Story 1 of 2

Building the Storyteller

How to build an AI specialist that selects
rather than summarizes — and why
narrative cognition matters for knowledge work.

March 2026  ·  Design Case Study
Context — canvas-specialists

Six specialists. One pipeline.

canvas-specialists is a bundle of AI specialist agents designed for structured knowledge work. Each specialist has a formally specified output contract, a multi-stage pipeline, and documented guardrails. They chain together: one specialist’s output is the next specialist’s input. Machine-parseable contracts enable reliable handoffs.

Researcher
Deep web research with source tiering and per-claim confidence
Formatter
Normalizes researcher output to canonical format
Data Analyzer
Labeled inferences from research — pattern, causal, evaluative, predictive
Writer
Structured documents with full coverage and inline citations
Competitive Analysis
Comparison matrices, positioning gaps, and win conditions
Storyteller
The newest specialist.
This is the story of how it was built.
The Analytical Chain

The most common pipeline:

Researcher
Formatter
Analyzer
Writer

This chain produces rigorous, well-cited documents. Sourced findings. Labeled inferences. Full coverage. Inline citations tracing every claim.

But the documents read like what they are:
structured data transformed into prose.

Act 1 — The Gap

“There are two modes of cognitive functioning,
two modes of thought, each providing
distinctive ways of ordering experience.”

— Jerome Bruner, Actual Minds, Possible Worlds, 1986

Paradigmatic Mode

Logical. Categorical. Evidence-first. Evaluated by truth and falsehood. This is what the pipeline speaks.

Narrative Mode

Sequential. Causal. Meaning-making. Evaluated by verisimilitude and coherence. This is what audiences need.

You cannot get from one to the other by reformatting.
It requires structural transformation.

What Happens Without It

Two failure modes.
Neither is acceptable.

The analysis ships as-is

The audience receives a rigorous document in the wrong cognitive mode. They don’t act on it — not because the evidence is weak, but because the format doesn’t engage the system that drives decisions. Stakeholders, boards, customers need to be moved, not just informed.

Someone rewrites it informally

A human transforms the analysis into a narrative by hand. Traceability disappears. You can’t tell what was selected, what was left out, or why. The editorial judgment that shaped the story is invisible — and unreviewable.

For stakeholder communications, change narratives, and user research synthesis — where the audience needs to be moved, not just informed — there was a gap.

Act 2 — The Design Insight

“The Writer covers everything.
The Storyteller selects.”

Selection is the creative act. The Writer surfaces all findings, flags gaps, and produces structured documents with full coverage. The Storyteller is authorized to choose which findings are load-bearing for the narrative arc and which ones to set aside.

What you leave out shapes the narrative
as much as what you include.

Framework Selection

Three independent axes.
No user picks a framework directly.

Axis Signal Framework / Effect
Primary Goal Decision needed SCQA (answer-first, Minto)
Persuasion required Three-act or Sparkline (Duarte)
Insight to deliver Kishōtenketsu (twist-based)
Change to communicate Story Spine (causal chain)
Audience Expert, time-constrained Reinforces SCQA
Skeptical, needs convincing Reinforces three-act
General audience Emotional-first framing
Tone Trustworthy Conservative register, evidence-forward
Dramatic Amplified tension, urgency foregrounded
Creative Experimental sequencing, unexpected angles
Persuasive Loss aversion activated, audience-as-hero

Framework determines structure. Tone determines register. Any framework in any tone. A trustworthy SCQA and a dramatic SCQA are the same shape delivered in different voices.

The 6-Stage Pipeline

From raw findings to clean prose.

0
Open with STORY OUTPUT
Header at byte 0. Non-negotiable.
1
Parse
Detect input type, extract findings, infer audience & tone
2
Find the Drama
Dramatic question + 3-axis framework selection
3
Select and Structure
7 transformation decisions. NARRATIVE SELECTION built here.
4
Draft
Clean prose. No citations. No markdown headers.
5
Quality Gate
11-item checklist. Max 2 revision cycles. Fails loud.

Stage 3 is where the craft lives. The 7 transformation decisions — dramatic question, protagonist, gap, evidence selection, sequencing, register, peak moment — are all made before a single word of prose is written.

Selection as Craft

Every finding gets a role.
Or a rationale.

Included Findings

Each one earns a narrative role:

  • Hook Opens the story, establishes the situation
  • Complication Creates the gap or tension
  • Evidence Supports the causal chain
  • Peak The single most powerful moment
  • Resolution Answers the dramatic question
Omitted Findings

Each one gets an explicit rationale:

  • Off-arc Doesn’t serve the dramatic question
  • Redundant Covered by a stronger finding
  • Undermines tone Conflicts with the chosen register
  • Insufficient Too thin to carry story weight

Nothing is silently dropped. Every finding from the Analyzer is accounted for.

Act 3 — The Most Important Design Decision

The NARRATIVE SELECTION record.

An editorial audit trail that makes narrative judgment transparent and reviewable.

NARRATIVE SELECTION

——————————————

Dramatic question: Why are Fortune 500 companies generating impressive AI pilots they cannot finish?

Protagonist: Enterprise leadership

Framework: SCQA | Board audience + decision context

 

INCLUDED FINDINGS

F1: 78% Fortune 500 have AI in production | role: hook

F4: 60% of AI projects fail at pilot stage | role: complication

F5: Root cause is missing exec sponsorship | role: peak

F3: Dedicated AI teams ship 3x faster | role: resolution

 

OMITTED FINDINGS

F2: 18-month avg prototype-to-production | rationale: off-arc

A reviewer can see exactly what was selected, what was cut, and why — something human writers rarely document.

How Findings Flow

From 15 findings to 8 with roles.

15
Findings from the Analyzer
8
Included
Each with a narrative role:
hook, complication, evidence, peak, resolution
7
Omitted
Each with a rationale:
off-arc, redundant, undermines-tone, insufficient
Story prose uses only the selected 8. Clean narrative. No citations. No markdown headers. Evaluated by coherence, not coverage.
A Deliberate Choice

Zero inline citations in the story body.

Inline citations activate the analytical reading mode and suppress narrative transportation — the psychological state where readers lose themselves in a story.

Green & Brock (2000)

Narrative transportation mediates all narrative persuasion. Interruptions — including citation markers — break absorption and reduce impact.

Slovic et al. (2007)

Adding statistics to a story reduces its emotional impact. More analytical detail is not better narrative.

The NARRATIVE SELECTION

Provides full traceability without breaking the story. Auditability lives in the editorial record, not in the prose.

Act 4 — The Quality Gate

11 items. 4 categories.
Zero tolerance on format.

Structural (6 items)
  • Dramatic question present
  • Protagonist identified
  • Stakes explicit, not implied
  • Causal chain (not just sequence)
  • Single peak moment near climax
  • Resolution answers the dramatic question
Craft (3 items)
  • Concreteness — no abstract claims without specifics
  • Curse of Knowledge check
  • Audience-as-protagonist framing
Auditability (1 item)
  • Every upstream finding accounted for in INCLUDED or OMITTED
Format (1 item — zero tolerance)
  • STORY OUTPUT header at byte 0
Mechanical Fail Triggers

<3 findings extracted
>50% omitted as insufficient-evidence
Upstream quality NOT MET

Max 2 revision cycles, then NOT MET.

Act 5 — Verification
13/13

Contract items verified.

5/5
Pipeline Stages Execute in Order
3/3
Framework Axes Documented
5/5
Findings Accounted For
0
Inline Citations in Prose

Two tests: a rich AnalysisOutput (5 findings, board audience, trustworthy tone) and a boundary case (single low-confidence tertiary finding, upstream NOT MET). The first produced a clean SCQA narrative. The second correctly returned NOT MET with specific diagnostics.

The Boundary Case

It fails loud. Not silent.

Rich Input — 5 findings, board audience

4 findings included with roles. 1 omitted (off-arc) with documented rationale. SCQA framework selected via 3-axis logic. Clean prose, zero citations. QUALITY THRESHOLD RESULT: MET

Sparse Input — 1 tertiary finding, upstream NOT MET

Pipeline halted at Stage 2. No protagonist identifiable. No stakes establishable. No narrative fabricated. 6 specific structural failures documented. QUALITY THRESHOLD RESULT: NOT MET

The boundary case didn’t silently produce a weakened story and present it as complete. It halted, named the 6 missing structural elements, and refused to proceed. That’s the design working.

What Changed

The pipeline gained a new terminal node.

Researcher
Formatter
Analyzer
Writer

Analytical chain — comprehensive coverage, inline citations

Researcher
Formatter
Analyzer
Storyteller

Narrative chain — selective framing, editorial transparency

Same first three steps. Different final specialist depending on the deliverable — document or narrative. The Storyteller speaks Bruner’s other mode.

Sources & Design Lineage

Sources

Cognitive foundation:

Craft principles:

Artifacts:

Primary contributor: Chris Park (design direction, all decisions) with Amplifier AI (research synthesis, implementation, verification)

The Storyteller adds a capability
the pipeline didn’t have:

the ability to select, frame, and structure
findings for narrative impact rather than
comprehensive coverage.

But shipping is not the same
as proving it works in a chain.

That question requires evidence,
not architecture.

Next: “How Do You Know?” — The blind test, the statistical reckoning, and the five-word question that changed everything.

Story 2 of 2
Next: “How Do You Know?” → More Amplifier Stories