Deep Research → Interactive Dashboard Pipeline

Takes a simple research topic from a user, autonomously discovers dimensions, metrics, subgroups, and taxonomies, then produces an interactive web dashboard.

Extension System — Specialized workflows (product comparisons, incident reviews, sprint analyses) are handled by pluggable extensions that auto-detect from user prompts and augment the pipeline. See extensions/registry.md.

ENVIRONMENT CHECK → INTERPRET → SURVEY → DISCOVER → RESEARCH → ANALYZE → BUILD → ENRICH → PRESENT
       ↓                ↓                     ↓                                ↓
  Hub exists?       Extension?         User Checkpoint                   Glossary (always on)
    ↓ no               ↓ yes
  FIRST-TIME SETUP  Extension Phase 1B

---

Phase 0: ENVIRONMENT CHECK

See hub-architecture.md for config schema, directory structure, and project registry.

Shell Variables

SKILL_ROOT is the directory containing this SKILL.md file. Set once per session:

GEN="<skill-root>/scripts/hub-gen.mjs"
HUB=$(node -e "try{console.log(JSON.parse(require('fs').readFileSync('<skill-root>/config.json','utf-8')).personalHubPath)}catch{process.exit(1)}")

Pre-check: If $GEN does not exist, halt with: hub-gen.mjs not found at $GEN — check skill installation.

Config (Two-Layer)

1. Pointer config (<skill-root>/config.json): Machine-local. Contains only personalHubPath. Written by hub-gen.mjs scaffold --init. Located relative to the skill's install directory (works for both global and workspace installs).
2. Portable config (<personalHubPath>/hub-config.json): Git-synced. Port, gitRepo, libraries, projects, telemetry.
3. Machine-local vite config (<personalHubPath>/.local-config.json): Gitignored. Library paths for Vite aliases.

No-Topic Invocation

If invoked without a topic: run detection/setup, start dev server, open browser preview, stop — do NOT proceed to Phase 1.

Detection

Track: Derive slug (kebab-case) from topic immediately. Batch in one command:

node $GEN $HUB track <slug> session-start 2>&1 | tail -1 && \
node $GEN $HUB track <slug> phase-start environment 2>&1 | tail -1

1. Check pointer config exists
2. If exists: Read personalHubPath → read hub-config.json → git pull (stash if needed) → sync libraries (git -C <localPath> pull for each with browseEnabled: true) → check dev server on port → if no topic, start server + stop; if topic, run batch then → Phase 1:

   node $GEN $HUB track <slug> phase-end environment 2>&1 | tail -1 && \
   node $GEN $HUB track <slug> phase-start interpret 2>&1 | tail -1

3. If not exists → First-Time Setup (runs ONCE per machine — clone or scaffold hub, configure libraries)

Key Principles

• No feature depends on another. Zero GitHub setup → fully functional local hub.
• Phase 0 should take ~2 seconds when hub exists.
• See hub-architecture.md § Telemetry for schema and formulas.

Build Log (Optional)

Phase timing via build-log.jsonl. write-meta reads build log → computes timing → writes session-end → preserves the file as a permanent committed artifact. AI provides zero timing data.

Opt-out: Phase timing is on by default. User can say "skip timing" at any point to disable for the current project.

1. First track batch: run with SafeToAutoRun: true. Note to user: "Phase timing enabled. Say skip timing to disable."
2. Auto-run succeeds → no further prompts needed.
3. Manual approval required → remind on each: "Say skip timing to stop these prompts."
4. On skip → stop all track commands. At write-meta, pass "phaseTiming": null. write-meta will set timingSource: "estimated" automatically (no build log).

When disabled: Skip all > **Track** blocks throughout the phases.

When enabled:

• track <slug> phase-start <phase> / phase-end <phase> at every boundary.
• track <slug> user-prompt discover checkpoint before and user-response discover checkpoint after.
• Every > Track block is exactly ONE run_command. Chain all calls with &&. This is not optional.
• $GEN and $HUB are set in the Shell Variables block above; up to 6 track calls per command.

node $GEN $HUB track <slug> phase-end <A> 2>&1 | tail -1 && \
node $GEN $HUB track <slug> phase-start <B> 2>&1 | tail -1 && \
node $GEN $HUB track <slug> phase-end <B> 2>&1 | tail -1 && \
node $GEN $HUB track <slug> phase-start <C> 2>&1 | tail -1

• Always set SafeToAutoRun: true on track commands. No other hub-gen.mjs subcommands qualify.
• When timing is enabled, all 9 phases (environment, interpret, survey, discover, research, analyze, build, enrich, present) must have both phase-start and phase-end events.

Derive slug in Phase 1 (kebab-case, same as add-project). For resumed sessions: track-read <slug> --json. See hub-architecture.md § Build Log for the full event table.

---

Phase 1: INTERPRET — Understand Intent

Track: Batched with phase-end environment from Phase 0 (no separate command needed if chained).

Parse natural language into research parameters:

1. Core topic — what is being studied?
2. Implied populations — map casual references to researchable groups
3. Time scope — temporal boundaries (do NOT choose segmentation yet)
4. Research intent — trends? comparison? impact analysis?
5. Research lenses — Population (demographics, health) · Behavior (prevalence, outcomes) · Industry (market, revenue) · Culture (shifts, media). Most topics combine 2-3.

Extension Detection

Scan prompt against extensions/registry.md triggers. If matched, load EXTENSION.md — it may inject Phase 1B before Phase 2.

Output: Topic, populations, time scope, intent, lenses, active extension. → Phase 2.

Track (batch):

node $GEN $HUB track <slug> phase-end interpret 2>&1 | tail -1 && \
node $GEN $HUB track <slug> phase-start survey 2>&1 | tail -1

---

Phase 2: SURVEY — Broad Landscape Scan

1. 3-5 broad searches: "[topic] research overview", "[topic] major studies statistics", "[topic] trends data [time scope]", "[population] demographics research", "[topic] systematic review OR meta-analysis"
2. Identify: Major datasets, commonly measured dimensions, known subgroups, temporal inflection points, data availability
3. Temporal segmentation — pick the lens where data looks different on either side of the boundary: inflection-point eras > policy eras > technology eras > decades > generations

Extension hook: Load Phase 2 instructions if active.

Track (batch):

node $GEN $HUB track <slug> phase-end survey 2>&1 | tail -1 && \
node $GEN $HUB track <slug> phase-start discover 2>&1 | tail -1

---

Phase 3: DISCOVER — Dimension Discovery

3A: Subgroup Discovery

Split test: >20% difference → SPLIT; <20% → MERGE. Max 2 levels, 5 subgroups. See subgroup-discovery.md.

3B: Core Metrics

Apply Standard Research Dimensions Framework per lenses. See research-dimensions.md.

3C: Taxonomy Discovery

Search academic taxonomies, cross-reference 2-3 sources, trim to top 25-30. See subgroup-discovery.md.

Extension hook: May override 3A-3C.

3D: User Checkpoint

Track (batch — run immediately before presenting checkpoint):

node $GEN $HUB track <slug> phase-end discover 2>&1 | tail -1 && \
node $GEN $HUB track <slug> user-prompt discover checkpoint 2>&1 | tail -1

Present research plan:

TOPIC · TIME AXIS · POPULATIONS · CORE METRICS · TAXONOMIES · PROPOSED SECTIONS · KEY DATA SOURCES

Ask: "Should I proceed, or adjust anything?"

HARD RULE: ALWAYS wait for explicit user response. Do NOT auto-approve for benchmarks, speed tests, or any reason. Silence does not count as approval.

After presenting, inform default visibility (personal if gitRepo exists, local if not). See hub-visibility.md. For libraries with contributeEnabled: true: if confirmEachShare: true (default), ask about sharing; if confirmEachShare: false (power-user pre-approval), auto-set visibility: "public" — no ask needed. Never push to a public library without user approval (either per-project ask or blanket pre-approval via confirmEachShare: false).

Track (batch — run immediately after user approves):

node $GEN $HUB track <slug> user-response discover checkpoint 2>&1 | tail -1 && \
node $GEN $HUB track <slug> phase-start research 2>&1 | tail -1

---

Phase 4: RESEARCH — Deep Data Gathering

1. For each metric × group × time period, search and extract statistics
2. Tag every data point with quality tier (T1 Gold → T4 Estimate)
3. Triangulate — 2-3 independent sources per data point
4. Fill gaps with T4 estimates, clearly marked
5. Gather taxonomy prevalence per group × period
6. Track all sources in structured citation list

See subgroup-discovery.md for tier definitions. Extension hook: Load Phase 4 if active.

Track (batch):

node $GEN $HUB track <slug> phase-end research 2>&1 | tail -1 && \
node $GEN $HUB track <slug> phase-start analyze 2>&1 | tail -1

---

Phase 5: ANALYZE — Story & Visualization

Decisions made HERE, after seeing the data.

• 5A: Key findings — surprises, group differences, trends, counterintuitive results → insight callouts
• 5B: Auto-select chart types per visualization-rules.md
• 5C: Structure: Broad → Deep → Compare → Explore → Context → Sources. Each section: title + subtitle + 1-3 charts + insight callout.

Extension hook: Load Phase 5 if active.

Track (batch):

node $GEN $HUB track <slug> phase-end analyze 2>&1 | tail -1 && \
node $GEN $HUB track <slug> phase-start build 2>&1 | tail -1

---

Phase 6: BUILD — Into Research Hub

Never create a standalone Vite project. See hub-architecture.md, build-templates.md, collections-architecture.md.

Bespoke Mode (default)

hub-gen.mjs handles structure; AI writes only creative content.

1. Read pointer config → get personalHubPath
2. Read hub-config.json for port and project list
3. Generate slug (kebab-case)
4. node scripts/hub-gen.mjs <hub-path> add-project --json '<payload>' — payload: slug, title, subtitle, query, lens, icon, accentColor, visibility, createdAt, sections. Creates directory, App.jsx, updates config + registry. Parse --json output for createdPaths.
5. AI writes creative content:
   • components/*.jsx — sections with charts, callouts. Import shared components from ../../../components/.
   • data/*.js — research data as ES module exports.
6. node scripts/hub-gen.mjs <hub-path> install-components
7. Start dev server if not already running (port from hub-config.json):

   lsof -i :<port> | grep -q LISTEN || npm run dev &

8. Open browser preview

Code hygiene rule: Only declare a variable (const, let) if it is referenced in the JSX return value or passed to a child component. Do not leave computed values that are unused. If a value was computed for exploration but isn't needed in the final render, delete it before moving to the next component.

Atomic write groups: Write files in three ordered batches to prevent Vite HMR from seeing partial project state:

1. All data/ files (researchData.js, any JSON)
2. All components/ files
3. App.jsx + registry update (sync-registry) + hub-config.json update

Track (batch):

node $GEN $HUB track <slug> phase-end build 2>&1 | tail -1 && \
node $GEN $HUB track <slug> phase-start enrich 2>&1 | tail -1

Shared Component Reference

Import path: ../../../components/ — that's 3 levels up from src/projects/<slug>/components/. NOT 2.

Component	Import	Props	Notes
InsightCallout	import InsightCallout from '../../../components/InsightCallout'	children, color	Colors: violet (default), amber, emerald, blue, red, cyan, orange, indigo, rose, teal, purple. No emoji/icon prop.
CustomTooltip	import CustomTooltip from '../../../components/CustomTooltip'	Recharts content	<Tooltip content={<CustomTooltip />} />
GlossaryTerm	import { GlossaryTerm } from '../../../components/GlossaryTerm'	term, glossary, children	term must exactly match a glossaryTerms key.

Template Mode (extension-driven)

When extension has output_mode: template → produces JSON data file. See collections-architecture.md.

Store Original Query

Always store user's original query in hub-config.json query field for hub UI display.

---

Phase 6B: ENRICH — Key Term Glossary

Scan built project text, identify domain/technical terms a general audience wouldn't know, wrap in <GlossaryTerm>. Enabled by default (glossaryEnrichment: true).

Density: terms = max(3, floor(wordCount / 250)). Max 2 per section. Never two in same sentence.
Priority: Acronyms → Jargon → Technical → Tribal. Skip common words, proper nouns.

Steps

1. Scan text in components (overview paragraphs, callouts, labels)
2. Rank by obscurity, apply density rules, distribute across sections
3. Generate definition + research prompt per term
4. Create data/glossaryTerms.js — { definition, researchPrompt } per term
5. Wrap in components: <GlossaryTerm term="X" glossary={glossaryTerms}>X</GlossaryTerm>

Enrich Guardrails

1. JSX only. Wraps go in components/*.jsx ONLY. NEVER in data/*.js.
2. Key-match required. term="X" must exactly match a glossaryTerms.js key. Verify before writing.
3. Single-pass editing. Plan ALL wraps first, execute in one multi_edit per file. No iterative fix passes.
4. No invented props. See Shared Component Reference above.

Track (batch):

node $GEN $HUB track <slug> phase-end enrich 2>&1 | tail -1 && \
node $GEN $HUB track <slug> phase-start present 2>&1 | tail -1

---

Phase 7: PRESENT — Validate & Deliver

Steps 1-5 must complete before 6-9.

1. node scripts/hub-gen.mjs <hub-path> validate --fix --build --json — 13 categories + vite build. Parse summary.passed.
   • Zero-warning gate: After --fix, re-run node $GEN $HUB validate --check. Exit code 0 = clean. Exit code 1 = errors (fix before proceeding). Exit code 2 = warnings remain (run node $GEN $HUB doctor then re-validate).
2. Fix remaining errors manually if any.
3. QA: charts render, tooltips work, findings match data, citations complete.
4. Extension QA if active.
   4b. If phase timing enabled: track <slug> phase-end present — must run before write-meta.

node $GEN $HUB track <slug> phase-end present 2>&1 | tail -1

5. Telemetry GATE: node scripts/hub-gen.mjs <hub-path> write-meta <slug> '<json>' — provide all required fields per hub-architecture.md § Telemetry. Missing fields = exit 1.
   • model: Use the exact model identifier if known (e.g. "claude-sonnet-4-5"). Use null if not retrievable — the card will not show it.
   • If timing enabled: DO provide phaseTiming estimates (tracker overrides with real values). write-meta sets timingSource automatically ("verified" if build log has session-end, "estimated" otherwise).
   • If timing skipped: Pass "phaseTiming": null. write-meta sets timingSource: "estimated" automatically.
6. Git sync: git add -A → git commit → git push
7. Library share: Only if visibility: "public". Push via GitHub API per hub-contribution.md.
8. Library sync: git -C <localPath> pull for pushed libraries.
9. Deliver: Open browser preview using the port from hub-config.json (not any default). Verify port: lsof -i :<port> | grep -q LISTEN && echo running || echo not running. Summarize.

---

Edge Cases

See hub-architecture.md.