inspirer

Think outside the box, backed by evidence. 12 provocation lenses, web-grounded research, scored and filtered recommendations.

Contents

• Quick Start
• Architecture
• Stage 1: FRAME
• Stage 2: DIVERGE
• Stage 3: RESEARCH
• Stage 4: SCORE
• Stage 5: CONVERGE
• Stage 6: DELIVER
• Depth Control
• Evolve-Loop Integration
• Output Schema
• Reference

Quick Start

# Basic — think creatively about a topic
/evo:inspirer "How should we handle real-time sync in a serverless app?"

# Control depth (QUICK = fast ideation, DEEP = thorough exploration)
/evo:inspirer "What features increase user retention?" --depth DEEP

# Specify number of lenses (default: 4 for STANDARD)
/evo:inspirer "Multi-agent coordination patterns" --lenses 5

# Output for evolve-loop consumption
/evo:inspirer "Improve eval infrastructure" --format evolve --depth QUICK

Parse arguments:

• First quoted string or remaining text → topic
• --depth QUICK|STANDARD|DEEP → research depth (default: STANDARD)
• --lenses N → number of provocation lenses (default: per depth level)
• --format full|brief|evolve → output format (default: full)

Architecture

Six-stage pipeline: FRAME → DIVERGE → RESEARCH → SCORE → CONVERGE → DELIVER

Input: User's topic/question
         │
    ┌────▼────┐
    │  FRAME  │  Parse topic, classify domain, detect constraints
    └────┬────┘
         │
    ┌────▼─────┐
    │ DIVERGE  │  Apply 3-5 provocation lenses → divergent questions
    └────┬─────┘
         │ 1-2 research questions per lens
    ┌────▼─────┐
    │ RESEARCH │  Web search each question (Smart or Default routing)
    └────┬─────┘
         │ Research findings with sources
    ┌────▼────┐
    │  SCORE  │  Create Inspiration Cards, score feasibility × impact × novelty
    └────┬────┘
         │ Scored cards with KEEP/DROP verdicts
    ┌────▼─────┐
    │ CONVERGE │  Rank by composite, diversity filter, select top 5-8
    └────┬─────┘
         │
    ┌────▼─────┐
    │ DELIVER  │  Output as Inspiration Report / brief table / evolve JSON
    └──────────┘

Why this pipeline? Creative divergence without research produces brainstorming fluff. Research without creative framing produces obvious answers. This pipeline forces creative questions first, then grounds every idea in evidence.

Stage 1: FRAME

Parse the user's topic and establish context for lens selection.

Step	Action
1. Parse topic	Extract the core question or challenge
2. Classify domain	code-architecture, product-strategy, technical-research, process-improvement, general
3. Detect constraints	Implicit limits (technology, team size, timeline, budget) mentioned in the topic
4. Check evolve-loop context	If invoked from evolve-loop, read benchmarkWeaknesses and failedApproaches from context. Otherwise skip.

Output: Problem Frame object:

{
  "topic": "<parsed question>",
  "domain": "<classified domain>",
  "constraints": ["<constraint 1>", "<constraint 2>"],
  "evolveContext": null | { "weaknesses": [...], "failedApproaches": [...] }
}

Stage 2: DIVERGE

Apply provocation lenses to the Problem Frame. Each lens generates 1-2 divergent research questions that push thinking beyond the obvious.

Lens Selection

Depth	Lenses	Selection Method
QUICK	3	1 random + 2 domain-matched
STANDARD	4	1 random + 3 domain-matched
DEEP	5	1 random + 4 domain-matched

Domain-to-lens matching: See reference/provocation-lenses.md for the full affinity matrix. Each domain has 4-5 high-affinity lenses ranked by relevance.

The 12 Provocation Lenses

#	Lens	Provocation Question
1	Inversion	"What if we did the exact opposite?"
2	Analogy	"What would this look like borrowed from {adjacent domain}?"
3	10x Scale	"What breaks at 10x the current load/complexity?"
4	Removal	"What if we deleted this entirely?"
5	User-Adjacent	"What problem will the user hit NEXT?"
6	First Principles	"Why does this exist? What fundamental constraint requires it?"
7	Composition	"What if we combined two unrelated things?"
8	Failure Mode	"How would this fail silently?"
9	Ecosystem	"What external tool/pattern makes this obsolete?"
10	Time Travel	"What will we wish we had in 3 months?"
11	Constraint Flip	"What if the biggest constraint were removed entirely?"
12	Audience Shift	"What if the primary user were someone completely different?"

Lenses 1-10 are from the evolve-loop research protocol. Lenses 11-12 are added for general-purpose topics beyond codebase analysis.

Divergent Question Generation

For each selected lens:

1. Apply the provocation question to the Problem Frame
2. Generate 1-2 concrete, searchable research questions
3. Tag each question with lens and domain

Example: Topic = "How to handle real-time sync in serverless?"

• Inversion lens → "What architectures deliberately avoid real-time sync and still succeed?"
• Ecosystem lens → "What managed services handle real-time sync so we don't build it?"
• 10x Scale lens → "What real-time sync approaches handle 10M+ concurrent connections?"

Stage 3: RESEARCH

Ground every divergent question in web research. Route queries based on depth.

Depth	Routing	Max Queries	Max WebFetch
QUICK	Default WebSearch (1-2 queries per question)	5 total	2
STANDARD	Smart Web Search for complex, Default for simple	8 total	4
DEEP	Smart Web Search for all questions	12 total	6

Smart Web Search protocol: Use smart-web-search.md 6-stage pipeline (intent classification → query transformation → execution → evaluation → refinement → synthesis).

Default WebSearch: Direct 1-2 keyword queries with year filter for volatile topics.

For each research result, capture:

• Source URL and title
• Key finding (1-2 sentences)
• Relevance to the divergent question (0.0-1.0)
• Recency (penalize results > 2 years old for technology topics)

Critical rule: Ideas without at least 1 supporting research result are scored 0.0 on feasibility and auto-dropped in Stage 4. No research = no recommendation.

Stage 4: SCORE

Convert research-backed ideas into Inspiration Cards — extended Concept Cards with actionable detail.

Inspiration Card Schema

{
  "id": "insp-NNN",
  "title": "<concise idea title>",
  "oneLiner": "<1-sentence pitch — why this matters>",
  "lens": "<which provocation lens generated this>",
  "researchBacking": [
    {"source": "<URL>", "finding": "<key finding>", "relevance": 0.0}
  ],
  "implementationSketch": [
    "<step 1>", "<step 2>", "<step 3>"
  ],
  "risks": ["<risk 1>", "<risk 2>"],
  "nextSteps": ["<immediate action 1>", "<immediate action 2>"],
  "feasibility": 0.0,
  "impact": 0.0,
  "novelty": 0.0,
  "composite": 0.0,
  "verdict": "KEEP|DROP"
}

Scoring Rubric

Dimension	0.0-0.2	0.3-0.5	0.6-0.8	0.9-1.0
Feasibility	Requires tech that doesn't exist	Major unknowns, high risk	Achievable with known tech + moderate effort	Straightforward, proven patterns
Impact	Negligible improvement	Nice-to-have	Meaningful improvement to key metric	Transformative, 10x improvement
Novelty	Already standard practice	Minor twist on existing	Fresh combination of known ideas	Genuinely new approach

Composite: composite = (feasibility + impact + novelty) / 3

Verdict: composite >= 0.5 AND researchBacking.length >= 1 → KEEP. Otherwise → DROP.

See reference/scoring-rubric.md for detailed examples.

Stage 5: CONVERGE

Filter and rank KEPT cards into the final recommendation set.

Step	Action
1. Remove DROP cards	Only KEEP cards proceed
2. Diversity filter	Max 2 cards per lens (prevents one lens dominating)
3. Rank by composite	Highest composite first
4. Select top N	QUICK: top 3-5, STANDARD: top 5-8, DEEP: top 8-12
5. Cluster by theme	Group related cards (e.g., "scaling" cluster, "simplification" cluster)

Stage 6: DELIVER

Output in the requested format.

Format: full (default)

Human-readable Inspiration Report:

# Inspiration Report: <topic>

## Problem Frame
- **Domain:** <domain>
- **Constraints:** <constraints>
- **Lenses applied:** <lens1>, <lens2>, <lens3>

## Top Recommendations

### 1. <title> (composite: 0.XX)
> <oneLiner>

**Lens:** <which lens>
**Evidence:** <source> — "<key finding>"

**Implementation sketch:**
1. <step 1>
2. <step 2>
3. <step 3>

**Risks:** <risk 1>, <risk 2>
**Next steps:** <action 1>, <action 2>

---
### 2. <title> ...

## Research Sources
| # | Source | Finding | Used By |
|---|--------|---------|---------|

## Dropped Ideas (for transparency)
| Idea | Lens | Composite | Drop Reason |
|------|------|-----------|-------------|

Format: brief

Compact table:

| # | Idea | Lens | Composite | One-Liner | Next Step |
|---|------|------|-----------|-----------|-----------|

Format: evolve

JSON compatible with evolve-loop Scout task selection:

{
  "conceptCandidates": [
    {
      "id": "insp-001",
      "title": "...",
      "targetFiles": ["..."],
      "complexity": "S|M",
      "feasibility": 0.0,
      "impact": 0.0,
      "novelty": 0.0,
      "composite": 0.0,
      "source": "inspirer",
      "lens": "<lens-name>",
      "researchBacking": ["<capsule-ref>"]
    }
  ]
}

Depth Control

Depth	Lenses	Queries	Token Budget	Duration	Best For
QUICK	3	3-5	~20K	~30-60s	Fast ideation, time-constrained brainstorming
STANDARD	4	5-8	~40K	~2-3 min	Balanced creativity + research depth
DEEP	5	8-12	~60K	~4-6 min	Architecture decisions, strategy sessions, thorough exploration

Default: STANDARD

Evolve-Loop Integration

When invoked from within the evolve-loop pipeline, the inspirer provides enhanced creative divergence.

Phase 1 Delegation

The orchestrator can delegate to inspirer at Step 2.5 (DIVERGENCE TRIGGER):

Trigger conditions (ALL must be true):

• strategy == "innovate" OR discoveryVelocity.rolling3 < 0.5
• Budget is GREEN (not YELLOW/RED)
• Lean mode is NOT active
• Strategy is NOT repair or harden

Invocation (in-process): /evo:inspirer [goal] --depth QUICK --format evolve --lenses 3

Invocation (subprocess-isolated, REQUIRED in production cycles):

echo "/evo:inspirer $GOAL --depth QUICK --format evolve --lenses 3" | \
    bash legacy/scripts/dispatch/subagent-run.sh inspirer "$CYCLE" "$WORKSPACE_PATH"

The runner enforces the inspirer profile (.evolve/profiles/inspirer.json) which restricts writes to the inspirer-output artifact only and disallows state/ledger/profile mutation. WebSearch and WebFetch remain enabled for research-grounded ideation. Legacy fallback: LEGACY_AGENT_DISPATCH=1 for one A/B cycle.

Result: Returned concept cards merge with standard gap-analysis cards and flow to Scout with +2 priority boost (same as research-backed concepts).

Standalone Use

Outside evolve-loop, the skill requires no pipeline infrastructure. It uses WebSearch and WebFetch tools directly.

Reference (read on demand)

File	When to Read
reference/provocation-lenses.md	Deep lens descriptions, domain affinity matrix, examples
reference/scoring-rubric.md	Detailed scoring criteria, card schema, worked scoring examples
reference/worked-examples.md	3 end-to-end pipeline examples