← Pilot Hub · Week 7
Stage 1 · Prompt · Artifact A

A Practical Prompt for financial_modeling

A single pasteable instruction Preston actually uses against the financial_modeling work: draft the Big Idea sentence for an Investment Committee preface, given fpa-quant's latest headline numbers. The smallest container that survives the job.

Source: financial_modeling repo· Job: Big Idea sentence draft· Gate: Light· Tier: 3

BLUF

Artifact A is a real prompt Preston runs against the financial_modeling work: given fpa-quant's headline numbers for the FY27 revenue walk, draft a Big Idea sentence for an Investment Committee preface. The Big Idea is Knaflic's one-sentence verdict — complete sentence, unique POV, names what is at stake. It is a recurring task, owned by one person, light verification gate. Stage 1 Prompt is the right container.

The full prompt sits in the next section. It is grounded in the same financial_modeling repo the H3 walkthrough tours: same agents (fpa-quant upstream, fpa-storyteller downstream in the orchestration), same references (Knaflic's Storytelling with Data), same voice rules (preston-writing skill). The prompt extracts one task from that surface and runs it standalone.

1. The Job — One Sentence, IC-Ready, Numerically Grounded

Inside the financial_modeling work, Preston regularly produces a Big Idea sentence — a complete-sentence verdict that opens the Investment Committee preface, the board memo headline, the Eric briefing. The sentence carries one unique point of view and names what is at stake. It cites the numbers that justify it.

The full orchestration produces this sentence through fpa-storyteller after fpa-quant finishes the math and fpa-insight-hunter flags the Pareto. Stage 4 work. For the moments between full runs — when Preston has the headline numbers from a fresh fpa-quant output and wants the sentence right now — a single prompt is enough. One person, one recurring task, one pasteable instruction.

If you can extract one specific sub-task from a Stage 4 orchestration and run it standalone, that sub-task lives at Stage 1. The prompt below is exactly that extraction.

2. The Prompt — Pasteable, Re-runnable, Grounded

The instruction lives in Preston's personal notes, a Cowork sidebar, or pasted at the top of a fresh Claude chat. The same text runs every time. The headline numbers paste in below. Claude returns the sentence and the diligence block. Preston edits one word, sends.

Big Idea Sentence Draft — Stage 1 Prompt (paste into Claude)
## Role
You are drafting a Big Idea sentence for an Investment Committee
preface, given fpa-quant's headline numbers from the latest
financial_modeling build. Big Idea = Knaflic's one-sentence
verdict: complete sentence, unique POV, names what is at stake.

## Voice (preston-writing rules)
- Lead with the number that matters. Bare, no adjectives.
- Active voice. Only people can be active subjects.
- One sentence, under 25 words.
- "For context" is the signature for benchmark comparisons.
- No em-dashes. No "X, not Y" / "not X but Y" / "isn't X — it's Y".

## Inputs (paste below before running)
- FY27 raise target: <paste from fpa-quant output>
- Current pipeline expected close: <paste>
- Implied gap (raise target minus expected close): <paste>
- Pathstone Base scenario implied draw rate (FY27): <paste>
- The "three numbers that matter" fpa-insight-hunter flagged: <paste>

## Output
1. One Big Idea sentence. Complete sentence. Names unique POV.
   Names what is at stake. Under 25 words.
2. One alternate framing if the first leans optimistic — the
   3P test (possible, plausible, probable) needs to clear all three.
3. Diligence block: list every number used and where each came
   from in the pasted input. Mark any framing pulled from the
   Knaflic Big Idea pattern rather than the input.

## What you do not produce
- Do not invent funder names, committee dates, or pipeline specifics
  beyond what is in the pasted input.
- Do not generate numbers from training memory. Every number traces.
- Do not write a paragraph. One sentence; one alternate.
- Do not use any of preston-writing's forbidden words.

## Verify before sending
The diligence block traces clean. The sentence reads in under ten
seconds. The numbers reconcile to the input. The 3P test clears.

That is the whole spec. Forty lines. The prompt encodes your ninety-second explanation of what a Big Idea sentence is — pulled from Knaflic and the preston-writing skill — and gives Claude a structured surface to apply it.

3. First Run — What the Conversation Looks Like

An illustrative 90-second exchange. Numbers and scenarios below are sized to be plausible for FY27 planning; the actual figures shift with each fpa-quant build.

Step 1 · Preston pastes the headline numbers (illustrative)

- FY27 raise target: $12.10M
- Current pipeline expected close: $9.35M
- Implied gap: $2.75M
- Pathstone Base scenario implied draw rate (FY27): 10.34%
- Three numbers that matter (fpa-insight-hunter):
  1. Draw bends from 10.34% in FY27 to 5.81% by FY29
  2. Operating revenue covers 78% of operating expenses
  3. The $2.75M gap is half UP renewals, half new partner growth

Step 2 · Claude returns the Big Idea and an alternate

Big Idea: FY27 is the trough year; the math says we earn our way back to a sustainable draw rate by FY29 if we close the $2.75M raise gap on schedule.

Alternate (more probable framing): The pipeline already covers $9.35M of the $12.10M FY27 target; closing the remaining $2.75M is what bends the draw curve from 10.34% back toward 5.81% by FY29.

Step 3 · Claude's diligence block (returned alongside)

  • "$12.10M FY27 target," "$9.35M pipeline," "$2.75M gap" — direct from input.
  • "10.34% draw rate," "5.81% by FY29" — direct from input.
  • "FY27 is the trough year" — implied from the draw-rate trajectory in input #5. The trough framing is Knaflic Big Idea structure.
  • "if we close the gap on schedule" — generated. The Knaflic verdict needs a conditional that names what is at stake.

Step 4 · Preston reads, edits, ships into the IC preface

  • Diligence traces clean. No fabricated numbers.
  • The alternate framing reads more like Preston's voice (bare numbers up front, conditional on the close). Picks the alternate.
  • Replaces "the pipeline already covers" with "we already see $9.35M in the pipeline" (Preston-writing agency rule — only people are active subjects; "pipeline" is an abstract noun).
  • Ships.
Total time: about ninety seconds. The light gate caught one preston-writing rule (abstract subject). The diligence block is the audit trail.

4. What Makes This Prompt Strong

Six disciplines are encoded into the prompt itself. None of them are AI tricks; all of them come from the same framework the financial_modeling agents follow.

Role names the specific job

  • "Draft a Big Idea sentence for an IC preface" is the job. "Help me write something" is the category — too vague for the model to act on.
  • The role definition includes the Knaflic source so the model knows what shape to produce.

Voice rules come from a real skill

  • The voice section pulls directly from preston-writing — bare numbers, active voice, only-people-as-subjects, forbidden constructions.
  • The rules are observable. "No em-dashes" is checkable; "be professional" is undefined.

Inputs are paste-bound

  • Five named inputs. Each one comes from a known source (the fpa-quant output, the fpa-insight-hunter brief).
  • Claude cannot fabricate the headline numbers because the prompt demands them as input.

Output shape is specific

  • One Big Idea sentence + one alternate framing + a diligence block. Three components, named.
  • The alternate is a Damodaran 3P-test safety net — if the first leans optimistic, the alternate catches it.

"What you do not produce" is explicit

  • The forbidden list shows up before the model produces anything. The constraints sit inside the spec so the model generates against them in real time.
  • Failure modes (cosmetic, costly, catastrophic) live in the don't-list.

Diligence block is required output

  • Claude returns a source trace for every claim. The author can audit before sending.
  • The block doubles as the light gate — read it first, scan the rest second.

The prompt encodes the discipline that financial_modeling's fpa-storyteller agent runs at full Stage 4 scale. The agent loads preston-writing, dc-cap-org-intelligence, executive-summary-formatter, funder-framing, plus references for Knaflic, Damodaran, Bridgespan, and BoardSource. This single prompt compresses the same rules into one pasteable instruction the designer runs by hand. Same expertise, smaller container.

5. Data Tier — Tier 3 Inputs, Strategic Working Surface

Tier 3 · Internal Strategy / Operational

What enters the prompt: the headline numbers from a fresh fpa-quant build, the three-numbers-that-matter list, the named scenario assumptions. These are the same numbers that reach the Investment Committee preface.

What stays out: named donor specifics with expected close dates, individual gift sizes by donor, anything tagged Tier 2 in the financial_modeling governance config. The Big Idea sentence summarizes; it does not enumerate donors.

If the question would require Tier 2 input: the prompt is the wrong container. The full fpa-storyteller agent inside the orchestration handles that surface — it loads the donor-level references behind the audit-trail gates.

6. Light Verification Gate — Three Checks, Thirty Seconds

Stage 1 verification is the designer's own read. Three checks, every time.

  1. Diligence trace. Read the block first. Every claim should trace to the pasted input or be marked as generated framing.
  2. Voice check. Read the sentence aloud. Bare numbers up front, active voice, no forbidden constructions. The sentence should sound like Preston wrote it.
  3. 3P scan. Possible. Plausible. Probable. The sentence (or the alternate) clears all three; if one fails, swap the framing or rewrite.

For higher-stakes drafts headed to the IC or Board, the full fpa-storyteller path inside financial_modeling adds standard-and-full-gate verification on top. The Audit Picker maps gates to artifact type: Open the Audit Picker →

7. When to Promote — Signals This Prompt Has Outgrown Itself

The Prompt is the right container while one person owns the Big Idea draft for one recurring surface. The signals to promote — to Stage 2 (a Skill the cohort or another agent could load) — are concrete:

The full fpa-storyteller agent inside financial_modeling is what happens at the top of the ladder — Stage 4. The Skill version (Artifact B) is the next rung up from here.

Stage 2 next — Artifact B walks an actual Skill that financial_modeling loads

The Artifact B teach-page walks preston-writing — the real SKILL.md at ~/Desktop/BRAIN/skills/skills/preston-writing/. The same voice rules this prompt cites are encoded there. Walking the skill end to end shows what makes a Stage 2 build effective: the description that triggers it, the body that holds the discipline, and the references folder where the examples live.

Most pilot builds will stop at Stage 2 or Stage 3. Stage 1 trains writing the spec for yourself; Stage 2 trains writing the spec so other people and other agents can load it.