trend-scout

Trend Scout

Interactive workflow for scouting trends across 4 dimensions with industry selection and bilingual support. Each trend discovered is later analyzed through the complete TIPS framework (Trend → Implications → Possibilities → Solutions). Produces configuration files for downstream value-modeler and trend-report skills.

Purpose

This skill enables users to:

1. Select an industry and subsector from a standardized taxonomy
2. Initialize a research project with semantic slug
3. Generate 60 trend candidates (5 per cell × 12 cells: 4 dimensions × 3 horizons)
4. Write the final trend list and produce configuration for downstream pipeline skills (value-modeler, trend-report)

Language Support

Full German and English support throughout. This skill follows the shared language resolution pattern — see $CLAUDE_PLUGIN_ROOT/references/language-resolution.md.

Two language concepts:

1. Interaction language — how the skill communicates with the user (prompts, status, questions). Determined by workspace .workspace-config.json language setting. All AskUserQuestion prompts, status messages, and instructions use this language.
2. Output language — what language deliverables are written in. Asked explicitly in Phase 0 with workspace language as default. Stored as project_language.

• Industry taxonomy presented in both languages
• Web research queries in both languages (global + DACH regions)
• User-facing messages in interaction language
• Output files respect project_language setting

Context Independence

This skill reads configuration from project files and generates all outputs to disk — it does not depend on prior conversation context. If invoked after trends-resume or other conversational setup, context compaction is safe and recommended before starting.

Before executing Phase 0, run /compact to free working memory. This skill dispatches a web research agent with 32+ searches (Phase 1) and generates 60 scored candidates with extended thinking (Phase 2) — both require substantial context for processing research signals and candidate scoring. Compacting early maximizes the context available for these heavy phases.

If /compact is unavailable or this is the first skill in the session (no prior context to reclaim), skip compaction and proceed directly.

Prerequisites

• Projects are stored relative to the workspace root ($PROJECT_AGENTS_OPS_ROOT, falling back to $PWD)
• Web access enabled for live trend research

Shell Execution Constraints

CRITICAL - Do NOT improvise shell commands:

1. All project initialization MUST use the provided scripts
2. NEVER generate inline bash code for slug generation or project creation
3. If a script is not found, report the error and ask user to verify installation
4. Do NOT attempt workarounds with inline $(...) command substitution

Path Variable Distinction:

Variable	Purpose	Example
CLAUDE_PLUGIN_ROOT	Plugin installation (scripts, skills)	~/.claude/plugins/marketplaces/cogni-trends
PROJECT_AGENTS_OPS_ROOT	Workspace root where projects live (optional, set by cogni-workspace)	User's workspace directory

IMPORTANT - Environment Variables:

• CLAUDE_PLUGIN_ROOT is automatically injected by Claude Code from settings.local.json
• PROJECT_AGENTS_OPS_ROOT is set by cogni-workspace's generate-settings.sh — if not present, scripts fall back to $PWD
• DO NOT source .workplace-env.sh - variables are already available at runtime

Script Locations (always use CLAUDE_PLUGIN_ROOT):

• ${CLAUDE_PLUGIN_ROOT}/skills/trend-scout/scripts/generate-project-slug.sh
• ${CLAUDE_PLUGIN_ROOT}/skills/trend-scout/scripts/update-industry-metadata.sh
• ${CLAUDE_PLUGIN_ROOT}/skills/trend-scout/scripts/finalize-candidates.sh
• ${CLAUDE_PLUGIN_ROOT}/scripts/discover-portfolio-markets.sh
• ${CLAUDE_PLUGIN_ROOT}/scripts/initialize-trend-project.sh

References Index

Read references only when needed for the specific task:

Reference	Read when...
$CLAUDE_PLUGIN_ROOT/references/data-model.md	Understanding entity schemas and project structure
references/industry-taxonomy.md	Presenting industry selection to user
$CLAUDE_PLUGIN_ROOT/references/language-resolution.md	Language detection and resolution pattern
references/i18n/messages-en.md	English user messages
references/i18n/messages-de.md	German user messages
references/methodology.md	Academic foundations (Ansoff, Rohrbeck, Rogers), full methodology explanation
references/dach-sources.md	DACH site-specific web searches (Phase 1)
references/funding-signals.md	VC/funding signal queries (Phase 1)
references/job-market-signals.md	Job market signal queries (Phase 1)
references/academic-api-queries.md	Academic API searches - OpenAlex, Semantic Scholar, arXiv (Phase 1)
references/patent-api-queries.md	Patent API searches - USPTO, Lens.org, EPO (Phase 1)
references/regulatory-feeds.md	Regulatory API searches - EUR-Lex, SEC EDGAR, FDA (Phase 1)
references/workflow-phases/phase-0-initialize.md	Project init + industry selection
$CLAUDE_PLUGIN_ROOT/references/dimension-personas.md	Persona catalog for dimension-specific research (Phase 1, Sprint 2)
references/workflow-phases/phase-2.5-review.md	Candidate review: stakeholder assessment, repair protocol, re-review
references/workflow-phases/phase-3-present.md	Writing final trend-candidates.md with scores
references/workflow-phases/phase-4-finalize.md	Finalizing output for downstream pipeline

Immediate Action: Initialize TodoWrite

MANDATORY: Initialize TodoWrite immediately with workflow phases:

1. Phase 0: Initialize Project + Industry Selection [in_progress]
2. Phase 0.5: Configuration Disclosure + Preliminary Grounding [pending]
3. Phase 1: Bilingual Web Research [pending]
4. Phase 1.5: Signal Curation (if thorough mode) [pending]
5. Phase 2: Generate Candidate Pool [pending]
6. Phase 2.5: Candidate Review (Stakeholder Assessment) [pending]
7. Phase 3: Write Final Trend List [pending]
8. Phase 4: Finalize Output + Pipeline Config [pending]

Update todo status as you progress through each phase.

---

Core Workflow

Phase 0 → Phase 0.5 → Phase 1 → Phase 1.5 → Phase 2 → Phase 2.5 → Phase 3 → Phase 4
   │          │           │          │           │          │          │         │
   │          │           │          │           │          │          │         └─ Write config + JSON, finalize
   │          │           │          │           │          │          └─ Write final trend-candidates.md
   │          │           │          │           │          └─ Stakeholder review + repair loop (max 2 iter)
   │          │           │          │           └─ Generate + score 60 candidates
   │          │           │          └─ Signal curation (thorough mode)
   │          │           └─ Web searches + academic/patent/regulatory APIs
   │          └─ Config disclosure + 3 grounding searches
   └─ Language detect, industry select, project init

Phase 0: Initialize Project + Industry Selection

Read references/workflow-phases/phase-0-initialize.md and $CLAUDE_PLUGIN_ROOT/references/language-resolution.md, then execute:

1. Detect interaction language: Read workspace language from .workspace-config.json (via ${PROJECT_AGENTS_OPS_ROOT}/.workspace-config.json or CWD). Set INTERACTION_LANGUAGE — use this for all user-facing messages from this point on. Load the matching i18n message catalog (messages-{INTERACTION_LANGUAGE}.md).
2. Ask user for output language: Present AskUserQuestion in the interaction language. Workspace language is the pre-selected default (e.g., "Deutsch (DE) <- Workspace-Standard" or "English (EN) <- Workspace default"). User can override. Set PROJECT_LANGUAGE from explicit choice. Do NOT skip asking — always confirm with user.
3. Portfolio discovery (optional): Scan workspace for cogni-portfolio projects with markets. If the primary workspace ($PROJECT_AGENTS_OPS_ROOT or $PWD) has no portfolio projects, perform a broader scan — check parent directories and common cloud storage locations (~/Library/CloudStorage, ~/OneDrive, ~/Documents). If still nothing found, ask the user if they have a workspace directory to scan. If portfolio found, offer user to pre-populate industry/subsector from a portfolio market. If selected, skip steps 4-6 and suggest a research topic from the market context. See Step 0.1c in references/workflow-phases/phase-0-initialize.md.
4. Load references/industry-taxonomy.md (skip if portfolio market selected)
5. Present industries with subsectors (bilingual) (skip if portfolio market selected)
6. Capture user selection via AskUserQuestion (in interaction language) (skip if portfolio market selected)
7. Capture research topic/focus (in interaction language) — with optional suggestion from portfolio market
8. Generate project slug: {subsector}-{topic}-{hash}
9. Initialize project via initialize-trend-project.sh in the current working directory under cogni-trends/
10. Update tips-project.json with full industry context (bilingual names, subsector, research_topic) — see Step 0.8b in phase-0-initialize.md. The update-industry-metadata.sh script only updates .metadata/trend-scout-output.json, so you MUST also update tips-project.json inline with jq (industry.primary, primary_en, primary_de, subsector, subsector_en, subsector_de, research_topic).
11. Update .metadata/trend-scout-output.json with industry context (and portfolio_source if applicable)

Required outputs:

• PROJECT_PATH, PROJECT_SLUG variables set
• PROJECT_LANGUAGE set from explicit user choice (de/en)
• INDUSTRY, SUBSECTOR selected
• RESEARCH_TOPIC captured
• Project structure initialized in current working directory under cogni-trends/

Phase 0.5: Configuration Disclosure + Preliminary Grounding

This phase serves two purposes: (1) show the user what research options are available before committing to expensive web research, and (2) perform 3 quick grounding searches to anchor subsequent query formulation in what the web actually contains.

Step 1: Configuration Disclosure

Present research configuration options via AskUserQuestion before any web research begins. This makes capabilities discoverable and lets users make informed cost/quality tradeoffs.

Use the interaction language for the prompt. Present these options:

EN: "Before starting research, please confirm your preferences:"
DE: "Bevor die Recherche startet, bestätigen Sie bitte Ihre Einstellungen:"

Options:
1. Research depth:
   a) Standard — ~32 web searches, fastest (default)
   b) Thorough — adaptive budget (~36-48 searches), better signal coverage per dimension
2. Preliminary grounding:
   a) Enabled — 3 broad searches to calibrate research queries (default, recommended)
   b) Skip — jump directly to full research
3. Confirm and start research

Store selections in tips-project.json under a research_config key:

{
  "research_config": {
    "depth": "standard|thorough",
    "grounding": true|false
  }
}

If the user selects defaults or just says "go" / "start" / "los", use: depth: "standard", grounding: true.

Step 2: Preliminary Grounding (if grounding enabled)

Execute 3 broad exploratory WebSearch queries inline (NOT delegated to agent). These ground subsequent Phase 1 query formulation in what the web actually contains about this subsector + topic.

The reason this matters: fixed query templates don't know what's dominating discourse for a given subsector. If the topic is "AI in healthcare" and the web is dominated by FDA regulation news, the current fixed queries miss this. Grounding surfaces dominant themes so Phase 1 queries can incorporate them.

Grounding searches:

1. "{SUBSECTOR_EN} {RESEARCH_TOPIC} trends challenges {CURRENT_YEAR}" (broad EN scan)
2. "{SUBSECTOR_DE} {RESEARCH_TOPIC} Herausforderungen Chancen {CURRENT_YEAR}" (DACH scan)
3. "{SUBSECTOR_EN} {RESEARCH_TOPIC} market outlook disruption" (future-oriented)

Derive {CURRENT_YEAR} from the system date (same pattern as web-researcher Step 0).

Process grounding results:

From the 3 search results, extract a grounding summary (~200 words) capturing:

• Dominant themes — what topics appear most frequently across results?
• Key organizations — which institutions, companies, or regulators are mentioned?
• Recent developments — what specific events, regulations, or product launches are current?
• Terminology — what specific terms or buzzwords appear that the generic query templates wouldn't use?

Write the grounding context to {PROJECT_PATH}/.metadata/preliminary-grounding.json:

{
  "timestamp": "ISO-8601",
  "searches_executed": 3,
  "grounding_summary": "~200 word summary of dominant themes, key organizations, recent developments, and terminology",
  "dominant_themes": ["theme1", "theme2", "theme3"],
  "key_organizations": ["org1", "org2"],
  "terminology_hints": ["term1", "term2", "term3"]
}

Set GROUNDING_CONTEXT variable to the grounding_summary string for passing to the web-researcher agent in Phase 1.

If grounding is disabled (user chose "skip"), set GROUNDING_CONTEXT = "" and skip the 3 searches.

Required outputs:

• Research configuration stored in tips-project.json
• GROUNDING_CONTEXT variable set (empty string if grounding skipped)
• .metadata/preliminary-grounding.json written (if grounding enabled)

Phase 1: Bilingual Web Research + API Queries (DELEGATED)

Context Efficiency: This phase is delegated to the web-researcher agent to prevent context depletion from 20+ WebSearch results. The agent returns a compact JSON summary (~500 tokens) while logging full results to .logs/.

Invoke the web-researcher agent:

Task:
  subagent_type: "cogni-trends:trend-web-researcher"
  description: "Execute bilingual web research"
  prompt: |
    Execute Phase 1 web research for trend-scout.

    PROJECT_PATH: {{PROJECT_PATH}}
    INDUSTRY_EN: {{INDUSTRY_EN}}
    INDUSTRY_DE: {{INDUSTRY_DE}}
    SUBSECTOR_EN: {{SUBSECTOR_EN}}
    SUBSECTOR_DE: {{SUBSECTOR_DE}}
    RESEARCH_TOPIC: {{RESEARCH_TOPIC}}
    MARKET_REGION: {{MARKET_REGION}}
    GROUNDING_CONTEXT: {{GROUNDING_CONTEXT}}
    RESEARCH_DEPTH: {{RESEARCH_DEPTH}}

Agent responsibilities:

1. Build 32 web search configurations (16 standard + 8 DACH site-specific + 4 funding + 4 job market)
2. Execute WebSearch for each config in parallel batches
3. Execute mandatory API queries (academic, patent, regulatory) with fallback handling
4. Extract and deduplicate trend signals
5. Classify signals by indicator type (leading/lagging) and source type
6. Write full results to {{PROJECT_PATH}}/.logs/web-research-raw.json
7. Return compact JSON with ~85 aggregated signals

Note: The web-researcher agent is self-contained with all search configurations and deduplication logic.

Process agent response:

The agent returns compact JSON with abbreviated fields for token efficiency:

{
  "ok": true,
  "signals": {
    "total": 85,
    "by_dimension": {...},
    "by_source": {"web": 48, "dach_site": 12, "funding": 8, "jobs": 6, "academic": 5, "patent": 4, "regulatory": 2},
    "by_indicator": {"leading": 38, "lagging": 47}
  },
  "items": [{"d": "dimension", "n": "name", "k": ["keywords"], "u": "url", "f": "freshness", "a": 5, "t": "type", "i": "leading", "lt": "12-24m"}]
}

Log file format (.logs/web-research-raw.json):

The log file uses full field names for debugging readability. Key structure:

{
  "metadata": {...},
  "searches_executed": {"total": 32, "successful": 30, "failed": 2, "by_category": {...}},
  "raw_signals_before_dedup": [
    {"dimension": "...", "signal": "...", "keywords": [...], "source": "url", "freshness": "...", "indicator_type": "leading|lagging", "lead_time": "...", "source_type": "..."}
  ],
  "api_queries_executed": {...}
}

To query the log file directly:

jq '.raw_signals_before_dedup[] | {dimension, signal, keywords, source}' .logs/web-research-raw.json

Set availability flag:

• Set WEB_RESEARCH_AVAILABLE = (response.ok == true)
• Do NOT expand or group signals — the trend-generator agent self-loads them from disk

Persist compact response for downstream fallback:

Write the agent's raw compact JSON response (the full response object including the .items array) to: {PROJECT_PATH}/phase1-research-summary.json

This file serves as a fallback for trend-report when .logs/web-research-raw.json is missing or incomplete.

Required outputs:

• WEB_RESEARCH_AVAILABLE flag set
• Signal data persisted to disk (agent will self-load from .logs/web-research-raw.json or phase1-research-summary.json)

Fallback Hierarchy:

1. Agent returns {"ok": false} — proceed to inline fallback research (below)
2. Agent unavailable (subagent dispatch fails) — proceed to inline fallback research (below)
3. Inline fallback research also fails — proceed with training-only generation (warning logged)

Inline Fallback Research (when web-researcher agent is unavailable):

If the web-researcher agent cannot be dispatched (e.g., nested subagent context, agent not found), perform a reduced set of web searches directly using WebSearch. This is less thorough than the agent's 32 searches but ensures candidates have some web grounding rather than being 100% training-only.

Execute these 12 targeted searches organized by source authority tier. The first 6 target authoritative institutional sources (CRAAP authority 4-5) to ensure the candidate pool has credible grounding. The remaining 6 broaden coverage.

Tier 1 — Authoritative institutional sources (run these first):

1. "site:fraunhofer.de {SUBSECTOR_DE} {RESEARCH_TOPIC} Studie 2025"
2. "site:ec.europa.eu OR site:eur-lex.europa.eu {SUBSECTOR_EN} {RESEARCH_TOPIC} regulation"
3. "site:bitkom.org OR site:{ASSOCIATION_DOMAIN} {SUBSECTOR_DE} {RESEARCH_TOPIC} 2025"
4. "{SUBSECTOR_EN} {RESEARCH_TOPIC} site:gartner.com OR site:mckinsey.com OR site:rolandberger.com"
5. "site:destatis.de OR site:bmwk.de {SUBSECTOR_DE} {RESEARCH_TOPIC} Statistik"
6. "{SUBSECTOR_EN} {RESEARCH_TOPIC} arxiv.org OR ieee.org OR sciencedirect.com 2024 2025"

For search 3, replace {ASSOCIATION_DOMAIN} with the subsector's primary industry association from references/dach-sources.md (e.g., vda.de for automotive, bvmed.de for healthcare).

Tier 2 — Broader market and signal sources:

7. "{SUBSECTOR_EN} trends {RESEARCH_TOPIC} 2025 2026"
8. "{SUBSECTOR_DE} {RESEARCH_TOPIC} Markt DACH Mittelstand"
9. "{SUBSECTOR_EN} {RESEARCH_TOPIC} market outlook DACH"
10. "{SUBSECTOR_DE} Digitalisierung {RESEARCH_TOPIC} Trend"
11. "{SUBSECTOR_EN} {RESEARCH_TOPIC} funding investment startups DACH"
12. "{SUBSECTOR_EN} {RESEARCH_TOPIC} patent filing 2024 2025"

For each search result, extract trend signals (name, keywords, source URL, freshness). Write the aggregated signals to {PROJECT_PATH}/.logs/web-research-raw.json in the same format the agent would produce, and to {PROJECT_PATH}/phase1-research-summary.json as compact fallback. Set WEB_RESEARCH_AVAILABLE = true.

Source tagging: When extracting signals, tag each with its source authority level based on domain:

• Authority 5: .gov, .eu, fraunhofer.de, ieee.org, arxiv.org, nature.com
• Authority 4: gartner.com, mckinsey.com, rolandberger.com, industry associations (.org)
• Authority 3: handelsblatt.com, reuters.com, industry trade publications
• Authority 2-1: commercial blogs, vendor sites, social media

This tagging flows into the trend-generator's CRAAP scoring — candidates grounded in authority 4-5 sources will score higher on the 15% Source Quality weight.

Phase 1.5: Signal Curation (OPTIONAL, DELEGATED)

When to run: Signal curation activates when the web research returned 20+ signals AND research depth is "thorough". Skip in standard mode or when signals are sparse (< 20).

Purpose: Rank the ~85 raw signals from Phase 1 into quality tiers (primary/secondary/supporting) before the trend-generator consumes them. This ensures the generator grounds its best candidates in the highest-quality signals rather than treating all signals equally.

Invoke the signal curator agent:

Task:
  subagent_type: "cogni-trends:trend-signal-curator"
  description: "Curate and rank web research signals"
  prompt: |
    Evaluate and rank Phase 1 web research signals for trend-scout.

    PROJECT_PATH: {{PROJECT_PATH}}
    RESEARCH_TOPIC: {{RESEARCH_TOPIC}}
    SUBSECTOR_EN: {{SUBSECTOR_EN}}

Process agent response:

The agent returns compact JSON:

{
  "ok": true,
  "total": 85,
  "tiers": {"primary": 25, "secondary": 40, "supporting": 20},
  "by_dimension": {"externe-effekte": 22, "neue-horizonte": 21, "digitale-wertetreiber": 20, "digitales-fundament": 22},
  "diversity_warnings": 0,
  "dimension_gaps": []
}

Set availability flag:

• Set CURATED_SIGNALS_AVAILABLE = (response.ok == true)

Adaptive follow-up (thorough mode only): If dimension_gaps is non-empty (dimensions with < 10 signals), execute 2-3 additional targeted WebSearch queries for each gap dimension using persona vocabulary. Write results to the raw signals file and re-run curation. This is a single retry — do not loop.

Fallback: If the agent fails or is unavailable, set CURATED_SIGNALS_AVAILABLE = false and proceed — the trend-generator will fall back to reading raw signals directly.

Phase 2: Generate Candidate Pool (DELEGATED)

Context Efficiency: This phase is delegated to the trend-generator agent to leverage Opus model's extended thinking for complex multi-framework reasoning. The agent returns a compact JSON summary (~600 tokens) while logging full candidate data to .logs/.

Invoke the trend-generator agent:

Task:
  subagent_type: "cogni-trends:trend-generator"
  description: "Generate 60 scored trend candidates"
  prompt: |
    Execute Phase 2 candidate generation for trend-scout.

    PROJECT_PATH: {{PROJECT_PATH}}
    INDUSTRY_EN: {{INDUSTRY_EN}}
    INDUSTRY_DE: {{INDUSTRY_DE}}
    SUBSECTOR_EN: {{SUBSECTOR_EN}}
    SUBSECTOR_DE: {{SUBSECTOR_DE}}
    RESEARCH_TOPIC: {{RESEARCH_TOPIC}}
    PROJECT_LANGUAGE: {{PROJECT_LANGUAGE}}
    WEB_RESEARCH_AVAILABLE: {{WEB_RESEARCH_AVAILABLE}}

Agent responsibilities:

1. Apply embedded scoring framework
2. Generate 60 trend candidates (5 per cell × 12 cells) using extended thinking (MANDATORY)
3. Apply multi-framework scoring (TIPS, Ansoff, Rogers, CRAAP)
4. Classify indicator types (leading/lagging) and diffusion stages
5. Validate subcategory balance (MIN 1 per subcategory per cell)
6. Validate portfolio balance (≥40% leading indicators)
7. Write full results to {{PROJECT_PATH}}/.logs/trend-generator-candidates.json
8. Return compact JSON summary

Process agent response:

The agent returns compact JSON:

{
  "ok": true,
  "candidates": {"total": 60, "by_source": {...}, "by_dimension": {...}},
  "scoring": {"avg_score": 0.65, "confidence": {...}, "indicator": {...}},
  "validation": {"passed": true, "warnings": []},
  "log": ".logs/trend-generator-candidates.json"
}

Prepare Phase 3 data files:

Execute data preparation script to generate compact candidate data:

bash "${CLAUDE_PLUGIN_ROOT}/skills/trend-scout/scripts/prepare-phase3-data.sh" "${PROJECT_PATH}"

This generates:

• .logs/candidates-compact.json (compact format for Claude reading)

Load compact candidate data:

Read {{PROJECT_PATH}}/.logs/candidates-compact.json to build trend-candidates.md.

Field mapping for compact format:

• d → dimension, h → horizon, n → name
• s → trend_statement, r → research_hint, k → keywords
• sc → score, ct → confidence_tier, si → signal_intensity
• src → source, url → source_url

Required outputs:

• CANDIDATES_BY_CELL loaded from log file
• TOTAL_CANDIDATES = 60
• SCORING_METADATA populated from agent response
• Validation status confirmed

Fallback Hierarchy:

1. Agent returns {"ok": false} — log error and halt workflow
2. Agent unavailable (subagent dispatch fails) — perform inline candidate generation (below)

Inline Fallback Generation (when trend-generator agent is unavailable):

If the trend-generator agent cannot be dispatched, generate the 60 candidates inline. This loses the benefit of extended thinking in a separate context, but still produces the required output.

Steps:

1. Load web research signals from {PROJECT_PATH}/.logs/web-research-raw.json or {PROJECT_PATH}/phase1-research-summary.json
2. Prioritize web signals for candidate creation: For each cell, first check if web signals exist for that dimension. Create candidates grounded in web signals first (mark as source: "web-signal" with the original URL), then fill remaining slots with training knowledge. Target: at least 50% of candidates should be web-sourced when signals are available.
3. Generate 60 candidates (5 per cell x 12 cells) following the same dimension/horizon/subcategory structure
4. Apply the scoring weights from references/scoring-framework.md — especially the training source caps (source_quality max 0.4, signal_strength max 0.3 for training-only candidates)
5. Validate subcategory balance: each cell must have MIN 1 candidate per subcategory. If violated, replace the lowest-scored candidate in the over-represented subcategory
6. Validate and repair horizon-intensity alignment: ACT candidates must have signal_intensity 4-5 (if < 4, set to 4). OBSERVE candidates must have intensity 1-2 (if > 2, set to 2). PLAN candidates: clamp to [2, 4]. This is a core Ansoff methodology constraint — a trend in the "act now" horizon must show strong signals, and a long-horizon OBSERVE trend must show weak/emerging signals.
7. Write results to {PROJECT_PATH}/.logs/trend-generator-candidates.json
8. Run prepare-phase3-data.sh to generate compact format

Important: Even in inline mode, enforce the scoring caps for training-sourced candidates. A training-only candidate with score: 0.78 signals a scoring cap violation — the theoretical max for a pure training candidate is ~0.60 after caps are applied.

Phase 2.5: Candidate Review — Stakeholder Assessment

Read references/workflow-phases/phase-2.5-review.md, then execute:

This phase evaluates the 60 candidates as a pool from three stakeholder perspectives before writing the final list. It catches set-level issues that per-candidate validation misses: duplicates across dimensions, subsector-generic filler, weak clustering, and scoring violations.

Three perspectives:

• Strategic Foresight Analyst — methodological soundness (horizon balance, signal grounding, leading indicators, diffusion spread, scoring integrity)
• Industry Domain Expert — subsector relevance (specificity, coherence, distinctiveness, DACH relevance, research hint quality)
• Downstream Pipeline Consumer — fitness for value-modeler/trend-report (TIPS expandability, theme potential, evidence enrichability, solution readiness, cross-dimension linkage)

Workflow:

1. Invoke trend-candidate-reviewer agent with iteration 1
2. Process verdict:
   • accept → proceed to Phase 3
   • reject → re-invoke full trend-generator, then re-review as iteration 2
   • revise → execute selective repair (cell regeneration, candidate replacement, scoring fixes)
3. After repair, re-invoke reviewer with iteration 2
4. If still not accepted after iteration 2 → force accept with issues logged

Max 2 review iterations. See phase reference for invocation templates and repair protocol.

Required outputs:

• .metadata/candidate-review-verdicts/v{N}.json — review verdict(s)
• Updated .logs/trend-generator-candidates.json (if repairs applied)
• Updated .logs/candidates-compact.json (regenerated after repairs)
• candidate_review metadata in execution state

Phase 3: Write Final Trend List

Read references/workflow-phases/phase-3-present.md, then execute:

Entry gate: Phase 2.5 must have completed with a review verdict of "accept" (clean or forced). Check that .metadata/candidate-review-verdicts/ contains at least one verdict file with verdict: "accept".

1. Write trend-candidates.md to {PROJECT_PATH}/ (project root) as the final trend list
2. Use bilingual template based on PROJECT_LANGUAGE
3. Include all 60 candidates organized by dimension and horizon with scores and metadata
4. Include source integrity summary and references

All 60 reviewed candidates are the final agreed list — no user selection step. Proceed directly to Phase 4.

Phase 4: Finalize Output

Read references/workflow-phases/phase-4-finalize.md, then execute:

1. Update consolidated trend-scout-output.json with all 60 candidates
2. Update tips-project.json with current timestamp (updated field)
3. Update trend-candidates.md frontmatter status to agreed
4. Log completion with next-step instructions
5. Recommend /trends-resume for the next session

Required outputs:

• .metadata/trend-scout-output.json - consolidated output (config + candidates)
• tips-project.json - updated timestamp
• trend-candidates.md status updated to agreed

---

Output Schema

trend-scout-output.json (Consolidated)

Location: {PROJECT_PATH}/.metadata/trend-scout-output.json

{
  "version": "1.0.0",
  "project_id": "automotive-ai-predictive-maintenance-abc12345",
  "project_name": "automotive-ai-predictive-maintenance-abc12345",
  "project_path": "/path/to/project",
  "project_language": "de",
  "created": "2025-12-16T10:30:00Z",

  "config": {
    "research_type": "smarter-service",
    "dok_level": 4,
    "industry": {
      "primary": "manufacturing",
      "primary_en": "Manufacturing",
      "primary_de": "Fertigung",
      "subsector": "automotive",
      "subsector_en": "Automotive",
      "subsector_de": "Automobil"
    },
    "research_topic": "AI-driven predictive maintenance",
    "organizing_concept": "ai-driven-predictive-maintenance"
  },

  "tips_candidates": {
    "total": 60,
    "source_distribution": {
      "web_signal": 28,
      "training": 32,
    },
    "web_research_status": "success",
    "search_timestamp": "2025-12-16T10:25:00Z",
    "scoring_metadata": {
      "avg_score": 0.68,
      "confidence_distribution": {
        "high": 12,
        "medium": 18,
        "low": 5,
        "uncertain": 1
      },
      "intensity_distribution": {
        "level_1": 4,
        "level_2": 6,
        "level_3": 10,
        "level_4": 12,
        "level_5": 4
      },
      "indicator_distribution": {
        "leading": 16,
        "lagging": 20,
        "leading_pct": 0.44
      },
      "diffusion_distribution": {
        "innovators": 3,
        "early_adopters": 8,
        "early_majority": 15,
        "late_majority": 8,
        "laggards": 2,
        "pre_chasm": 11,
        "post_chasm": 25
      },
      "scoring_framework_version": "1.0.0"
    },
    "source_integrity": {
      "training_capped": true,
      "training_with_corroboration": 8,
      "training_without_corroboration": 24,
      "avg_training_score": 0.48,
      "avg_web_signal_score": 0.72
    },
    "items": [
      {
        "dimension": "externe-effekte",
        "dimension_de": "Externe Effekte",
        "subcategory": "regulierung",
        "subcategory_en": "Regulation",
        "subcategory_de": "Regulierung",
        "horizon": "act",
        "horizon_de": "Handeln",
        "sequence": 1,
        "trend_name": "EU AI Act Compliance",
        "keywords": ["ai-act", "regulation", "2024"],
        "rationale": "Immediate deadline pressure",
        "source": "web-signal",
        "source_url": "https://ec.europa.eu/...",
        "freshness_date": "2024-12",
        "score": 0.82,
        "confidence_tier": "high",
        "signal_intensity": 5,
        "indicator_classification": {
          "type": "leading",
          "lead_time": "12-24 months",
          "source_type": "regulatory"
        },
        "diffusion_stage": {
          "stage": "early_majority",
          "estimated_adoption": 0.25,
          "crossed_chasm": true
        }
      }
    ]
  },

  "execution": {
    "workflow_state": "agreed",
    "current_phase": 4,
    "phases_completed": ["phase-0", "phase-0.5", "phase-1", "phase-1.5", "phase-2", "phase-2.5", "phase-3", "phase-4"],
    "agreed_at": "2025-12-16T11:45:00Z",
    "candidate_review": {
      "iterations": 1,
      "final_verdict": "accept",
      "final_score": 85,
      "cells_regenerated": 0,
      "candidates_replaced": 0,
      "scoring_fixes_applied": 0,
      "forced_accept": false
    }
  },

  "downstream_integration": {
    "source_type": "trend-scout",
    "auto_load_candidates": true,
    "auto_configure_research_type": true,
    "auto_configure_dok_level": true,
    "auto_configure_language": true
  }
}

---

Dimension Matrix

Each dimension is used to scout trends. Each trend discovered in any dimension is then analyzed through the complete TIPS framework (T→I→P→S).

Each dimension has 3 subcategories to ensure balanced trend discovery across all aspects:

Dimension	Subcategory	German	Focus	Trend Anchors
externe-effekte	wirtschaft	Wirtschaft	Market forces, competition, economic factors	Multikrise, Digital Transform, Net Neutral
externe-effekte	regulierung	Regulierung	Policy, compliance, legal frameworks	CSR-D/LKSG, EU AI Act, EU Data Act
externe-effekte	gesellschaft	Gesellschaft	Demographics, societal shifts	Demografie, De-Coupling, De-Carbonisation
neue-horizonte	strategie	Strategie	Business model direction, strategic goals	Nachhaltigkeit, Resilienz, OPs Excellence
neue-horizonte	fuehrung	Führung	Leadership approaches, organizational change	Business Agility, Open Leadership, Purpose
neue-horizonte	steuerung	Steuerung	Governance, analytics, control systems	Trends Driven, Risk Management, Predictive KI
digitale-wertetreiber	customer-experience	Customer Experience	Customer touchpoints, engagement	Digital First, Omnichannel, Metaverse
digitale-wertetreiber	produkte-services	Produkte & Services	Offerings, product innovation	Smartification, Digital Twin, Digital Ecosystem
digitale-wertetreiber	geschaeftsprozesse	Geschäftsprozesse	Operations, process optimization	Hyperautomate, Smart Manufacturing, Digi Supply Chain
digitales-fundament	kultur	Kultur	Organizational culture, mindset	New Work, Employee Wellbeing, Data Culture
digitales-fundament	mitarbeitende	Mitarbeitende	Workforce, skills, talent	Digital Workplace, Up/Reskilling, Talent Management
digitales-fundament	technologie	Technologie	Tech infrastructure, platforms	Cyber Security, Data Platforms, Industry X-Cloud

Balancing Rule: Each cell (dimension × horizon) must have at least 1 candidate from each subcategory. With 5 candidates per cell and 3 subcategories, this ensures complete coverage with flexibility.

---

Error Handling

Scenario	Response
Industry not selected	Cannot proceed - prompt user
Project init fails	Exit with error details
All web searches fail	Continue with training-only (warning)

---

Integration with Downstream Pipeline

After trend-scout completes, the user proceeds with one of two paths:

Option A — Trend Report (simpler path)

Invoke /trend-report directly to generate a narrative TIPS trend report:

1. trend-report auto-discovers this project via tips-project.json
2. Enriches each trend with web-sourced quantitative evidence
3. Produces the CxO-level report organized by investment themes

Option B — Value Modeling (recommended, full pipeline)

Invoke /value-modeler to build T→I→P→S relationship networks and ranked solution templates:

1. value-modeler discovers the project and loads .metadata/trend-scout-output.json
2. Auto-configures research_type, DOK level, language from .config
3. Loads candidates from .tips_candidates.items
4. Builds relationship networks, investment themes, and solution templates
5. After value-modeler completes, invoke /trend-report for the full CxO narrative

---

Debugging

Logging

# Log file location
${PROJECT_PATH}/.logs/trend-scout-execution-log.txt

# View phase transitions
grep "\[PHASE\]" "${PROJECT_PATH}/.logs/trend-scout-execution-log.txt"

# View validation results
grep "\[VALIDATION\]" "${PROJECT_PATH}/.logs/trend-scout-execution-log.txt"

Common Issues

1. "Industry not selected" - User must select from taxonomy
2. "Project init failed" - Check workspace root ($PROJECT_AGENTS_OPS_ROOT or current directory) is writable