This is an example of a fully-populated
cortex.mdfrom a real Claude Code user. Routes have been generalized — your stack will look different.Use this as a shape reference, not a copy-paste target. The framework (Routing Protocol, Effort Calibration, Self-Learning Loop) is identical to
templates/cortex.md. The Workflow Registry and Decision Shortcuts below show what it looks like when you populate it with a real, opinionated stack.Convention used below: locally-cloned tools live at
~/tools/<name>/with a thin CLI wrapper at~/.claude/bin/<name>.
Installed: Yes Strengths: Multi-agent orchestration, verify/fix loops, parallel execution, team mode Best for: Complex builds, planning, multi-step tasks needing quality gates, research
| Pattern | Use When |
|---|---|
ralplan "task" |
Important work needing planning + verify/fix loops. Default for non-trivial work. |
autopilot: task |
Well-defined, lower-risk tasks. One-shot autonomous. |
ralph: task |
Must be bulletproof. Won't stop until verified complete. |
/team N:executor "task" |
Parallel execution, multiple perspectives. |
/deep-interview |
Requirements unclear. Clarify before building. |
ultrawork |
High-throughput parallel task completion. |
/ccg |
Tri-model orchestration (Claude + Codex + Gemini). Auto-fires per Effort Calibration. |
/oh-my-claudecode:external-context |
Web research and documentation lookup. |
/oh-my-claudecode:self-improve |
Autonomous code improvement with benchmarking loops. |
/oh-my-claudecode:verify |
Turn "it works" into concrete evidence. |
/oh-my-claudecode:deep-dive |
2-stage: trace root cause then define requirements. |
Installed: Yes Strengths: Context rot prevention, spec-driven development, phase-based project management, autonomous execution Best for: Structured multi-phase builds, long sessions, preventing quality degradation, full project lifecycle
| Pattern | Use When |
|---|---|
/gsd-new-project |
Initialise a new project with spec-driven planning. |
/gsd-plan-phase |
Create detailed phase plan with verification. |
/gsd-execute-phase |
Execute all plans in a phase with wave-based parallelisation. |
/gsd-autonomous |
Run all remaining phases autonomously (discuss → plan → execute per phase). |
/gsd-do "task" |
Route freeform text to the right GSD command. |
/gsd-discuss-phase |
Gather context through adaptive questioning before planning. |
/gsd-debug |
Systematic debugging with persistent state across context resets. |
/gsd-map-codebase |
Scan and index codebase structure for context. |
/gsd-code-review |
Review changed files for bugs, security, quality. |
/gsd-ship |
Create PR, run review, prepare for merge. |
/gsd-resume-work |
Resume from previous session with full context restoration. |
/gsd-fast |
Execute trivial task inline — no subagents, no planning overhead. |
Installed: Yes Strengths: Code-focused workflows, PR pipelines, review cycles, LSP integration Best for: Feature implementation, code review, PR workflows
| Pattern | Use When |
|---|---|
feature-dev |
End-to-end feature implementation with structure. |
code-review |
Review code for quality, bugs, security. |
pr-review-toolkit |
PR-specific review with checklist. |
ralph-loop |
Persistent loop until task passes. |
code-simplifier |
Reduce complexity, clean up code. |
commit-commands |
Structured git commits. |
security-guidance |
Continuous 3-layer security review as code is written (per-edit pattern check → per-turn model review → agentic review on commit/push). |
Installed: Yes Strengths: Multi-model debate for spec hardening, cross-LLM adversarial review Best for: Hardening specs, API contracts, schemas, security designs before implementation Requires: At least one external model API key
| Pattern | Use When |
|---|---|
/adversarial-spec "spec" |
Harden a spec/contract through multi-model adversarial debate. |
Installed: Yes — protocol skill + cross-family reviewer CLIs Strengths: Cross-family review of build-phase diffs at checkpoints — structured findings, judge synthesis with dissent noted, spec-conformance checking Best for: Catching builder drift, logic bugs, and risky diffs before they land. The BUILD-phase counterpart to CCG's plan validation.
| Pattern | Use When |
|---|---|
| Medium — single critic | Default for L2/L3 builds with moderate risk: one cross-family reviewer on the finished diff (spec-check or devil's-advocate mode) |
| High — panel + judge | Any high-risk trigger: auth / payments / migrations / data deletion / agent routing, diff >300 LOC or >5 files, weak tests, architectural or API-contract change, external integrations, security exposure |
| Rival test generation | Reviewer proves a flaw with a failing test — objective, no debate needed |
| Skip | Small local diffs, mechanical changes, well-tested code |
Rules: advisory by default — BLOCK escalates to a human, never auto-rejects; one canonical review packet per checkpoint (spec + plan + diff + tests + uncertainties); findings carry severity, file:line, evidence, and a concrete failure scenario; max 3 reviewers, one is the default; verdicts are judged (PASS / CAUTION / BLOCK with dissent), never raw transcripts.
How it fits: CCG cross-checks plans before building; Sentinel cross-checks diffs before they land.
Installed: Yes (brew install gitleaks)
Source: github.com/gitleaks/gitleaks
Strengths: Fast, deterministic secret scanner (regex + entropy) for git history, working tree, and staged changes. No model in the loop — reproducible pass/fail.
Best for: Catching hardcoded secrets (API keys, tokens, private keys) before they land in a commit or PR.
| Pattern | Use When |
|---|---|
gitleaks git -v |
Scan the full git history of a repo for leaked secrets. |
gitleaks dir -v |
Scan a directory / working tree (files on disk, not git history). |
gitleaks protect --staged |
Scan staged changes before commit — the pre-commit gate. |
| pre-commit hook / CI action | Wire into pre-commit or GitHub Actions so every commit and PR is scanned. |
How it fits: the deterministic first line — cheap, reproducible, never hallucinates. Pair with a Security Engineer agent and continuous model-based review for defense in depth.
Installed: Yes — data, legal, enterprise-search enabled
Strengths: First-party plugins for non-code knowledge work
Best for: SQL / dashboards, contract triage, cross-tool search
| Plugin | Use When |
|---|---|
data |
SQL + dashboards + self-validation over a dataset — pairs with DuckDB / Supabase MCPs |
legal |
Contract / NDA triage and review |
enterprise-search |
Cross-tool search over email / chat / docs |
product-management |
PRDs, roadmaps, prioritisation (install on demand) |
bio-research |
PubMed / genomics / BioRender (install on demand) |
Installed: Pending — claude plugin marketplace add anthropics/claude-tag-plugins, then install services on demand
Strengths: 18 first-party SaaS connector plugins, one per service; credentials injected by the runtime, not embedded
Best for: Giving the agent direct, routable access to issue trackers, CRM, data warehouses, and observability
| Plugin | Use When |
|---|---|
jira / linear / asana |
Issue and project tracking |
salesforce / hubspot |
CRM — accounts, contacts, deals, pipeline |
datadog / grafana / pagerduty / sentry |
Observability and incident (pairs with SRE / Incident Responder agents) |
bigquery / snowflake / redshift |
Cloud data warehouse SQL (pairs with the data plugin) |
confluence / notion / google-drive |
Docs and knowledge base (dedupe against Cloud MCPs) |
Effort tier default: L1 to install or query; the task using it sets the real tier.
Installed: Yes Strengths: First-party document creation / editing skills Best for: Programmatic Word reports, Excel models, PowerPoint decks, PDF manipulation; canonical playbooks for shipping MCPs and skills
| Skill | Use When |
|---|---|
docx |
Create / read / edit Word documents — reports, memos, letters, templates |
pdf |
Read / extract / merge / split / fill / encrypt PDFs, OCR scanned PDFs |
pptx |
Create / read / edit slide decks — pitch decks, presentations, speaker notes |
xlsx |
Create / read / edit spreadsheets — formulas, charts, clean messy tabular data |
mcp-builder |
Canonical playbook when shipping an MCP server |
skill-creator |
Build / edit / benchmark Claude skills with eval support |
Chains naturally: backtest output → xlsx model → pptx deck → pdf export.
Installed: Yes Strengths: First-party finance skills with Excel / PowerPoint integration Best for: Financial modelling, equity research, earnings analysis, pitch decks, meeting prep
| Plugin | Use When |
|---|---|
financial-analysis |
DCF, comps, LBO, 3-statement models, competitive analysis, deck QC |
model-builder |
Build live Excel models — DCF, LBO, 3-statement, comps |
pitch-agent |
Comps + precedents + LBO → branded pitch deck, end to end |
market-researcher |
Sector / theme → industry overview, competitive landscape, peer comps |
earnings-reviewer |
Earnings call + filings → model update → note draft |
meeting-prep-agent |
Briefing pack before client meetings |
equity-research |
Earnings analysis, initiating coverage reports (install if doing equity research) |
Installed: Yes
| Tool | Use When |
|---|---|
obsidian-skills |
Vault operations — Markdown / Bases / JSON Canvas via the Obsidian CLI. Powers "document builds in a notes vault as you go" |
claude-mem |
Automatic cross-session memory: captures sessions, compresses, injects relevant context |
last30days |
"What's new in the last 30 days" across Reddit / X / YouTube / HN — recency complement to deep-research |
defending-code harness |
Security work: /threat-model, /vuln-scan, /triage, /patch (Anthropic official; sandboxed pipeline needs Docker/gVisor) |
Installed: Yes (280+ specialists across 20+ divisions)
Strengths: Deep domain expertise per agent
Best for: Pair with orchestrators via combo syntax. Run /agents to list.
| Domain | Agent(s) |
|---|---|
| Backend / API / architecture | Software Architect, Backend Architect |
| Frontend / UI | UI Designer, Frontend Developer |
| Database | Database Optimiser |
| Security | Security Engineer |
| AI / ML | AI Engineer, Voice AI Integration Engineer |
| DevOps | DevOps Automator |
| UX | UX Researcher |
| Brand / copy / content | Brand Guardian, Content Creator |
| Growth | Growth Hacker, Agentic Search Optimizer (AEO/GEO) |
| Finance | Finance Tracker, Investment Researcher, Tax Strategist, FP&A Analyst |
| Onboarding / minimal change | Codebase Onboarding Engineer, Minimal Change Engineer |
| Legal | Legal Client Intake, Legal Document Review, Legal Billing |
| HR / recruitment | HR Onboarding, Recruitment Specialist |
| Cross-functional | Chief of Staff |
| Localisation | Language Translator |
Installed: Yes Strengths: Framework-specific expertise with live docs Best for: React, Next.js, Three.js, Stripe, Python tooling, frontend design
| Skill | Use When |
|---|---|
tessl__senior-frontend |
React, Next.js, TypeScript, Tailwind work |
tessl__nextjs-developer |
Next.js 14+ App Router, server components |
tessl__stripe-integration |
Payment processing, subscriptions |
tessl__modern-python |
Python project setup with uv, ruff, ty |
tessl__frontend-design |
High-quality UI implementation |
tessl__3d-web-experience |
Three.js, WebGL, interactive 3D |
Installed: Yes — cloned at ~/tools/hyperframes/
Source: github.com/heygen-com/hyperframes
Strengths: Write HTML → render MP4. GSAP timelines, Tailwind browser runtime, Three.js / Lottie / Anime.js adapters
Best for: Product intros, social ads, animated explainers, motion graphics from CSV / PDF / URL / repo, captions synced to TTS
| Skill | Use When |
|---|---|
hyperframes |
Author video compositions — scenes, animations, captions, transitions |
hyperframes-cli |
Dev loop — init, lint, inspect, preview, render, doctor |
hyperframes-media |
TTS narration, transcription, background removal |
website-to-hyperframes |
URL → video. Product tour / social ad / promo from a link |
Chains naturally: landing page → website-to-hyperframes → 30s promo. Backtest results → hyperframes chart-race video.
Installed: Yes Source: github.com/safishamsi/graphify Strengths: Turns mixed content (code, docs, papers, images, videos) into a queryable knowledge graph Best for: Research-heavy projects, cross-content reasoning, onboarding doc-heavy codebases
| Pattern | Use When |
|---|---|
/graphify . |
Build a knowledge graph for the current folder |
/graphify ./path --update |
Merge new content into an existing graph |
/graphify ./path --watch |
Auto-sync as files change — long-running work |
/graphify query "..." |
Semantic question across all indexed content |
/graphify path "NodeA" "NodeB" |
Find shortest conceptual path between two nodes |
/graphify explain "Concept" |
Surface god-nodes and surprising connections |
Rituals: onboarding (/graphify . + /gsd-map-codebase), pre-refactor dependency check ("what depends on X beyond direct imports?"), long-running project sync (--watch), local-vs-external research ("what's our prior work on X?" then deep-research). Skip on repos under ~20 files, single-file fixes, and throwaway scripts.
Installed: Trial — head-to-head vs Graphify not yet settled Source: github.com/Egonex-AI/Understand-Anything Strengths: Code-only knowledge graph with guided architecture tours, layer visualisation (API / Service / Data / UI / Utility), and business-domain mapping. Tree-sitter for deterministic structure + a five-agent LLM pipeline for semantics. Best for: Onboarding onto a single, code-only repo where guided tours beat a concept graph
| Pattern | Use When |
|---|---|
/understand |
Scan the repo and build the graph (incremental updates supported) |
/understand-dashboard |
Interactive web dashboard to explore the graph visually |
/understand-chat |
Ask natural-language questions about the codebase |
/understand-domain |
Extract business domains, flows, process steps |
/understand-onboard |
Generate an onboarding guide / guided tour |
/understand-diff |
Blast-radius analysis of current changes — pre-refactor impact check |
vs Graphify: Graphify spans code + docs + papers + media; Understand Anything is code-only but goes deeper on code (tours, layers, domains). Mixed-content repos → Graphify. Pure-code repos → trial Understand Anything.
Installed: Yes (Kronos, QuantMind)
Strengths: Pre-trained foundation models and extraction frameworks for domains where generic LLMs underperform
Best for: Zero-shot baselines, feature generation, fine-tuning on proprietary data
Convention: each lives at ~/tools/<name>/ with a wrapper at ~/.claude/bin/<name>
Source: github.com/shiyu-coder/Kronos What it does: Open-source foundation model for OHLCV candlestick prediction. Trained on 45+ exchanges. Variants: mini → small (default) → base → large.
| Pattern | Use When |
|---|---|
kronos predict <ticker> [--lookback N --horizon N] |
Zero-shot OHLCV forecast |
kronos backtest <ticker> --strategy <name> |
Evaluate predictions vs historical |
kronos finetune <dataset.csv> |
Adapt to a proprietary universe / alt data |
Don't use for: single-asset alpha generation (use as a feature, not an oracle), regime shifts outside the training distribution, news/event-driven trading (price-only, no text awareness).
Source: github.com/LLMQuant/quant-mind What it does: Ingests financial research at scale — arXiv papers, SEC filings, news — and turns it into a structured, queryable knowledge base via an LLM extraction pipeline. Status: Trial — confirm output quality and per-paper cost before committing further.
| Pattern | Use When |
|---|---|
paper_flow(<source>) |
Extract structured knowledge from a single paper / filing / news item |
batch_run(<sources>) |
Ingest a batch of sources into the knowledge base |
Don't use for: graphing content you already own (use Graphify), financial modelling on extracted data (use the finance plugins), open-ended web research with citations (use deep-research).
Installed: Yes — ~/design-systems/
Source: github.com/VoltAgent/awesome-design-md
Strengths: 58+ real-world DESIGN.md specs (Stripe, Linear, Vercel, etc.) as plain markdown
Best for: Giving a frontend build a real design language instead of generic AI aesthetics
| Pattern | Use When |
|---|---|
Copy DESIGN.md into project root |
Starting a new UI build — agents auto-detect it |
| Browse the library | Picking a design system to base the project on |
Design pipeline: DESIGN.md (tokens) → Figma MCP (get_design_context) → frontend-design / tessl__senior-frontend skill or Frontend Developer agent → animate → polish / normalize → critique / audit → Chrome DevTools MCP for visual QA.
Installed: Yes (infrastructure layer — consumed by the systems above) Strengths: Live docs, AI web search, page extraction, faster edits, data access Best for: Powering research, documentation lookup, and edit performance
| MCP | What It Does | Auto-Used By |
|---|---|---|
| Context7 | Live docs for 10k+ packages | documentation-lookup skill, tessl query_library_docs |
| Exa | AI-native web search | deep-research, external-context skills |
| Firecrawl | Full page extraction to markdown | deep-research (Exa finds → Firecrawl extracts) |
| Scrapling | Adaptive scraper with anti-bot bypass (Cloudflare, DataDome) | Fallback when Firecrawl fails on bot-walled pages |
| Morph | Faster edits + semantic code search | Executor agents, edit operations |
| Supabase | DB queries, auth, storage, schema management | Database Optimiser agent, direct SQL |
| DuckDB | In-process OLAP — SQL over local Parquet / CSV / JSON | Backtest output analysis, ad-hoc tabular exploration |
| cortex-queue | Human-in-the-loop decision queue — enqueue_decision() / await_decision() / ask_human() |
Long-running unsupervised loops that hit ambiguity |
| Slack | Send/read messages, manage channels | Team communication, notifications |
Installed: Pending — brew install --cask tsouth89/toolport/toolport (or the install script)
Strengths: Local-first MCP gateway with lazy tool discovery — advertises four compact meta-tools instead of every server's full catalog
Best for: Cutting the per-session token cost of a large MCP surface; one shared server config across clients
| Pattern | Use When |
|---|---|
| Route MCP servers through the gateway | Many servers are registered and full catalogs bloat context every request |
toolport_search_tools "<query>" |
The agent should discover the right tool on demand rather than loading all upfront |
Effort tier default: L1 to install / query; L2 to re-wire the session's MCP config through it.
Installed: Yes Strengths: Direct access to productivity tools, communication, design, knowledge
| MCP | Use When |
|---|---|
| Figma | Design-to-code, inspecting mockups, Code Connect |
| Notion | Knowledge base, docs, project wikis |
| Gmail | Email context, drafting replies |
| Google Calendar | Scheduling, availability checks |
| Canva | Visual assets, presentations, social graphics |
| Granola | Meeting transcripts and notes |
| Scholar Gateway | Academic research, citations |
| GitHub | Issue triage, PR workflows |
Installed: Yes Strengths: Browser debugging, performance auditing, accessibility testing Best for: Visual QA, Lighthouse audits, network debugging, a11y checks
Installed: Optional local integration, not bundled with Cortex. Requires the installed ask-argon skill to be exposed in the current session, its configured local bridge, and an authenticated Antigravity CLI with the required account entitlement.
Strengths: Bounded Argon second opinions returned to the current Codex task for parent review.
Best for: Explicit user requests for Argon, a Gemini second opinion, or a Codex-to-Argon call.
| Pattern | Use When |
|---|---|
$ask-argon |
The user explicitly requests Argon, a Gemini second opinion, or Codex-to-Argon, and the skill is currently exposed. Read the installed skill and use its configured bridge location and invocation. |
Routing gate: Registry presence alone is not callability or authorization. If the skill is unavailable, report the limitation rather than installing or invoking a guessed command. Keep native delegation and configured host model defaults for ordinary subtasks. This route is not an automatic council, reviewer, or fallback for uncertainty or higher effort tiers.
Model and usage: Follow the installed skill's explicitly pinned models: gemini-4-argon-medium by default, or --effort high for gemini-4-argon-high when deeper reasoning is useful. Cortex tiers do not select these models. Usage is subscription-backed and subject to account quotas, not guaranteed free or unlimited. On authentication, entitlement, quota, or model-selection errors, report the error and stop dependent work. Do not silently substitute another model, fall back to an API, buy credits, upgrade plans, or enable API billing.
Context and review: Send only the minimum task-authorized prompt context, without secrets. The bridge uses a fresh temporary workspace with no project files automatically supplied. This is context isolation, not a security sandbox; cached credentials and CLI configuration remain accessible. Do not grant project-directory access unless requested. The parent Codex task reviews the output and retains integration, correctness, and final judgment. Only an accepted explicitly pinned model request with a successful response verifies an Argon call; a model list or answer self-identification does not.
Strengths: Zero overhead Best for: Quick edits, lookups, single-file changes, casual questions, git operations
| Task Type | Default Route | Tier |
|---|---|---|
| Plan / architect | OMC > /ccg pre-plan cross-check → ralplan > Software Architect | L3 |
| Build feature (backend) | OMC > ralplan > Backend Architect | L3 |
| Build feature (frontend) | OMC > ralplan > Frontend Developer + tessl | L3 |
| Build (well-scoped, brief locked, single specialist) ✓ | OMC > autopilot > [specialist] | L3 |
| Build (small bounded edits, ≤100 LOC, single domain) ✓ | Direct > batched-edits | L2 |
| Build checkpoint review (pre-commit / risky diff) | Sentinel > single critic (Medium), or panel + judge (High) | L2–L3 |
| Multi-phase project build | GSD > /gsd-new-project then /gsd-autonomous | L3 |
| Long session / context-heavy | GSD > /gsd-do (context rot prevention) | L2 |
| Resume previous work | GSD > /gsd-resume-work | L2 |
| Code review | ECC > code-review | L2 |
| Explicit request for Argon / Gemini second opinion / Codex-to-Argon | Argon > ask-argon, optional exposed Codex skill; bounded prompt and parent review, registry gates above apply |
L2 |
| PR workflow | ECC > pr-review-toolkit | L2 |
| Hardcoded secret scan (pre-commit / CI gate) | Direct > gitleaks (deterministic) — pair with Security Engineer for deep review | L1 |
| Security audit | OMC > /ccg attack-surface pass → ralplan > Security Engineer + defending-code harness | L4 |
| Harden spec / API contract | Adversarial Spec > /adversarial-spec | L4 |
| Library / framework docs | Context7 (direct) or tessl query_library_docs |
L1 |
| Research / analysis | OMC > deep-research (Exa + Firecrawl) | L2 |
| Recency research (last-30-days, social) | last30days plugin | L1 |
| Vault note ops / build docs as you go | obsidian-skills plugin | L1 |
| Scrape a bot-walled page | Scrapling MCP (Firecrawl fallback) | L1 |
| Query parquet / CSV / JSON columnar | DuckDB MCP | L1 |
| Build UI with design system | Copy DESIGN.md from the design-systems library + frontend-design skill |
L3 |
| Design-to-code | Figma MCP > get_design_context + Frontend Developer |
L3 |
| Word / PDF / PPT / Excel deliverable | Anthropic skill (docx / pdf / pptx / xlsx) — auto-triggers on file type |
L1 |
| Build an MCP server | mcp-builder skill + OMC > ralplan > Backend Architect |
L3 |
| Build / edit / benchmark a Claude skill | skill-creator skill |
L2 |
| Financial model — DCF / LBO / comps / 3-statement | model-builder or financial-analysis + xlsx skill |
L3 |
| Pitch deck end-to-end | pitch-agent + pptx skill |
L3 |
| Earnings analysis / model update | earnings-reviewer + xlsx / docx |
L2 |
| Sector research → peer comps | market-researcher (or deep-research for broader scope) |
L2 |
| SQL / dashboards over a dataset | Knowledge Work > data plugin + DuckDB / Supabase |
L2 |
| Contract / NDA triage | Knowledge Work > legal plugin |
L2 |
| Cross-tool search (email / chat / docs) | Knowledge Work > enterprise-search plugin |
L1 |
| Video from a URL (product tour, promo, social ad) | Hyperframes > website-to-hyperframes |
L2 |
| Motion-graphics video (explainer, chart race, captions+TTS) | Hyperframes > hyperframes + hyperframes-media |
L2 |
| Quick bug fix | Direct execution | L1 |
| Trivial task (no overhead) | GSD > /gsd-fast or Direct execution | L1 |
| Small surgical fix / "touch as little as possible" | OMC > autopilot > Minimal Change Engineer | L1 |
| Database work | Supabase MCP + Database Optimiser agent | L2 |
| Visual QA | Chrome DevTools MCP | L2 |
| Requirements unclear | OMC > /ccg framing pass in parallel with /deep-interview | L3 |
| Must be perfect | OMC > /ccg pre-gate → ralph | L4 |
| Debug (persistent across resets) | GSD > /gsd-debug | L3 |
| Root cause investigation | OMC > /oh-my-claudecode:deep-dive | L3 |
| Prove it works (evidence) | OMC > /oh-my-claudecode:verify | L2 |
| Research cross-content (code + papers + media) | Graphify index + OMC deep-research | L3 |
| Onboard onto doc-heavy codebase | Graphify + /gsd-map-codebase + Codebase Onboarding Engineer | L3 |
| Onboard onto a single, code-only repo | Understand Anything /understand + /understand-dashboard (trial) + /gsd-map-codebase |
L2 |
| Pre-refactor dependency check | Graphify query + lsp_find_references |
L3 |
| Overnight / unsupervised loop hits ambiguity | cortex-queue > enqueue_decision + await_decision |
L2 |
| Financial OHLCV forecast / quant baseline | Kronos > predict (+ backtest) | L2 |
| Fine-tune forecasting model on private data | Kronos > finetune | L3 |
| Financial paper / filing ingestion at scale | QuantMind (trial) | L2 |
| Cross-functional ops | OMC > autopilot > Chief of Staff | L2 |
| Multi-domain build (≥2 disciplines) | OMC > /team > [Specialist A ∥ Specialist B ∥ …] → reconciler | L3 |
| Cross-cutting audit (security + perf + a11y) | OMC > /team > parallel domain auditors → /ccg reconciler | L3 |
| Full-stack feature (API + UI + data) | OMC > /team > [Backend Architect ∥ Frontend Developer ∥ Database Optimiser] → /ccg reconciler | L3 |
The patterns that have proven valuable across many projects:
- CCG auto-firing on irreversible actions — the single biggest catch for "wait, are you sure" moments
- Sentinel-style checkpoint review — CCG validates the plan, a cross-family critic validates the diff. Two different failure modes, two different gates
- Domain breadth → multi-agent fan-out by default — prevents single-specialist tunnel vision. Shape is a separate decision from tier
- The anti-inflation rule — the log showed ~67% of early routes were L3, most of them reflexive bumps. L3 must name a concrete failure mode L2 would produce
Direct > batched-edits @ L2for tightly-scoped multi-file work — when routing overhead exceeds the actual work- Three confidences, not one — "high route, high tier, low spec" is the canonical signal to stop and interview before executing
- Self-learning log — at ~30 routes, two patterns auto-surfaced that were being done instinctively but never codified.
cortex hintstarted producing useful signal well before the theoretical threshold
Your patterns will be different. The framework is universal; the routes are personal.