5 packs · 56 files · 34 skills

Foundations: the packs that run Corey itself

These packs are about Corey rather than your business: how it plans, builds, checks and ships work, how to keep it running on a machine you own, and how to build and improve the MCP servers and apps it is made of. You rarely ask for them by name. They are why a job gets checked before it is called done, and why the scheduled work still runs when you close the laptop.

How the work gets done

The AI SDLCDiscuss, Plan, Execute, Verify, Ship

Reference · 7 files · 2 sources · researched 09/09/2026 · review by 09/03/2027

What it covers

The loop - the five steps, the artefact each ends in, and what "done" means at each.
Acceptance criteria - criteria that can fail, and the one test that tells you.
Blind verification - why the verifier is denied the plan, and what to do when blind verification is impossible.
What a check proves - three ways a green check proves nothing, each of which costs a day.
Gates and escalation - pre-flight, revision, escalation, abort, named before the work rather than improvised after it.
Orchestration - the orchestrator routes and does not execute, and the contract every dispatched agent owes.
Context budget - why quality degrades before any warning fires, and when to checkpoint.

What Corey can do with it

Corey reads it whenever a job touches The AI SDLC, whichever skill is running.

Try asking

Talk me through the loop for The AI SDLC, and flag anything that changes often.
What do I need to get right about acceptance criteria before I start?

What it will not do

Do anything public, costly or irreversible without your yes.

Always-on Macrunning Corey's scheduled work on a machine you own

Reference · 12 files · 5 sources · researched 04/10/2026 · review by 04/04/2027

What it covers

Choosing a host - the three tiers, when a Mac beats the cloud, laptop or desktop, how much memory, and using an existing laptop as an interim host.
Setup playbook - the build, in order, with the check that proves each step. Start here.
Power and sleep - sleep, Low Power Mode, the battery charge limit, the lid, and starting again after a power cut.
Login session and restarts - what has to stay running (not a terminal), FileVault against auto-login, and the restart plan.
Scheduling with launchd - LaunchAgents over cron, the bare PATH that breaks every job, the wrapper every job should go through, and catch-up after sleep.
Claude on a host - signing Claude Code in for unattended use, the desktop app as the bridge, and why an exit code is not an outcome.
Remote access - the Claude app bridge first, keeping it alive, then NordVPN Meshnet (free) as the private device network, with the alternatives, Remote Login and the firewall.
Monitoring - the off-box dead-man, the host heartbeat, outcome alerts, and what to log while you decide on hardware.
And 3 more topics.

What Corey can do with it

Corey reads it whenever a job touches Always-on Mac, whichever skill is running.

Try asking

Talk me through choosing a host for Always-on Mac, and flag anything that changes often.
What do I need to get right about setup playbook before I start?

What it will not do

Quote a price, limit or threshold from memory. Figures that move are kept in a dated ledger and re-checked against the primary source first.
Do anything public, costly or irreversible without your yes.

MCP servers and apps

MCP serversbuilding and operating one

Reference · 11 files · 2 sources · researched 10/09/2026 · review by 10/03/2027

What it covers

Server shape - tools, resources and prompts, what each is for, and how to choose. The choice is a routing decision, not a taste one.
Tool design - names, schemas, titles, annotations, and the fact that the description is the routing surface. Includes the two name rules the SDK will not enforce for you.
Transports - stdio against streamable HTTP, session-bearing against stateless, and exactly what a per-request server forbids.
Auth at the boundary - where authentication belongs, why every gate fails closed, and the trap that the listing is a separate seam from the call.
Listing and visibility - tools/list is the per-request seam. Hiding is not a permission, and hiding the wrong tool is silent and total.
Errors and refusals - a refusal must carry the next action. One wording per condition, no stack traces, and why this is a large share of what a reviewer sees.
Testing a server - in-process, over stdio, and over the wire; and what a local pass does not prove.
Deployment - the Host and Origin allowlist that 421s every proxied request when you get it wrong, plus readiness and the deploy gate.
And 2 more topics.

What Corey can do with it

Mcp app builder - Take an idea to a shipped MCP App - interactive HTML a server serves as a ui:// resource and a host renders inside the conversation. Decides whether it earns a UI at all, then stamps, builds and hands off to verification.
Mcp app verify - Prove a built MCP App actually works - static conformance, a protocol handshake against the reference host, and a rendered check in a real client - and emit a scorecard rather than an opinion.
Mcp lifecycle - Build an MCP server or MCP App end to end so it ships instrumented, eval-tested and ready for the weekly review, or audit an existing one against that bar.
Mcp weekly review - Weekly improvement pass across every MCP server the owner runs - reads usage signals, gaps and feedback, measures last week's changes, researches the top findings, and opens draft PRs and issues to add, rewrite, remove or research tools and knowledge files.

Try asking

Build an MCP App
Verify the MCP App
Build an MCP

What it will not do

Quote a price, limit or threshold from memory. Figures that move are kept in a dated ledger and re-checked against the primary source first.
Do anything public, costly or irreversible without your yes.

MCP Appsinteractive UI for MCP servers

Reference · 13 files · 2 sources · researched 04/09/2026 · review by 04/03/2027

What it covers

Protocol and resource contract - the extension id, the ui:// scheme, the MIME type, capability negotiation, and why the text fallback is mandatory rather than polite.
Meta fields - the whole _meta.ui surface, and the thing everyone gets wrong: which _meta object each field belongs to.
Jsonrpc surface - every ui/* request and notification in both directions, with what each is actually for.
App lifecycle - the App class, the handlers-before-connect rule, teardown, and the sandbox proxy.
Host context and theming - the CSS-variable contract, theme, fonts, safe areas, display modes, and how to look native in a host you cannot see.
Security csp and auth - the deny-by-default CSP model, CORS and stable origins, authorization at the HTTP boundary, step-up auth, and four questions the spec does not answer.
Patterns - app-only tools, polling, chunked reads, streaming partial input, model-context updates, view state, and pausing offscreen work.
Serverless deployment - the one nobody else documents. Why the official examples cannot run on a serverless runtime, and what to do instead.
And 4 more topics.

What Corey can do with it

Mcp app builder - Take an idea to a shipped MCP App - interactive HTML a server serves as a ui:// resource and a host renders inside the conversation. Decides whether it earns a UI at all, then stamps, builds and hands off to verification.
Mcp app verify - Prove a built MCP App actually works - static conformance, a protocol handshake against the reference host, and a rendered check in a real client - and emit a scorecard rather than an opinion.
Mcp lifecycle - Build an MCP server or MCP App end to end so it ships instrumented, eval-tested and ready for the weekly review, or audit an existing one against that bar.
Mcp weekly review - Weekly improvement pass across every MCP server the owner runs - reads usage signals, gaps and feedback, measures last week's changes, researches the top findings, and opens draft PRs and issues to add, rewrite, remove or research tools and knowledge files.

Try asking

Build an MCP App
Verify the MCP App
Build an MCP

What it will not do

Never branch on getUiCapability(). It cannot work on a server that builds a fresh instance per request, and it fails with no error at all.
Quote a price, limit or threshold from memory. Figures that move are kept in a dated ledger and re-checked against the primary source first.
Do anything public, costly or irreversible without your yes.

MCP operationsinstrumenting, analysing and improving a live MCP server

Reference · 13 files · 5 sources · researched 29/09/2026 · review by 29/12/2026

What it covers

App telemetry
Customer telemetry
Definition of done
Event contract
Health and slos
Instrumentation
Intent and outcomes
Knowledge standard
And 4 more topics.

What Corey can do with it

Mcp app builder - Take an idea to a shipped MCP App - interactive HTML a server serves as a ui:// resource and a host renders inside the conversation. Decides whether it earns a UI at all, then stamps, builds and hands off to verification.
Mcp app verify - Prove a built MCP App actually works - static conformance, a protocol handshake against the reference host, and a rendered check in a real client - and emit a scorecard rather than an opinion.
Mcp lifecycle - Build an MCP server or MCP App end to end so it ships instrumented, eval-tested and ready for the weekly review, or audit an existing one against that bar.
Mcp weekly review - Weekly improvement pass across every MCP server the owner runs - reads usage signals, gaps and feedback, measures last week's changes, researches the top findings, and opens draft PRs and issues to add, rewrite, remove or research tools and knowledge files.

Try asking

Build an MCP App
Verify the MCP App
Build an MCP

What it will not do

Quote a price, limit or threshold from memory. Figures that move are kept in a dated ledger and re-checked against the primary source first.
Do anything public, costly or irreversible without your yes.

The running corey skills

Skills are the step-by-step jobs that lean on these packs. You start one by asking for the job in plain English.

Approval packBuild a human-approval pack for a finished deliverable: a review page with screenshots or a rendered preview, a plain-English explanation, a comment thread with Approve / Request- changes, and a linked implementation issue.
Assign cabsPopulate team.md with the Accountability Map - every seat and the Critical Accountability Bullets (CABs) it owns, grouped by department - by walking the value engines and asking "who does this?" for each critical system.
AuditScore an OS install against the 4 C's (Context, Connections, Capabilities, Cadence) and report the gaps, naming the one to close next and the skill that closes it. Read-only; reports, never fixes.
BrainFreeform brain-dump capture plus the compounding page memory layer: save a stream-of- consciousness thought verbatim, dissect it into discrete units, and fan each out to the right destination.
Build scorecardDraft a Project's scorecard to the Rule of Three, back-test each metric against recent history, set monthly targets, and wire manual weekly status.
ClassifierThe single layer-routing decision for the OS.
Claude pushbackPush back, ask one focused question, or propose a better path when the owner's request is ambiguous, contradicts a prior decision, or has a single high-cost interpretation.
Claude solutionsForce every problem you surface into a solution shape before you send it: a finding, a single recommended path with its reason, and the action you have taken or are about to take.
Command centerBuild a tailored Command Center (a personal control-center dashboard) for a business, discovery-driven from what that business actually does.
Company hqBuild and publish the read-only Company HQ surface: a sidebar-driven static command centre rendered from generated snapshots, with a data-health signal and a gap-fill interview loop.
FeedbackSend feedback about the engine itself up the chain to the vendor.
File execution ticketDispatch a code change to the right execution lane - direct commit, async coding agent, browser IDE, or desktop editor - and file the ticket with the right template and label.
Get idea goingThe front door for "I have an idea" and "help me build this" - takes a raw idea, routes it by shape, and stands it up one step at a time. A venture goes to evaluate, scaffold and spec; a capability - one tool a model calls and a user looks at - goes to mcp-app-builder, which decides whether it earns a view at all.
Grounded answerProduce output grounded in the business's own knowledge base with a required citation.
Host checkCheck that a Mac is set up to run Corey's scheduled work unattended, and say what is missing for that machine and who has to fix it. Read-only.
IdeasThree-mode skill for the business's idea pipeline.
Level upThe operating loop that surfaces the single next automation. Asks five questions, one at a time, then proposes the one highest-leverage thing to automate next and routes it. Read-only; proposes, never builds.
Map value engineDraft a Value Engine for a Project - the ordered stages by which it wins work and delivers it - and mark its Power Stages.
Mcp app builderTake an idea to a shipped MCP App - interactive HTML a server serves as a ui:// resource and a host renders inside the conversation. Decides whether it earns a UI at all, then stamps, builds and hands off to verification.
Mcp app verifyProve a built MCP App actually works - static conformance, a protocol handshake against the reference host, and a rendered check in a real client - and emit a scorecard rather than an opinion.
Mcp lifecycleBuild an MCP server or MCP App end to end so it ships instrumented, eval-tested and ready for the weekly review, or audit an existing one against that bar.
Mcp weekly reviewWeekly improvement pass across every MCP server the owner runs - reads usage signals, gaps and feedback, measures last week's changes, researches the top findings, and opens draft PRs and issues to add, rewrite, remove or research tools and knowledge files.
New projectAdd a new Project to the owner's workspace the canonical way - questions first, then provision end-to-end.
New ventureTake a raw business idea from "here's my idea, what do you think?" to a scaffolded, GitHub- backed venture ready to build.
OnboardThe first-run "Meet Corey" onboarding interview. The primary entry point is the "Get Started" prompt; also.
ProfileBuild the owner's professional profile from their career data, and tailor it to a specific role or job application.
ReorganiseRepair and tidy an existing Corey setup that drifted or was never structured properly - the fix for a folder where Corey "just started working" with no spine.
Scaffold projectCreate a new Scalable-OS Project folder tree to the canonical schema - the OS-layer slots, the domain layer, and the working folders, each with an _index.md front door, initialised as its own git repo.
Setup businessThe "set up a business" build - the progressive company context builder Corey offers after onboarding.
Setup guideWrite or update a for-dummies setup guide for getting a tool, service, account, or integration working - Hetzner, Doppler, a new API, a new account.
Skill builderThe meta-skill that turns a process the owner already does by hand into a reusable skill: it interviews them, then stamps and verifies a generic skill folder from the engine's own templates. Lets an owner grow the OS without the author.
ThoughtDraft a short-form thought (one sharp post - the kind for LinkedIn, X, or a microblog) in the owner's brand voice.
Verify setupVerify a Corey setup is healthy and safely backed up - the confidence check for a customer's OS.
Write playbookDocument one Power Stage as a playbook via the Three Ds (Define the outcome, Design the steps, Deploy to a named Owner with a 90-day review), using the Extract capture technique so the steps come from the person who does the work, not from memory.

Questions

5 knowledge packs, 56 files in all: how the work gets done (The AI SDLC, Always-on Mac); mcp servers and apps (MCP servers, MCP Apps, MCP operations).

No. Ask for the job and Corey reads the packs it touches. You can name one to be sure, for example "use your The AI SDLC pack for this".

Every pack shows the date it was researched and the date it must be reviewed by (the earliest here is 29/12/2026). Figures that move fast are kept in a dated changes ledger and re-checked against the primary source before Corey quotes them.

No. The packs are reference material. Corey uses them to do the work and explain it, and anything public, costly or irreversible waits for your approval.

Where next

Corey is free for individuals and takes about two minutes to add to Claude.

Get started →