view-builder

View Builder

A topic comes in. A view comes out — one that is multi-lens, data-connected, uncertainty-honest, and tells the user what to notice going forward.

The skill's edge comes from three things most takes skip:

1. Exhaustive first-principles enumeration of lenses before any research. Don't let search results anchor the frame.
2. A persistent lens library that grows across every view, so you never re-discover a lens you've used before.
3. Parallel multi-agent research — one subagent per lens — so breadth is cheap and depth per lens stays high.

---

What a "view" is (and isn't)

A view is not a neutral encyclopedia summary, and it is not a single-source take. It is an opinionated but epistemically honest synthesis that:

• Identifies what the topic actually is and what goal the view serves for the user
• Runs the topic through many lenses — discovered via first-principles thinking plus a growing library plus meta-research for niche angles
• Connects each lens to real data where data exists, drawn from primary sources wherever possible
• States what parts of the topic are locally persistent (data-rich, testable) vs. humanistic (many variables, little persistence)
• Lands on a "this slightly edges out" recommendation without overclaiming
• Flags the interesting things to watch for — variables and frontier shifts other people miss
• Preserves uncertainty instead of compressing it away to sound confident

Underlying style: goal-driven reasoning under high uncertainty, across many domains, with a live model of a changing world. Bandwidth > pattern-matching.

---

The working loop — phase index

Twelve phases. Each phase lives in its own file under phases/. Read the phase file when you begin that phase; don't try to hold the whole protocol in your head at once. Phases back-reference each other by number, so if a phase says "as set up in Phase 3," find phases/phase-03-*.md.

Do not skip phases 1, 2, 4, 7, 9, and 10 — those are where the skill's quality comes from. Phase 9 is a three-track source-level analysis (fact-check, narrative corroboration, actor/incentive profiling) run in parallel; skipping it means shipping claims whose sources haven't been interrogated.

The loop is designed to be systematic and multi-agentic. When subagents are available (via the Task tool), Phase 6 (per-lens research) and Phase 9 (source analysis) MUST be parallelized. This is the whole point of the architecture: breadth without serial slowness.

#	Phase	File	One-line summary
1	Orient	phases/phase-01-orient.md	Dialogue with the user to nail down topic, goal, audience, depth, priors, option space. Push back on weak frames before running the machine.
2	First-principles brainstorm	phases/phase-02-brainstorm.md	Enumerate 20-30+ lenses from scratch. No research, no library — think first so search results don't anchor the frame.
3	Consult the lens library	phases/phase-03-lens-library.md	Add lenses from the seed library and the user's personal library that weren't in the first-principles brainstorm.
4	Meta-research	phases/phase-04-meta-research.md	2-4 targeted searches to surface niche / novel lenses the brainstorm missed. Lens discovery, not content research.
5	Prioritize with user	phases/phase-05-prioritize.md	Show the full candidate set, group by category, and prioritize 10-20 lenses to actively research.
6	Parallel per-lens research	phases/phase-06-parallel-research.md	One subagent per lens, all spawned in parallel. Chase primary sources, not high-level summaries.
7	Mid-research check-in	phases/phase-07-mid-check-in.md	Short dialogue surfacing surprises, conflicts, gaps, and new lenses before synthesis.
8	Synthesize the view	phases/phase-08-synthesize.md	Write the view artifact with strict inline citation discipline. Preserve uncertainty, don't paper over conflicts.
9	Source-level analysis	phases/phase-09-source-analysis.md	Three parallel tracks: fact-check (9a), narrative corroboration (9b), actor/incentive profiling (9c). Auto-fix and consolidate into a Source analysis block.
10	Metacognitive check	phases/phase-10-metacognitive.md	Self-audit for over-anchoring, tail-as-baseline, rounded uncertainty, missed frontier shifts, emotional bias, citation gaps.
11	Update the lens library	phases/phase-11-update-library.md	Append new lenses from this run to the user's persistent library, with hygiene rules.
12	Deliver	phases/phase-12-deliver.md	Share via computer:// link. Surface TL;DR, distinctive findings, source-analysis headline, and new library entries.

---

On style and tone

Action-enabling, not fence-sitting. Pick an edge even when uncertainty is high. But do not fake confidence — "X edges out, and I'd flip if Y" beats "X is the answer."

Avoid:

• Running only obvious lenses
• Letting search results anchor the lens set (that's what Phase 2 protects against)
• Treating tail examples as base rates
• Pattern-matching prior cases without checking the variables
• Frontier-blindness (stale model) and prior-stickiness (refusing to update)
• Citing a secondary summary when a primary source is one click away

Reach for:

• Bandwidth over pattern-matching
• Primary sources over high-level reads
• Specific data over vague trends
• Historical variables over historical analogies
• Niche / unusual lenses over canonical ones
• Library growth — every view should teach you something to reuse