/nexus

Nexus Package Recipe Reference

Purpose: Generalized document-package generator — turn a single theme/idea/problem into a comprehensive, cross-functional, multi-format file tree (Markdown, CSV, JSON, YAML, SQL, HTML/CSS, Mermaid) + zip, for any of a registry of domain presets. One shared engine; per-domain blueprints swap the directory layout, the role→skill mapping, the distinctive outputs, the traceability anchor, and the risk gates. Read when: User invokes /nexus package (optionally domain=<preset>), or asks for a complete documentation package in a domain covered by the preset registry below (research plan, AI-adoption plan, legal/compliance pack, media-operation kit, growth-experiment plan, generic project package, …).

venture is the startup preset of this recipe. /nexus venture ≡ /nexus package domain=startup. The startup blueprint's full per-file detail lives in reference/venture-recipe.md; this file is the canonical engine + preset registry.

Generated-content language follows the CLI output-language config; this reference and all recipe instructions are English. File names, IDs, schema keys, and code stay English.

Contents

• Overview
• Relationship to venture and other recipes
• Invocation
• The Shared Engine (domain-agnostic Phase 0-6)
• Generalized Traceability Anchor
• Domain Preset Registry
• Preset Auto-Detection
• Per-Preset Blueprints (startup / generic / research / ai-adoption / legal / saas / media / growth)
• Risk Gates
• Conditional Inclusion
• AUTORUN Chain Template
• Failure Escalation
• Cost and Latency Profile

Overview

package factors the venture recipe into engine + blueprint. Every domain package shares the same skeleton — Phase 0 framing → Phase 1 research → Phase 2 spine (canonical entity-id barrier) → Phase 3 parallel doc tracks → Phase 4 overview synthesis → Phase 5 integrate/validate → Phase 6 zip. Only the blueprint varies: which directories exist, which specialist owns each track, which files are distinctive, what the traceability anchor is, and which risk gates fire.

This keeps Nexus to one recipe instead of a dozen near-identical ones (Core Rule: minimum viable chain; avoid recipe sprawl). Adding a new domain = adding one row to the Preset Registry, not a new recipe.

Relationship to venture and other recipes

Invocation

Optional condition fields (the collection's "common input sheet") populate the framing contract; missing → documented assumptions in 00_*/assumptions.md: theme / background / problem / audience / region / objective / deliverable_use / budget / timeline / team / existing_assets / tech_pref / monetization_kpi / avoid / legal_safety_notes / depth.

The Shared Engine (domain-agnostic Phase 0-6)

Identical to venture's engine; only the blueprint binds differ.

Format syntax lint (Phase 5→6, engine-generic, all presets): JSON via python3 -m json.tool or jq; YAML via python3 -c "import yaml,sys;yaml.safe_load(...)"; SQL via basic parse; HTML/CSS via structural check; CSV via header presence. Results recorded in validation_report.md before the Phase 6 zip.

package_contract (Phase 0 emit):

package_contract:
  preset: startup | generic | research | ai-adoption | legal | saas | media | growth
  theme: <normalized one-paragraph theme>
  depth: lite | standard | raise | full
  modes: [...]
  entity_anchor: <id prefix per preset, e.g. F- / H- / UC- / R- / P- / E->
  output_language: <from CLI config>
  web_grounding: available | unavailable
  risk_flags: [legal | medical | finance | safety | public-equity | none]
  assumptions: [(field, assumed_value, why), ...]
  output_dir: <preset>_package
  zip_name: <preset>_package_<slug>.zip   # [A-Za-z0-9_-] only

Generalized Traceability Anchor

venture's "feature_id (F-001) barrier" generalizes to a canonical entity-id barrier — Phase 2 fixes the domain's primary entity list before any Phase 3 track runs, and every track references existing IDs only (never mints new ones). Phase 5 fails the package on dangling references or unmapped primary entities.

Universal Grounding Gate (all presets — not just research)

The entity-id barrier guards structural integrity; this gate guards factual integrity. Every preset generates factual claims (market sizes, adoption stats, "studies show", competitor numbers, salary ranges), and an AI document package is exactly where plausible-but-fabricated numbers slip in. So claim-grounding is a cross-preset Phase 5 gate, not a research/career-only rule:

• Every factual claim is one of three things, explicitly:
  1. Sourced — cites a row in references.md (live source when web_grounding == available).
  2. Assumption — marked inline ASSUMPTION — confirm and logged in 00_*/assumptions.md with the assumed value + why.
  3. Research-to-do — when grounding was unavailable, enumerated in research_todo.md as a lookup, not stated as fact.
• A bare number or external-fact stated as fact with none of the three fails Phase 5 validation — for all presets, not only research. Fabricated market/statistic/citation claims presented as established fact are the package equivalent of a hallucinated source.
• Internal design content (the user's own roadmap, the proposed feature set, opinions, recommendations) is exempt — the gate targets externally-checkable facts, not the plan's own propositions.
• Preset-specific grounding rules layer on top, not replace: research requires reproducibility+ethics files; legal adds the lawyer-review disclaimer. The universal gate is the floor every preset clears.

validation_report.md reports the count of: sourced claims / assumptions / research-to-dos / ungrounded-fact failures (must be 0 to ship).

This gate is the origin of — and is generalized to all doc-producing runs by — reference/doc-quality-protocol.md (W4). The full document discipline (W1 reader contract per document, W3 freshness metadata, W8 cross-doc single source of truth via the traceability anchor, W12 reader-path check) applies to every package; Phase 5's cross-doc consistency + traceability matrix are its W8 implementation.

Domain Preset Registry

Preset Auto-Detection

When domain= is omitted, Phase 0 resolves the preset from the theme. Detection is deterministic:

1. Explicit domain= always wins. Skip detection.
2. Score the theme's English intent against the signal table below; pick the highest-scoring preset.
3. Apply the precedence rules when more than one row matches.
4. Apply the confidence gate — confirm or fall back when the match is weak or tied.

Theme → Preset signal table

Precedence rules (multiple matches)

• legal vs other: if the deliverable is the legal pack → legal. If legal is one concern inside a broader build → the broader preset (startup/saas/…) with its legal track, not the legal preset.
• startup vs saas: saas only when the product itself is software-sold-as-a-service (platform, tenancy, pricing tiers). Otherwise startup.
• startup vs growth: launching something new → startup. Improving something already shipped → growth.
• growth (this preset) vs kaizen/growth-acceptance/experiment (other recipes): package domain=growth produces a planning package only. If the user wants to execute an improvement → route to kaizen; to gate a merge with growth proof → growth-acceptance; for a single A/B test design → experiment direct. State the chosen routing in the report.
• ai-adoption vs saas: building/selling an AI product → saas. Adopting AI for internal operations → ai-adoption.
• research vs deep-research (recipe): a full research operations package (design + literature + data + analysis + validation dirs) → research preset. A single cited report → deep-research direct.

Confidence gate

Per-Preset Blueprints

Each preset's distinctive outputs (beyond README/manifest/validation_report) and per-file conventions. The per-file Markdown convention (Purpose / Intended readers / Assumptions / Body / MVP-or-current treatment / Future expansion / Next steps / Related files) applies to all presets. CSV/JSON/YAML/SQL/HTML/CSS must be real, loadable, syntactically valid.

startup

Full 14-directory blueprint with per-file → agent mapping: reference/venture-recipe.md. Distinctive: one_page_pitch, design_tokens.json, index.html LP, database_schema.sql, api_design_openapi.yaml, test_cases.csv, backlog.csv, growth_experiments.md, risk_register.md.

generic

Lightest preset (the collection's "universal"). Distinctive: 00_overview/{summary,goals,assumptions,decision_summary,90_day_action_plan}, 01_research/{context,stakeholder_map,current_state,benchmark,references,research_todo}, 02_strategy/{strategy,positioning,success_metrics,roadmap,risk_register}, 03_design/{requirements,workflows,information_architecture,templates,diagrams_mermaid}, 04_execution/{action_plan,backlog.csv,timeline,owners,raci_matrix}, 05_assets/{examples,copy_templates,data_templates.csv,mock_data.json}, 06_evaluation/{test_cases.csv,checklist,metrics,review_process}, 07_operations/{operating_model,governance,maintenance,improvement_cycle}. No brand/LP. Use when the domain doesn't match a richer preset.

research

Distinctive: research_question.md, hypothesis.md, methodology.md, limitations.md, literature_review.md, source_matrix.csv, annotated_bibliography.md, coding_scheme.md, findings_template.md, peer_review_checklist.md, reproducibility_checklist.md, bias_and_ethics.md. Phase 1 may invoke the deep-research skill (via the Skill tool — it is a harness skill, not a skills/-path spawnable agent) for grounded literature. Anchor: hypothesis_id H-001 ↔ methodology ↔ finding ↔ source_matrix.csv row. Validity gate: every claim traces to a source or a research_todo entry; reproducibility + bias/ethics files mandatory.

ai-adoption

Distinctive: prompt_library.md, role_based_prompts.md, evaluation_prompts.md, ai_evaluation_cases.csv, hallucination_test_cases.csv, forbidden_use_cases.md, human_review_rules.md, rag_source_plan.md, model_selection.md, rollout_plan.md. Anchor: use_case_id UC-001 ↔ prompt ↔ ai_evaluation_cases.csv (EVAL-001) ↔ human_review_rule. Mandatory gates: every use case has ≥1 eval case + a human-review rule; high-risk use cases (medical/finance/legal) carry forbidden-use entries.

legal

Distinctive: privacy_policy_draft.md, terms_of_service_draft.md, ai_usage_policy.md, cookie_policy.md, moderation_policy.md, nda_template.md, vendor_contract_checklist.md, data_processing_agreement_outline.md, legal_risk_register.md, compliance_checklist.md, lawyer_review_points.md, data_rights_matrix.md, consent_management.md, incident_response.md, breach_response_checklist.md. Anchor: risk_id R-001 ↔ mitigation ↔ policy clause ↔ lawyer_review_point. HIGH-RISK gate (mandatory): every legal document carries an explicit "draft only — not legal advice — professional/legal counsel review required" disclaimer; Phase 5 fails the package if any draft lacks a lawyer_review_points.md reference. Surface the expert-review requirement in the final report.

saas

= startup blueprint re-keyed to the SaaS dirs, defaulting to b2b-saas + ai-product mode overlays. Distinctive: package_type_catalog.md, generation_flow.md, prompt_orchestration.md, agent_workflow.md, validation_engine.md, model_selection.md, zip_export.md, integration plans (notion_export.md, github_integration.md, figma_linear_jira_plan.md), database_schema.sql, api_design_openapi.yaml, pricing.md, package_quality_score.md, ai_eval_cases.csv. Reuse startup's tech + GTM tracks; add the AI-system track from the ai-adoption preset.

media

Distinctive: content_pillars.md, editorial_calendar.csv, episode_ideas.md, article_templates.md, tone_of_voice.md, visual_guidelines.md, thumbnail_guidelines.md, channel_strategy.md, seo_strategy.md, newsletter_strategy.md, sponsorship_plan.md, product_funnel.md, production_workflow.md, scripts.md, analytics_dashboard_plan.md. Anchor: content_pillar_id P-001 ↔ episode/article ↔ channel ↔ metric. Coverage note (known gap): no single skill owns editorial strategy — assembled from cue (scripts/episodes) + prose (articles/tone) + field (SEO) + vision/muse (visual) + funnel/pulse (monetization/analytics). Lossy vs a dedicated editor skill; flag to user. Risk: copyright/attribution/ad-disclosure — add canon[legal] (light) for sponsorship/ad-compliance review.

growth

Distinctive: funnel_analysis.md, friction_points.md, opportunity_matrix.md, growth_hypotheses.md, experiment_backlog.csv, ab_test_plan.md, success_metrics.md, prioritization_rice.csv, improved_flows.md, copy_tests.md, ui_variants.md, event_tracking_plan.md, dashboard_spec.md, analysis_template.md, learning_log.md. Anchor: hypothesis_id H-001 / experiment_id E-001 ↔ metric ↔ learning-log entry. Dedup discipline: this preset produces a planning package only. If the user wants to actually implement an improvement → route to kaizen; to gate a merge with growth proof → growth-acceptance; for a single A/B test design → experiment direct. State this routing in the final report when growth is selected.

Risk Gates

Conditional Inclusion

AUTORUN Chain Template

Nexus AUTORUN package domain=<preset> theme="<X>" depth=<...> mode=<...>
  ── Phase 0 Framing ──────────────────────────────────
  → parse theme + condition fields → resolve preset (auto-detect if omitted)
  → web_grounding check → risk-flag scan
  → clarify gate (≤3 Qs only on domain-unidentifiable / high-risk; else assume)
  → emit package_contract (entity_anchor per preset)
  ── Phase 1 Research ─────────────────────────────────
  → preset.research_skills (web-grounded | research_todo)
       research preset → may invoke deep-research
  ── Phase 2 Spine [BARRIER] ──────────────────────────
  → preset.spine_skills → ═══ EMIT canonical entity-id table (F-/H-/UC-/R-/P-/E-) ═══
       → bind to every Phase 3 track
  ── Phase 3 Parallel Doc Tracks (waves, entity-id-bound) ─
  → preset.track_skills (disjoint files; ≤7/hub per wave)
  ── Phase 4 Overview Synthesis (post-tracks) ─────────
  → spark + scribe (+ magi if depth≥raise)
  ── Phase 5 Integrate + Validate ─────────────────────
  → attest/judge(traceability matrix per preset anchor + cross-doc consistency)
  → universal grounding gate (ALL presets) → fail on ungrounded external fact (not sourced / ASSUMPTION / research-to-do)
  → risk gate (legal/ai-adoption presets) → fail on missing disclaimer/eval/review
  → Nexus: document_manifest.csv + validation_report.md + README.md + syntax lint
  ── Phase 6 Package ──────────────────────────────────
  → write tree (UTF-8) → zip -r <preset>_package_<slug>.zip
  → unzip -l test → secrets/PII scrub
  → report: zip path, file count, main contents, validation, caveats + routing notes

Resume

Checkpoint-resume (7 phases): persist the Phase 0 framing contract, the Phase 2 frozen entity list, and each Phase 3 track's completed files at its boundary. Resume re-enters at the first incomplete track — the entity list is never re-derived on resume, since re-minting IDs would break every cross-reference already written.

Termination Bound

N/A — package is a non-loop recipe. Phases 0-6 run once; Phase 5 validation either passes or fails the package with a named defect list. Validation failures are reported, not iterated over — the operator decides whether to re-run. Preset blueprints inherit this.

Output Report — Package Manifest Report (named)

Emitted inside NEXUS_COMPLETE on top of the base ## Nexus Execution Report, and mirrored on disk as document_manifest.csv + validation_report.md + README.md:

• Preset + blueprint — which domain preset ran, and (for an auto-detected domain) the detection score that chose it
• File inventory — every generated artifact with its owning track and format
• Traceability matrix — primary entities vs the documents that reference them; dangling references and unmapped entities listed (must be empty to ship)
• Grounding tally — sourced claims / assumptions / research-to-dos / ungrounded-fact failures (must be 0)
• Risk-gate disposition — which gates fired and how each was satisfied
• Gap notes — any novel domain produced as generic, with its #TODO(agent) promotion note

Scale

10-30 agents, single pass, mid-to-high cost — preset-dependent: the Phase 3 parallel doc-track count is the multiplier (typically 5-8 tracks), Phase 1 research adds 2-4, Phase 5 validation 2-3. No cycle multiplier (see Termination Bound).

Failure Modes Prevented

Preset blueprints inherit this section — reference/venture-recipe.md (the startup preset) adds no failure modes of its own; per-preset risk gates are tabulated in § Risk Gates above.

Failure Escalation

Cost and Latency Profile

Same guardrails as venture (5+ agent chain confirmation, full-depth confirmation, no-secrets scrub). When a native Dynamic Workflow substrate is available, delegate the Phase 3 parallel sweep to it and keep Nexus as the entity-id-contract + validation layer.

Coverage Map vs the source collection

The remaining source domains now have first-class presets backed by a dedicated owner skill (career/learning/hiring were retired with their owner skills). A novel domain outside these still falls back to generic with a surfaced gap note; promote it once a dedicated skill exists (propose via architect).