The Claude Cowork Stack

Project Instructions, Skills, Scheduled Tasks, Plugins & MCP — when to use each layer and how they work together

Day 2: Agentic AI Interactive 5 Layers Claude Cowork

🏗️ Why a Stack?

Building an AI-powered workflow in Claude Cowork isn't one thing — it's layers. Each layer solves a different problem. You can use one layer alone, or combine all five for a fully autonomous agent.

📏

Project Instructions

Global rules that apply to every conversation in your project. Like a standing brief you never have to repeat.

🧠

Skills

Saved prompt templates that auto-activate when the task matches. Like a specialist you call in by name.

⏰

Scheduled Tasks

Auto-triggers that run on a schedule or event. Like setting a recurring alarm for your agent.

🤖

Agents

Autonomous sub-tasks Claude delegates to. Like handing off a piece of work to a focused colleague.

💡
The key insight: Project Instructions set the rules. Skills define the expertise. Scheduled Tasks trigger automatically. Agents work independently. Plugins/MCP connects to your data. Together = an autonomous workflow.

🎯 The Stack at a Glance

📏
Project Instructions
Always-on rules
🧠
Skills
Saved templates
⏰
Scheduled Tasks
Auto-triggers
🤖
Agents
Delegated workers
↕️
🔌
MCP — Model Context Protocol
Connects AI to databases, APIs, and files

🔍 Explore Each Layer

Click a layer to see details, file location, when to use it, and a real example.

📏
Project Instructions
🧠
Skills
⏰
Scheduled Tasks
🤖
Agents
🔌
Plugins / MCP
🔄
If you've used Kiro IDE before: the same architecture, different names.
Kiro Claude Cowork What it does
.kiro/steering/*.mdProject InstructionsAlways-on global rules
.kiro/skills/*/SKILL.mdSaved SkillsReusable prompt templates
.kiro/hooks/*.jsonScheduled TasksAuto-triggers on schedule or event
.kiro/agents/*.mdSub-agents (Agent tool)Delegated parallel workers
.kiro/settings/mcp.jsonPlugins / MCPConnections to external data & tools

📏 Project Instructions — Global Rules

Project → Project Instructions (notepad icon)

Project Instructions are always-on rules applied to every conversation inside your Cowork project. They're like a standing brief — you write them once and Claude follows them automatically without you ever repeating them.

When to use:

  • Rules that apply to ALL tasks (currency, PII, company name)
  • Formatting standards your team always follows
  • Compliance constraints that never change
  • Context that every conversation needs (market, audience)

When NOT to use:

  • Task-specific instructions (use Skills instead)
  • Large reference documents (use manual-inclusion steering)
  • Anything that only applies sometimes
Example: 
---
inclusion: always
---
# AnyCompany Rules
1. All amounts in SGD unless specified
2. Never include PII — use [REDACTED]
3. Risk ratings: GREEN/AMBER/RED only
4. Cite specific data for every claim

🧠 vs 🔌 Skills vs MCP — When to Use Which?

This is the most common question on Day 3. Skills and MCP both make AI more capable — but they do completely different things. Here's the clearest way to think about it:

🧠

Skills = What to say

Skills tell the AI how to think and what to produce. They're prompt templates with personas, output formats, guardrails, and decision rules.

Analogy: A recipe that tells the chef what dish to make, what ingredients to use, and how to plate it.
🔌

MCP = Where to get data

MCP gives the AI access to external data and systems. It connects to databases, APIs, and files so the AI can query real data instead of relying on what you paste in.

Analogy: The pantry and fridge that the chef pulls ingredients from. Without it, the chef can only cook with what you hand them.

🎧 AnyCompany Example: Case Context Summarizer

Watch how Skills and MCP play different roles in the same workflow:

StepWhat happensPowered byWhy
1. Get dataQuery D365 + datalake for case + booking + Pax/Dax history🔌 MCPAI needs live access to case data it can't see otherwise
2. AnalyseApply the structured Case Summary template (Symptom · Severity · Booking · Action · Next Step) with SOP citations🧠 SkillAI needs instructions on what format and SOP rules to follow
3. Cross-checkQuery open_safety_flags table for any prior unresolved Dax / Pax safety records🔌 MCPAI needs additional context from a different datalake table
4. DecideApply SOP rules: explicit DUI keyword = P1 auto-trigger, refund >SGD 200 = TL escalate, ambiguous = flag for human🧠 SkillAI needs the SOP rules to make the right call
5. ReportGenerate the structured case summary with SOP citations and post back to D365 case notes🧠 SkillAI needs the output template to match QA-audit format
💡
The pattern: MCP fetches the data. Skills process the data. They alternate throughout a workflow — MCP brings ingredients, Skills cook the meal.

❓ Quick Decision: Skill or MCP?

If you need the AI to...UseExample
Follow a specific output format🧠 Skill8-section risk report with GREEN/AMBER/RED
Query a database🔌 MCPPull D365 case + Pax history from datalake
Use a specific persona🧠 Skill"You are a senior IRT Team Lead briefing leadership..."
Read files from a folder🔌 MCPScan sops/ directory for updated SOP articles
Apply business rules and guardrails🧠 SkillDUI keyword = P1 auto, refund >SGD 200 = TL escalate
Send a Slack message or call an API🔌 MCPPost alert to #irt-safety-escalations channel
Generate a structured report🧠 SkillCredit committee narrative with bull/bear case
Access live data it can't see🔌 MCPCheck compliance status in real-time
🔑
One-liner: Skills = the brain (how to think). MCP = the hands (how to reach data). You almost always need both for a production workflow. A skill without MCP can only work with data you paste in. MCP without a skill gives raw data with no structure.

🧭 Which Layer Do I Need?

Answer the questions below to find the right layer for your use case.

Does this rule/behavior apply to EVERY conversation, or only specific tasks?
Every conversation
It's a global standard
Specific tasks only
Only when doing certain work
🔄
Try different paths! Click through the decision tree multiple times to see how different needs map to different layers. The tree resets when you click "Start over".

🎬 Real-World Scenarios

See how the layers combine for common AnyCompany Support workflows. Click a scenario to see the breakdown.

A new IRT case arrives. The system pulls case + booking + Pax history, drafts the structured stakeholder brief, and posts back to D365.

📊 Side-by-Side Comparison

AspectSteeringSkillsHooksAgentsMCP
ScopeGlobal — all tasksOn-demand — specific tasksEvent-driven — automaticDelegated — independentData layer — connections
ActivationAlways loaded in projectAuto-activates when task matchesFires on schedule or eventSpawned for parallel tasksAvailable when plugin installed
Where it livesProject Instructions fieldYour Cowork accountScheduled Tasks panelWithin active conversationPlugins marketplace
Who sets it upYou (business user)You (business user)You (business user)Claude (automatic)IT team + you activate
AnalogyStanding brief / house rulesRecipe cardKitchen timer / doorbellTeam member you delegate toKitchen equipment
Best forStandards, constraints, contextReusable structured outputsAutomation & schedulingReview, validation, parallel workLive data access
AnyCompany Support example"Cite SOP article ID; redact PAX names"Case Summarizer skill"Every hour, run the KB-gap report"Quality-check reviewer skillD365 / datalake / SOP corpus
💡
Key distinction — Skills vs Plugins: A Skill tells Claude how to work (the recipe). A Plugin gives Claude access to data (the ingredients). You write the Skills — that's your domain expertise. Your IT team installs the Plugins — that's the plumbing. Both are needed for a fully automated workflow.

🔗 The Full Pipeline

Here's how all 5 layers work together in a complete autonomous workflow:

1
Project Instructions load global rules
Every conversation starts with your market's standards already active: default currency, no PAX/DAX/MEX names, cite SOP article ID, escalate ambiguous to TL.
Project Instructions
2
Scheduled Task fires automatically
It's Monday 8am. The scheduled task fires: "Check the IRT case queue for last week's escalations and run the case-summarizer skill on each one — produce a KB-gap report by Friday."
Scheduled Task
3
Skill activates with the saved template
Claude recognises "case-summarizer" and applies the saved skill — the structured template (Symptom · Severity · Booking · Action · Next Step) with SOP citation rules, severity threshold logic, and ambiguity flags.
Skill
4
Plugin reads files from connected folder
The D365 / datalake Plugin reads the IRT case queue. Cases are pulled directly with full transcripts and booking history — no manual download or copy-paste needed.
Plugin / MCP
5
Sub-agent runs quality check
Claude spawns a quality-check sub-agent that independently verifies: SOP citation correct? Severity matches the rule? PII redacted? Returns PASS or FAIL with reasons before the summary is written to D365.
Agent
✓
Result: Verified exception report, automatically
A complete, quality-checked weekly KB-gap report — delivered every Monday morning without anyone asking. The TQM team sees only the cases that surface SOP gaps worth fixing.
🎯
You don't need all 5 layers to start. Begin with Project Instructions alone (5 minutes). Add a Skill when you have a template worth saving. Add a Scheduled Task when you want it to run automatically. Ask your IT team about Plugins when you need live data access. Build up one layer at a time.

🧠 How Claude Cowork Remembers — Two Mechanisms

Most AI chat tools treat every conversation as a blank slate — you ask a question, it answers, and when you close the window everything is gone. Claude Cowork works differently. It has two persistent memory layers that load at the start of every conversation, so Claude already knows your rules and your context before you type a single word.

Understanding both layers — and the important difference between them — determines how you structure your work and control your costs.

📏 You write this

Project Instructions

Static — you write it once, Claude follows it in every conversation.

The standing brief for your project. Think of it as the briefing document you hand to any new colleague before they start work — the rules that apply no matter what task they're doing.

  • Always-on global rules (currency, PII, escalation thresholds)
  • Your role and team context
  • Output style and format standards
  • Compliance constraints that never change
You are in full control. Edit it anytime. It only changes when you change it.
🤖 Claude builds this

Auto-Memory

Dynamic — Claude writes individual facts during conversations, when it judges something worth preserving.

This is not an automatic session summary. Claude doesn't log everything — it makes selective judgment calls: "this preference, decision, or context fact is durable enough to carry forward." Most conversations produce no memory writes at all.

  • Your preferences and working style ("always use tables, not bullets")
  • Project decisions and context ("invoice threshold changed to SGD 30K")
  • Feedback patterns ("CFO prefers one-page summaries")
  • Team structure and stakeholder context
Fully editable. Memory files are plain Markdown on your computer — read, edit, delete, or ask Claude to update them anytime. You stay in control of what Claude carries forward.
📓
The notebook analogy: Project Instructions is the printed rulebook pinned to your office wall — it's always visible, it doesn't change unless you reprint it. Auto-memory is the notebook on your desk — Claude jots things down as it learns, and picks up that notebook each time you start a new conversation. Both are always in the room. The rulebook sets the standards. The notebook tracks what Claude has learned about you specifically.
💡
The practical difference between the two: If you want Claude to always use SGD — put it in Project Instructions (stable rule). If Claude has noticed you prefer tables over bullet points from past sessions — that goes in auto-memory (learned preference). You manage the rules. Claude manages what it has learned.

✏️ Working With Memory — Read, Edit, Delete

A common question: "Can I control what Claude remembers? Can I remove something it stored?" Yes — fully. Memory files are plain Markdown on your computer, and Claude can read and modify them on your instruction.

What you want to doHow to do itExample
See what Claude knowsAsk Claude directly"Summarise what you know about me and this project from memory."
Add a fact explicitlyTell Claude to remember it"Remember: our AP escalation threshold changed to SGD 50K from 1 July."
Correct a stale factTell Claude to update"Update your memory — the escalation threshold is now SGD 50K, not SGD 30K."
Remove somethingTell Claude to forget it"Forget the preference about bullet points — I've changed my mind."
Edit a file directlyOpen the .md file in any text editorMemory files live in a /memory/ folder — plain Markdown with frontmatter. Delete a line, save the file.
Clean up accumulated clutterAsk Claude to consolidate"Consolidate my memory — merge duplicates, remove anything outdated, keep it concise."
⚠️
What auto-memory does NOT do automatically:
  • It does not create a summary of every conversation
  • It does not log your project progress session by session
  • It does not track action items or decisions unless Claude judges them worth preserving
  • Most conversations produce no memory writes at all
If you want a project log, decision record, or session summary — ask Claude to write one explicitly and save it. Don't assume it's happening automatically.
📋
Practical pattern for project teams: At the end of a significant session — a key decision made, a policy updated, a stakeholder conversation — take 30 seconds to say: "Save to memory: we decided [X] today because [Y]. This replaces our previous approach of [Z]." This turns auto-memory into an intentional project log, not just a passive preference tracker.
📋
Conversation management, token costs, when to start fresh, memory hygiene, and team collaboration patterns are covered in the Agentic Workflow Cheat Sheet → Conversations & Costs — the daily reference card.