Designer → Agent. Eight questions that turn "I'd like Claude to do X" into a build the team actually inherits.
Preston Magouirk · DC CAP Enterprise AI Leadership Pilot · Spec the work.
The job stops being "prompt the model" and starts being "spec the work." Eight questions every team leader asks the first time they realize the build, the skill, the project, and the orchestration are the same conversation at four different scales.
Same plain-answer rhythm as Week 5 and Week 6. Heavier focus on choosing the smallest container the work actually needs.
The agent has no expertise.
The designer has the expertise.
Claude predicts the next word that fits the pattern. The pattern is whatever you encoded. A generic spec produces generic output. A spec that carries your ninety-second explanation of what "good" looks like produces output that carries it too. The failure mode lives in the spec the designer wrote.
Every higher stage strictly contains the lower ones. Promotion is cheap. Start small. Move up only when reality demands it. Most pilot builds will stop at Stage 2 or Stage 3 — that is the right call.
SKILL.md any agent can load on trigger. The team needs consistency; the expertise gets encoded once.my-skill/ ├── SKILL.md ← name + description │ + instructions (<500 lines) ├── scripts/ ← deterministic code ├── references/ ← loaded on demand └── assets/ ← templates & outputs
preston-writing · funder-framing · dc-cap-org-intelligence · sf-demo-dataMy Project ├── System instructions ← role, voice, scope ├── Reference files ← curated, anchored ├── Governance config ← tier + escalation └── Workflow doc ← how the team uses it
An architect designing a community garden does not start with city planning. A fence is enough when one neighbor owns the work. A vault is the right move when something has to be secured for years. A building is the right move when many people use the space the same way. A city is the right move when the work crosses neighborhoods.
The job picks the container. The architect who starts every project with city planning ships nothing.
One person, one recurring task — a Prompt is enough. The team does the same thing repeatedly — Skill. Multi-step workflow with files and governance — Project. Crosses roles, needs an audit trail — Orchestration. The ladder is strict; every higher stage contains the lower ones, so promotion is cheap.
The signal to move up: the prompt is getting copy-pasted into team chats, or the team is asking "how do we do this consistently?" That is the moment to promote. Most pilot builds will stop at Stage 2 or Stage 3.
The takeaway: Start small. Promote when the work demands it. Knowing when to stop is the discipline this week trains.
A working restaurant kitchen runs on mise en place — every ingredient prepared, portioned, named, and placed within the cook's reach before service starts. The chef who knows the menu can list the mise en place in ninety seconds. The chef who does not produces a mise that fights the dish all night.
The expertise lives in the prep. The cook's pan-handling is execution; the prep is where the dish was already decided. When the prep is right, the work flows. When the prep is short, every plate fights the gap.
Claude has no expertise of its own. The build encodes your knowledge in a form the model can apply at scale. If you cannot explain to a teammate, in ninety seconds, what "good" looks like for this job — the spec is too thin. Specifics in, specifics out.
Two reps before you build: write the ninety-second explanation first, then convert to instructions; borrow expertise where you do not have it — that is exactly what the skill library is.
The takeaway: The failure mode lives in the spec. The model executes what the spec asks for.
A surgeon does not also run anesthesia, also scrub, also count instruments, also chart, also close. The team splits the work into named roles with explicit handoffs. The anesthesiologist signals "ready to incise." The circulating nurse counts before close. Each handoff is a state contract: this output reaches that input.
Pre-surgical-team operating rooms had a single surgeon doing everything. The mortality rates were what you'd expect.
A simple build has one role. A complex one has several, each with one job and an explicit handoff to the next. The signal to split: the role description reads like a small org chart inside one instruction. The signal to keep it together: the workflow is one person's standard move, end to end.
Inside the financial_modeling H3 build: fpa-lead dispatches, fpa-data-engineer ingests, fpa-quant models, fpa-storyteller writes. Each output a state file the next agent reads.
The takeaway: One job per role. Explicit handoffs between them. The Lead never overwrites; the Lead routes.
A constitution carries the durable rules — the structure that holds across every cycle. The day's news carries the world those rules apply to. A republic that rewrites its constitution every news cycle has no rules. A republic that updates its news from a 250-year-old document has no facts.
The split is the point. The durable rules hold; the world updates. Both surfaces work because they live in different places and refresh on different cadences.
Instructions: the durable spec. Role, voice, format, the rules that hold across every conversation. Short and prescriptive — if a rule needs three paragraphs to explain, the rule is doing too much.
Files: the world the rules apply to. Anything that refreshes weekly. Curated, named, front-loaded and back-loaded with what matters most. The middle of a long file fades; the top and bottom stay sharp (Week 6 Q4 was a real cost).
The takeaway: Instructions are the rules. Files are the world the rules apply to.
A symphony does not have one player doing strings, woodwinds, brass, and percussion. Each section is its own named expertise. The score chains them — each section enters where the score calls it in, and the conductor routes the timing. No section knows about another section's fingering; the chain handles the integration.
The same musician can also play in a smaller chamber group, where the score is different and only three sections are called. The expertise travels with the player; the chain travels with the score.
Most real builds chain two or three pieces of named expertise. A grant draft chains funder-framing + preston-writing. A board brief chains executive-summary-formatter + data-interpreter.
The same skill loads inside many agents. dccap-brand loads in fpa-frontend AND fpa-viz inside the financial_modeling build. Same SKILL.md. Two different lanes. No duplication.
The takeaway: Name the chain. The chain is the inheritance.
A Swiss Army knife with 35 tools is heavier, more expensive, and worse at almost every task than a knife with three tools. Survival experts carry fewer tools by deliberate practice. The reason is that every additional tool is one more surface that has to be maintained, sharpened, and remembered. Most expedition packing lists explicitly cap tool count.
The discipline is not minimalism for its own sake. It is the recognition that capability has a cost — to the user, to the operator, to the system.
Most builds need less than people think. Web search, file read, code execution — each tool adds a surface that has to be governed. Default to the minimum. Add a tool when the workflow cannot run without it. Tools that ship off by default ship off for a reason.
For each tool you plan to enable, write down two answers: what is the worst case if this fires incorrectly? and who catches it before it ships? If either answer is unclear, the tool stays off until you can answer both.
The takeaway: The smallest tool set is the safest tool set.
Failure Modes and Effects Analysis (FMEA) started at NASA in the 1960s and became the standard for high-reliability engineering. Before any system ships, the team enumerates the ways it can fail, sorts by severity (cosmetic / costly / catastrophic), names the early-warning sign for each, and assigns the check that catches it. Apollo, the Space Shuttle, modern commercial aviation all run on this discipline.
The teams that fail FMEA are not the teams that miss a failure mode. They are the teams that ship without writing the failure-mode map at all.
Good is what your team would call "ready to ship" with you out of the room. Failure modes fall into three buckets. Cosmetic: formatting drift, voice slip — caught at light gate, costs minutes. Costly: wrong number, wrong audience — caught at standard gate, costs trust. Catastrophic: FERPA leak, fabricated funder fact — caught at full gate, costs the relationship.
Write the failure-mode map before you ship. One row per failure mode, with the bucket, the early-warning sign, and the gate that catches it.
The takeaway: A build without a failure-mode map is a build that ships its failures.
A building code is written once, but it earns its keep over decades. It names the durable rules (structural-wall standards, fire-egress paths), the inspection cadence (occupancy permit, annual fire inspection), and the escalation path (variance request, code-board review). The code is what keeps the structure standing through three owners and two renovations.
Cities without code revisits — without inspections, without iteration — produce buildings that look fine for ten years and start failing in year fifteen. The code is the cadence.
Every build needs three explicit lines: a verification cadence (light / standard / full gate + who runs it + before what ships), an iteration cadence (what triggers a revisit — new data, new team member, quarterly refresh), and a governance line (data tier + escalation path + restricted topics).
A build ships with no review cadence — six weeks later the data has moved, the team has changed, and the outputs are confidently wrong. Write the three lines into the build itself.
The takeaway: A build's longevity is the discipline of its cadence.
DC CAP's financial modeling tool combines three live data streams into one audit-trail-grade view of the organization's financial position now and through FY35. The arc of the build is the arc of the eight questions answered at the upper rung of the ladder.
Development raise targets across individual, corporate, and philanthropic sources. Finance historical-and-projected actuals. Investment-market projections from Pathstone. Three refresh cadences, one reconciled model.
fpa-lead · fpa-architect · fpa-data-engineer · fpa-quant · fpa-insight-hunter · fpa-analyst · fpa-strategist · fpa-modeler · fpa-viz · fpa-storyteller · fpa-frontend.
Source-cell provenance · Parser hard-gate · Runtime cross-foots · Defaults alignment. Latest build: 53/53 checks green.
The orchestration is a process the designer authored. The agents have no expertise. The designer has the expertise. The failure mode lives in the spec the designer wrote.
Each artifact teach-page walks the same eight questions concretely. Pick the artifact whose stage matches the work you would actually ship for your team in the next two weeks.
fpa-quant headline numbers in, Investment Committee-ready Big Idea sentence out. One pasteable prompt. Light gate. Tier 3.
v2.6 of the voice skill any host loads — grants, LinkedIn, board memos, fpa-storyteller. Pushy description, fast-path routing table, evidence trailers under every rule, eleven differentiated samples in reference/. Standard gate.
Thought experiment: compress the Stage 4 system into one Claude Project. What fits, what chains as Skills, what cannot collapse. Standard gate.
The real production build. 3 data streams · 11 Opus agents · 13 references · 4 verification layers. Full gate. Live behind Cloudflare Access.
No 1:1 pairings this week. Three drop-in office-hours blocks across Weeks 7–8. Calendar-only. Bring your build.
A, B, C, or D. The stage that matches what you would actually ship in the next two weeks. Most of us land at Stage 2 or 3 — that is the right call.
One paragraph each. The eight questions are the same at every stage. If Question 2 surfaces a gap in your own expertise to spec — that gap is the most useful finding of the week.
Office hours: drop in with your sketch. AI Friday: walk one through together. Week 8 closes with the showcase — three slides each, June 4. The H3 financial_modeling tour runs live.