/insight-angles

A lens-based, multi-pass discovery engine. Where /insight-explore asks "what are the ways to build this?" and /insight-adversarial attacks for flaws via personas, insight-angles asks "what frames, connections, assumptions, and second-order effects are not yet visible?" — and finds them by pointing the right lenses at the subject. It can run each lens as a cold sub-agent (fresh context, no anchoring) — independence is what makes discovery work, the same reason /cold-review outperforms inline /verify at finding things — and picks cold vs inline per subject, asking only when it isn't sure. Read-only and advisory — it surfaces, never decides or changes anything.

Structured execution spec: protocol.yon. Read it for the canonical rules and step sequence; this file is explanation. The two must stay in sync — if you edit one, update the other and refresh the @STAMP date.

The lens roster (seed, not ceiling)

17 families, grouped by what they perturb. Core = candidate for most subjects; Specialist = domain-triggered. Always also derive dynamic, subject-specific lenses (see below) — the roster is a seed.

8 core (Aesthetic/Elegance included — it's good practice across UX, DX, design, ergonomics) + 9 specialist. A run still fires only 4–6 — the roster's job is to make the picking rich.

Dynamic-lens derivation (the roster is a seed)

For every run, also derive subject-specific lenses from three named sources, then validate:

1. Domain-native — what would an expert in the subject's home discipline always ask here?
2. Foreign-transplant — what does a deliberately distant discipline ask that nobody here is?
3. Live-trend — from Classify's current-trends read, what recent shift / live debate reframes the space?
4. Validate — keep a candidate only if it generates a question the 17 families don't. Otherwise drop it as redundant.

Typed connections

Each carries type · the link · why it matters · what it enables: analogical · causal · structural · tension (productive paradox) · emergent (A+B→C) · dependency · resource · temporal.

Scoring — novelty × relevance

Every angle/connection is tagged: novelty (obvious / fresh / surprising, relative to current trends) · relevance (decision-relevant & actionable?) · so-what (what changes if true). Rank by novelty × relevance; the strongest unseen angle is the highest combined; flag high-novelty / low-relevance as "interesting but not actionable" rather than dropping it.

Steps

1. Frame — state the subject and the dominant framing to escape. Light context only.
2. Classify — domain · subject-type · maturity · current trends (calibrates novelty, drives selection).
3. Pick lenses — select roster families for the subject-type, plus derive dynamic lenses (above). Log picked + skipped (+ why).
4. Layered passes — run the picked lenses (see venue below):
   • Layer 1 — each lens → angles + typed connections, each scored.
   • Layer 2+ — add 2–4 lenses, informed by the prior layer (re-read prior angles through new lenses; link findings; revise scores). Deeper than default depth is HIL-gated.
5. Connect — consolidate typed connections across layers, especially cross-layer links.
6. Score & Synthesize — finalize scoring; emit the angle map + synthesis (strongest angle · key connection · most-fragile assumption). If nothing non-obvious surfaced, say so — never manufacture insight.

Venue & depth — context-adaptive

Cold fan-out is the discovery powerhouse (independence = the cold-review effect), but not every subject needs it. Select the venue case-by-case from context / corpus / task; proceed when certainty is near-max, ask the handler only when it isn't. Explicit overrides: "use agents" / "cold" forces fan-out; "inline" / "quick" forces inline.

• Cold (fan-out) when: high ambiguity, high stakes, novel subject, many lenses (anchoring risk), or independence clearly matters.
• Inline when: well-bounded, low stakes, few lenses, cost-constrained — or forced under a *-RESOLVED marker.

The three modes (selected adaptively, not by a fixed default):

• standard — cold fan-out: one fresh leaf sub-agent per picked lens. No shared context = no anchoring (the cold-review effect). Orchestrator collects + synthesizes. State the fan-out size up front; HIL-gate if it exceeds ~6 agents.
• quick — inline single pass (cheap). Forced when orchestrated under a *-RESOLVED marker (a lens-agent is a leaf — runs one lens, returns findings, spawns nothing — which keeps the depth guard intact).
• deep — cold fan-out across 2–3 layers; the orchestrator threads each completed layer's findings into the next layer's briefs (compounding at the orchestrator, independence in the agents). HIL-gated (cost).

Before dispatch, bind the worker and total invocation to finite deadlines and resource limits, with a named monitoring owner, observable progress, a cancellation handle and terminal success/failure/timeout outcomes. Use native controls; include read-only scope and no-child instructions in research briefs. On failure or timeout, cancel owned work and report retained partial results and unchecked coverage. If these controls are unavailable, do not dispatch. Use the existing bounded inline fallback when permitted, labeling it non-independent; empty or failed returns never count as completed coverage.

Cold lens-agents are leaf workers: each gets a primed brief (subject + its one lens [+ prior-layer findings in deep mode]), runs cold, returns its angles/connections, and orchestrates nothing further.

Runtime: detect capabilities, not product names. When the active runtime exposes isolated workers (for example, Claude's Agent or Codex collaboration agents when available), cold fan-out may be used. Without isolated workers, all modes fall back to inline — the adaptive selector degrades gracefully since inline is always available.

The angle map (output)

Selected lenses (+ dynamic ones + skipped & why) → angles grouped by lens, each scored → typed connections (each with why-it-matters) → hidden assumptions (most-fragile flagged) → candidate-expansions → second-order effects → synthesis → layer trail.

Rules

• MUST surface ≥1 genuine connection or unseen angle, or explicitly state none found — never an empty map, never manufactured insight.
• MUST derive dynamic subject-specific lenses (3 sources + validate), not only pick from the roster.
• MUST keep each layer informed by the prior layer's findings (layers compound; not independent re-runs).
• MUST select venue (cold vs inline) per subject from context/corpus/task; proceed when certainty is near-max, ask the handler only when it isn't; honor explicit overrides ("use agents" → cold, "inline" → inline).
• Cold lens-agents MUST be leaves — run one lens, return findings, spawn nothing.
• MUST stay in perspective/connection space: a frame is not a flaw (/insight-adversarial) and not an option (/insight-explore).
• MUST make each connection assessable (type + why), and judge novelty relative to current trends.
• MUST log lens selection (picked + skipped + why).
• MUST NOT exceed default depth (extra layers, or fan-out > ~6 agents) without explicit handler consent — state the cost and offer.
• MUST NOT make changes, write code, or execute — ideation only.
• Under a *-RESOLVED marker: run quick (inline) — never spawn further sub-skills.
• SHOULD prefer the non-obvious; flag high-novelty / low-relevance rather than dropping it.

Self-improvement → roster discovery

This skill's SIP is specialized: usage grows the roster. When a dynamic lens this run produced a high novelty × relevance angle the roster lacked, SIP proposes a diff adding it to the lens roster + selection map (gated via /ask-gate, never auto-applied). When a roster lens consistently yields nothing for a subject-type, SIP proposes pruning/remapping it. The 17 families are the seed; real subjects reveal the rest.

Boundary

• /insight-explore — generates solution options. insight-angles surfaces frames & connections; its candidate-expansions are seeds, not finished options.
• /insight-adversarial — attacks from personas for flaws. insight-angles widens via lenses (same multi-pass shape; personas attack, lenses reveal).
• /insight-critique — reviews a concrete output. insight-angles explores a subject's representation.
• /insight-cross-examine — the deliberation engine that orchestrates insight-angles (in quick mode), then assesses and recommends. Use cross-examine for a decision; insight-angles to just see more.
• /cold-review — outside review of work artifacts against objectives. insight-angles borrows its cold-agent independence, but for angle discovery on a subject, not artifact review.

Human output. This skill's handler-facing output obeys the human-output contract (human-output/SKILL.md).

Next skills. On completion, run the Next Skills protocol (next-skills/SKILL.md): surface the next-skills recommendations from front-matter for the caller to pick. Offer only — never auto-invoke.

Self-improvement. On completion, run the Self-Improvement Protocol (self-improve/SKILL.md): apply its signal taxonomy and threshold to this run, including a dynamic lens worth promoting into the roster; surface any surviving proposal for handler approval. Conservative — silent otherwise. Never auto-apply.