wide pattern. The KPI Framework page on the hub carries the gate definitions.
The three-horizons frame from Week 2 closes the loop. Each ship-it artifact lives at one horizon. Some will scale up; most stay where they are. That is the right answer.
Counselor quick-drafts. The Monday-morning prep. The prompt template that runs every week. Artifact A lives here. Most Stage 1 work stays here on purpose.
The Skill the team loads. The Project the unit operates. Artifact B and Artifact C live here. Most pilot builds land at H2 — that is what the four-stage ladder makes accessible.
The financial_modeling orchestration. The AI Development Office pipeline. The Compensation Benchmarking flow. Artifact D points at this horizon. Rare; takes a year to mature.
Same five-construct instrument from Week 1. Your pre-launch baseline is saved. The post-pilot delta is yours to see — and ours to read across the cohort. Twenty minutes, open whenever you're ready.
Five constructs covered: understanding (what AI is and is not), delegation (when and how to hand work to Claude), description (the prompt patterns), discernment (the data-tier judgment), diligence (the verification discipline). Same scale at both ends — your numbers move because the same instrument captures what changed.
The assessment is for you first. The cohort summary informs how we scale, what we extend, where the next pilot lands.
The pilot closes. The work continues. Three things to do the week after.
Open the artifact you shipped. Run it on the real task. Read the diligence block first. Edit. Send. That single rep is what makes the build stick.
Pick one teammate who would use the same build. Walk them through it in five minutes. Watch where they stumble. That is the spec gap to fix next.
Put the iteration cadence on your calendar. The build that survives twelve months is the build whose refresh date already lives in someone's schedule.