LIVING DOCUMENT • UNFROZEN • PATTERNS + TOOLS + IDEAS

Build Toolbox

The execution plan is frozen. This page is not. It captures patterns discovered, tools installed, and ideas queued — everything that makes the build faster, smarter, or better without changing WHAT we build.

≡ How This Fits

The execution plan says WHAT to build. This page says HOW to build it better.

Frozen

5 Spec Documents

Architecture, features, schema, trust, timeline. No changes.

→

Frozen

Execution Plan

150 tasks, 5 phases, 3 tracks. Build sequence locked.

→

Living

Toolbox (This Page)

Patterns, tools, skills, ideas. How to execute faster and better.

→

Future

Post-Launch Optimizations

AI citations, A/B tests, new verticals. After build is complete.

⚙ Active Build Stack

Tools and hooks running in every Claude Code session. Installed 2026-04-14.

Hooks (Always On)

✓ GateGuard — forces investigation before every first edit
✓ continuous-learning-v2 — captures patterns per-project
✓ strategic-compact — suggests /compact at phase boundaries
✓ claude-mem — automatic memory + semantic search (plugin)

Orchestration

✓ Conductor — parallel agents in git worktrees
✓ continuous-claude — autonomous sequential PR loops
✓ Superpowers — dev methodology (brainstorm/plan/TDD/debug)
◆ Agent Teams — watching (Anthropic research preview)

Skills (45 Total)

9 new — seo, accessibility, browser-qa, context-budget, brand-voice, gateguard, blueprint, plankton-code-quality, click-path-audit
18 ECC — domain knowledge skills
18 project — pipeline stages + CampaignForge-specific

Patterns to Steal

7 patterns from 7 repos

🧠

Process-Type Specialization

Source: spacebot (spacedriveapp) • 2K stars

High Now All Repos

Instead of one monolithic agent doing everything, define distinct process types: Channels (coordinate), Branches (think), Workers (execute), Compactor (manage context), Cortex (synthesize memory). Each has constrained tools and a clear lifecycle.

Benefit: Conductor agents become purpose-specific. A "channel" agent coordinates while "worker" agents build sitemap, schema markup, and E-E-A-T pages in parallel. Thinking (planning) never blocks execution. Matches the Superpowers brainstorm → plan → execute flow but with true concurrency.

Maps to: Conductor parallel work blueprint skill superpowers:subagent-driven-development Tasks 1.22-1.25 (Agent Executor)

🔍

Automatic Memory + Semantic Search

Source: claude-mem (thedotmack) • 55.9K stars • INSTALLED

High Active All Repos

Every tool call, edit, and observation is automatically captured in SQLite + Chroma vector DB. At session start, relevant past context is auto-injected. Semantic search across all past sessions via MCP tools.

Benefit: Conductor agents start with relevant past context — no more cold starts. Cross-session continuity without manual STATUS.md curation. Every insight from a contentforge session is searchable from a campaignforge-app session. 10x token savings via progressive disclosure (index first, details on demand).

Maps to: Every session Conductor cold-start problem Cross-repo context

📊

AI Citation Auditing

Source: seomachine (TheCraigHewitt) • 6.2K stars

High Post-Launch ContentForge

Generate 100+ commercial-intent prompts users would ask AI tools about education financial aid. Test against ChatGPT, Perplexity, and Claude. Map which sources get cited. Identify gaps where degreesources.com tools should appear but don't.

Benefit: Free traffic at zero CAC. Our 10 interactive tools are exactly the type of structured content AI search cites. "EFC Calculator" and "Scholarship Finder" are real queries users ask AI. Being cited feeds the trust flywheel: AI citation → organic traffic → better platform quality signal → lower CPMs on paid. No other EDU affiliate is optimizing for this.

Maps to: degreesources.com tools Schema markup (1.58) Post-launch SEO New skill: /ai-citation-audit

🎯

Typed Memory with Relationship Edges

Source: spacebot (spacedriveapp) • 2K stars

Medium Phase 2+ campaignforge-app

Not all memories are equal. 8 memory types (Fact, Preference, Decision, Identity, Event, Observation, Goal, Todo) with relationship edges (RelatedTo, Updates, Contradicts, CausedBy, PartOf). A Decision that Contradicts a previous Fact is more useful to surface than two unrelated Facts.

Benefit: When building the platform's Winners Vault and campaign retro system, relationship-aware memory surfaces "this angle worked because it contradicted the previous assumption" instead of flat lists. Informs the pipeline's cross-campaign learning loop.

Maps to: Winners Vault Campaign Retro (Stage R) Obsidian pattern recognition

🛠

Transactional Actions with Preview

Source: spacedrive (spacedriveapp) • 37.8K stars

Medium Phase 2 campaignforge-app

Every file operation is a previewable, durable "transactional action." You see what will happen before it happens, and you can roll back. CQRS pattern: actions (writes) separate from queries (reads).

Benefit: The platform's approval workflow (Tasks 2.12-2.13) becomes: user sees a preview of what each pipeline stage will produce, approves or adjusts, stage executes approved output. Extends the existing --dry-run pattern into a full UI-level preview system.

Maps to: Pipeline --dry-run Approval persistence (2.13) Strategy browser (2.12) Copy workspace (2.14)

🎤

Local Voice Cloning & TTS

Source: voicebox (jamiepine) • 17.4K stars

Low Stage 4 campaign-forge

5 TTS engines running locally (Qwen3-TTS, Chatterbox, HumeAI TADA). Voice cloning from seconds of audio. REST API on localhost. Zero API costs, no rate limits, unlimited generation with crossfade.

Benefit: Replaces ElevenLabs cost for Stage 4 creative generation. UGC voice variations for ad creative without per-generation fees. Phone capture → voice clone → ad narration variations. The REST API integrates directly into generate-voiceover.py as a provider alongside Google TTS.

Maps to: Stage 4 Creative Director generate-voiceover.py UGC content strategy

📱

Client/Server Agent Architecture

Source: opencode (anomalyco) • 143K stars

Low Phase 3+ campaignforge-app

Separate the agent brain from the interaction surface. Server runs the AI + tools + LSP locally, clients (TUI, desktop, mobile, web) connect to it. Multiple clients can monitor the same running agent.

Benefit: When the platform scales to a 3-4 person team, the Agent Executor runs as a server process on Coolify. Multiple team members monitor the same pipeline run. Mobile monitoring becomes possible. Validates the architecture already in the platform spec.

Maps to: Agent Executor (1.22) Coolify deployment (1.29) Team scaling (Phase 3)

Design Research

Queued — Critical gap identified

The Gap

The current degreesources.com design was done with basic prompts, not informed by competitive research or conversion data. For an affiliate site where the tools ARE the conversion funnel, design directly impacts tool completion rates, trust perception, CTA click-through, and platform quality scores.

🎨

Competitive Design Audit → Design Brief

Plan: docs/design-research-plan.md • 4 phases • 11 competitor sites

High Next Session ContentForge

Scrape and analyze 11 competitor/reference sites across 4 tiers: government tools (studentaid.gov), high-authority content (NerdWallet, Forbes Advisor), direct competitors (scholarships.com, fastweb), and best-in-class tool UX (calculator.net, smartasset.com). Extract layout patterns, trust architecture, tool UX, conversion architecture, and performance benchmarks.

Benefit: Produces a locked design brief (same treatment as the 5 approved specs) that every BUILD-C visual decision references. Replaces guesswork with data-driven design. The brief covers: layout templates, component library specs, tool UX patterns, trust signal placement, CTA hierarchy, mobile spec, and performance budget.

Maps to: BUILD-C SEO (1.57-1.59) Quality gates (1.60-1.64) All 10 Svelte tool islands Astro layouts

📝

SEO Content Pipeline Patterns

Source: seomachine (TheCraigHewitt) • 6.2K stars

Medium Post-Launch ContentForge

Context files architecture (voice, SEO guidelines, competitor data, internal link map, AI citation targets) that every content command reads. Opportunity scoring with 8 weighted factors for content prioritization. Content lifecycle directories with per-stage validation.

Benefit: When expanding content (new articles, new tools, new verticals), a scoring model prioritizes what to build next by search volume, conversion intent, competition, and AI citation potential. Context files ensure voice and SEO consistency across all generated content.

Maps to: content-brief.json Content expansion planning New vertical content sites

Obsidian Pattern Pipeline

Adapted for CampaignForge-specific extraction

The Adaptation

The Obsidian repo analysis workflow already does excellent pattern recognition. It now has 5 project-specific filters so that every extracted pattern is scored for CampaignForge relevance: Conversion Architecture, Pipeline Orchestration, Multi-Agent Coordination, Platform Trust Signals, and Operational Scale.

Filter 1

Conversion
Architecture

Tool UX, CTAs, trust signals, mobile patterns

Filter 2

Pipeline
Orchestration

Stages, validation gates, context, cost tracking

Filter 3

Multi-Agent
Coordination

Decomposition, merge, shared state, routing

Filter 4

Platform
Trust Signals

E-E-A-T, schema, performance, accessibility

Filter 5

Operational
Scale

Multi-tenant, config-driven, white-label, analytics

Full adaptation spec: docs/obsidian-pattern-recognition-adaptation.md. Repos queued for analysis: spacebot, claude-mem, opencode, spacedrive, voicebox, plus 4 competitor sites from design research.

Installed Tools Reference

Full inventory of build acceleration tools

Tool	Type	Status	What It Does	Maps To
claude-mem	Plugin	Active	Automatic memory capture + semantic search via SQLite/Chroma	Every session
continuous-claude	CLI	Installed	Autonomous sequential PR loops with cost caps	Bulk tasks (tests, articles)
Conductor	App	Active	Parallel Claude Code agents in git worktrees	BUILD-C + BUILD-P parallel
GateGuard	Hook	Active	Fact-forcing pre-edit investigation gate (+2.25 quality)	Every Edit/Write/Bash
continuous-learning-v2	Hook	Active	Project-scoped instinct extraction with confidence scoring	Pattern learning
strategic-compact	Hook	Active	Suggests /compact at logical phase boundaries	Long sessions
seo	Skill	Installed	Technical SEO audit, schema markup, Core Web Vitals	BUILD-C 1.57-1.59
accessibility	Skill	Installed	WCAG 2.2 AA implementation + audit checklist	BUILD-C 1.63
browser-qa	Skill	Installed	4-phase automated visual QA via Chrome DevTools MCP	BUILD-C 1.60-1.64
brand-voice	Skill	Installed	Source-derived voice profile for DegreeSource	OPERATE 0.19-0.20
blueprint	Skill	Installed	Cold-start execution plans for Conductor agents	Task decomposition
context-budget	Skill	Installed	Token overhead audit across skills/MCP/agents	Run once to baseline
click-path-audit	Skill	Installed	State interaction bug finder for interactive tools	10 Svelte tool islands
plankton-code-quality	Skill	Needs Setup	Write-time auto-format + lint + Claude-powered fix	Activate with linters
Voicebox	App	Future	Local voice cloning + 5 TTS engines, zero API cost	Stage 4 creative gen

Discovery Log

When patterns were found and logged

2026-04-14 — Initial Toolbox Session

Research

Reviewed full ECC Universal Package (183 skills). Installed 9 new skills + 3 hooks. Researched 8 repos (andrej-karpathy-skills, claude-mem, voicebox, spacedrive, spacebot, superpowers, opencode, seomachine). Extracted 7 architectural patterns. Identified design research gap. Installed claude-mem plugin + continuous-claude CLI. Created Conductor setup guide, patterns-to-steal doc, design research plan, and Obsidian pattern recognition adaptation.