Agentic Workflow Cheat Sheet — AnyCompany Support Workshop

The 4 Layers of AI Automation

Click any layer to learn what it does, when to use it, and see the analogy.

🏠

Steering

Global rules that apply to everything your AI does

📋

Skills

Instructions for specific tasks — your prompt templates, upgraded

🔔

Hooks

Auto-triggers — when X happens, do Y

🔌

MCP

Connections to external systems and data

🏠 Steering = House Rules
Rules that apply to every conversation, every task, every output. Like house rules for guests: "Default to SGD," "Never include real PAX / DAX / MEX names," "Flag refunds "Always use SGD," "Never include real customer names," "Flag amounts over $50K."gt;SGD 200 for TL review." You set these once, and they govern everything your AI does — no matter which skill is running.

How They Connect

The 4 layers work together as a stack. Click any node to see how data flows through the system.

🔔 Hook (Trigger): Something happens — a new file arrives, a user asks a question, a scheduled time is reached. The hook detects this event and kicks off the workflow. Like a doorbell: someone arrives, the system responds.

Which Do I Need? — Click a Scenario

Click any scenario to see which layers you need and why.

💬 "I want consistent AI responses across my team"

Everyone gets different quality from AI. You want standards.

→ Steering only. Write 5-7 global rules (currency, format, PII handling). Takes 5 minutes. Every AI interaction follows your standards.

📝 "I want AI to do a specific task really well"

You have a repeating task (risk assessment, report, review) and want reliable quality every time.

→ Steering + Skill. Your Day 1 prompt template becomes a skill. Add activation rules and guardrails. AI knows exactly what to do when asked.

⚡ "I want it to run automatically without me asking"

You don't want to manually trigger the AI — it should react to events.

→ Steering + Skill + Hook. The hook watches for a trigger (new file, new data, scheduled time). When it fires, the skill runs automatically. You only see the output.

🔌 "I want AI to read from our database / call our APIs"

The AI needs access to live data — not just what you paste in.

→ All 4 layers + tech team. MCP connects the agent to external systems. You design the steering + skills. Your tech team sets up the MCP connections. This is the full autonomous agent.

🏁 "I just want to start — what's the minimum?"

You're overwhelmed. What's the absolute simplest first step?

→ Steering + 1 Skill. That's it. Write your house rules (5 min) + convert your best prompt into a skill (15 min). You now have a reusable, consistent AI tool. Add hooks and MCP later when you're ready.

💬 "I want to manage conversations and control costs"

When to start fresh, how memory works across sessions, and how to keep token costs low.

→ See the Conversations & Costs section below — context windows, model limits, fresh start signals, memory hygiene, and team collaboration patterns.

Decision Guide

Quick reference — what each layer does, who owns it, and the analogy.

	Layer	What It Does	Analogy	Who Owns It	Effort
🏠	Steering	Global rules — currency, PII, format, escalation thresholds	House rules for every guest	You (business)	5 minutes
📋	Skills	Task-specific instructions — your prompt template + activation rules + guardrails	Recipe card for a specific dish	You (business)	15-30 minutes
🔔	Hooks	Auto-triggers — "when X happens, run Y"	Doorbell — someone arrives, action starts	You + tech team	5-10 minutes
🔌	MCP	Connections to databases, APIs, file systems	Kitchen equipment — oven, fridge, mixer	Tech team	Hours-days

💡 Key insight: You own the first two layers (steering + skills). That's where 80% of the value comes from. Hooks and MCP are the "nice to have" that your tech team can add later. Don't wait for MCP to get started.

⚠️ Common mistake: Trying to build all 4 layers at once. Start with steering + 1 skill. Prove it works. Then add complexity. The first step is always the hardest — make it small.

Getting Started — Your First 20 Minutes

The exact steps to go from "I have a good prompt" to "I have a working agent."

1️⃣

Write Your Steering (5 min)

Open a new Claude Cowork project. In the project instructions, write 5-7 rules:

• Default currency: SGD
• Never include real customer names
• Flag amounts over $50K
• Output format: use markdown tables
• Always show reasoning before conclusion

2️⃣

Convert Your Template to a Skill (15 min)

Take your Day 1 prompt template. Add:

• When to use: "Run this when asked to summarise an IRT case"
• Guardrails: "Never assign P1 without explicit SOP-trigger keywords"
• Output format: "Always include severity, booking, action taken, and SOP citation"

3️⃣

Test It (5 min)

Feed sample data into your project. Check:

• Does steering apply? (correct currency, no PII)
• Does the skill produce consistent output?
• Is the quality good enough to share?

If yes → you have a working agent. If no → adjust rules and test again.

🎯 That's it. Steering + 1 Skill = your first agent. It's not fancy, but it's consistent, reusable, and shareable. Everything else (hooks, MCP, multi-agent) is an upgrade you add when you're ready — not a prerequisite.

Where This Works — Claude Cowork Mapping

Here's exactly how each layer translates to the tool you're using today.

Layer	In Claude Cowork	How To Do It	Available?
🏠 Steering	Project Instructions	Write your global rules in the "Custom Instructions" field when you create a Project	✅ Today
📋 Skills	Project Knowledge + Structured Prompts	Upload your prompt template as a Knowledge file, or add structured instructions ("When asked about X, follow these steps...")	✅ Today
🔔 Hooks	Scheduled Tasks	Set a recurring schedule (daily/weekly/monthly) and Claude runs your prompt automatically. Time-based, not event-based — think "every Monday at 9am" rather than "when a file arrives."	✅ Partial
🔌 MCP	Connectors + Computer Use	Connectors link to Microsoft 365 (Outlook, OneDrive, SharePoint, Teams). Computer Use lets Claude interact with desktop apps directly. For live database/API access, your tech team uses Kiro IDE or Claude Code.	✅ Partial

💡 Good news: The two layers that deliver 80% of the value (Steering + Skills) work in Cowork today. Hooks and MCP have partial equivalents — Scheduled Tasks for automation, Connectors for integrations. For full event-driven hooks and custom API connections, your tech team uses developer tools like Kiro IDE or Claude Code.

How Different AI Tools Handle Automation

The same concept — "make AI run without me asking" — shows up differently across tools. Here's how hooks translate.

Concept	Claude Cowork	Kiro IDE / Claude Code	Key Difference
🔔 Auto-trigger	Scheduled Tasks Time-based: "Every Monday at 9am, generate my report"	Hooks Event-based: "When a file is saved, run lint"	Clock vs Event. Cowork runs on a schedule you set. Hooks react to things that happen in real-time.
🏠 Global rules	Project Instructions Custom instructions per project	Steering files .md files in .kiro/steering/	Same idea, different format. Both set "always-on" rules. Steering files can be conditional (only apply to certain file types).
📋 Task instructions	Knowledge files + Prompts Upload docs, write structured prompts	Skills Reusable prompt packages with activation rules	Same idea, more structure. Skills add "when to activate" and "what NOT to do" on top of the instructions.
🔌 External access	Connectors (M365, Google) Computer Use (desktop apps)	MCP servers Custom API/database connections	Pre-built vs custom. Cowork has ready-made connectors. MCP lets developers build connections to anything.

🎯 The takeaway: Different tools implement the same 4 concepts differently. Claude Cowork is designed for knowledge workers (you) — simpler, visual, pre-built. Developer tools like Kiro offer more power and flexibility but require technical setup. Start where you are, upgrade when you need to.

Conversations & Costs — Your Daily Reference

Context windows, when to start fresh, memory hygiene, and team collaboration. The practical guide to using Claude Cowork efficiently every day.

📐 Context Window — Know Your Model's Limit

Model	Context window	Practical implication
Claude Opus 4.7 (Cowork default)	1,000,000 tokens	~10–20 full working days. Hard limit rarely reached. Cost discipline still matters.
Claude Sonnet 4.6	200,000 tokens	~1–2 days of active work. Hit faster if pasting large documents.
Claude Haiku 4.5	200,000 tokens	Same — but cheaper per token, so cost management matters more here.
No progress bar in Cowork. Check the model name at the bottom of the conversation for the limit. Overflow signal: Claude gives inconsistent answers or "forgets" earlier context. In Claude Code CLI: type `/status` for exact % usage.

💵 Bedrock Pricing for Anthropic Claude — Singapore (ap-southeast-1)

All prices in USD per 1M tokens. Verified 2026-06-16 against Anthropic's authoritative pricing data and AWS docs. Cache write/read multipliers cross-verified on the Bedrock pricing page (1.25× / 2× / 0.1×).

Model	Input	Output	Cache Write 5min	Cache Write 1h	Cache Read
Claude Opus 4.7	$5.00	$25.00	$6.25	$10.00	$0.50
Claude Opus 4.6	$5.00	$25.00	$6.25	$10.00	$0.50
Claude Sonnet 4.6	$3.00	$15.00	$3.75	$6.00	$0.30
Claude Sonnet 4.5	$3.00	$15.00	$3.75	$6.00	$0.30
Claude Haiku 4.5	$1.00	$5.00	$1.25	$2.00	$0.10

⚠️ Singapore reality check (verified via aws bedrock list-foundation-models): Claude 4.x models are NOT available for direct in-region inference in ap-southeast-1. They route only via cross-region inference profiles like global.anthropic.claude-opus-4-7. Claude Opus 4.7 is hosted in US East (N. Virginia), Asia Pacific (Tokyo), Europe (Ireland), and Europe (Stockholm) — Singapore calls cross to one of these. Pricing applies at the source region rate; no surcharge for cross-region routing. For data-residency-sensitive workloads, this matters.

📊 Bedrock Pricing Tiers — VERIFIED Anthropic Support

Anthropic Claude models on Bedrock support only Standard, Reserved, and Batch. Flex (50% off) and Priority (75% premium) tiers are not available for Claude — they apply to OpenAI, DeepSeek, Qwen, and Amazon models only.

Tier	Price multiplier	Anthropic Claude support	When to use
Standard (On-Demand)	1.0×	✅ All Claude models	Default. Anything interactive.
Batch	0.5× (50% off)	✅ Opus 4.5/4.6, Sonnet 4.5/4.6, Haiku 4.5 ❌ Opus 4.7 (not yet supported)	Async, S3-based. Overnight bulk PAC tagging, weekly KB-gap reports.
Reserved	Fixed/month commitment	✅ Opus 4.5/4.6, Sonnet 4.5/4.6, Haiku 4.5	99.5% uptime SLA. Production workloads at scale (≥100K input TPM).
Flex / Priority	0.5× / 1.75×	❌ Not available for Claude	Reserved for OpenAI, DeepSeek, Qwen, Amazon Nova.

⚡ Effort Levels — How Cowork's Slider Maps to Cost

Cowork's "Effort" setting (Low / Medium / High / Max in the conversation footer) controls Claude's extended thinking — how many tokens Claude uses for internal reasoning before answering. Verified from the Anthropic SDK source: thinking tokens are billed as output tokens at the standard model rate. There is no surcharge for thinking — but on Opus 4.7 at $25/M output, the cost adds up fast.

Effort	Approx thinking tokens	Cost on Opus 4.7	Cost on Sonnet 4.6	When to use for GS Support
Low (default)	~0–1,024	~$0.025/msg	~$0.015/msg	99% of daily work — emails, summaries, simple extractions
Medium	~4,000–8,000	~$0.10–0.20/msg	~$0.06–0.12/msg	Risk assessments, narrative outputs needing nuance
High	~16,000–32,000	~$0.40–0.80/msg	~$0.24–0.48/msg	Complex regulatory analysis, credit committee briefs
Max	up to ~64,000+	~$1.50–2.00/msg	~$0.90–1.20/msg	Rarely justified. Major policy decisions, complex compliance scenarios.

⚠️ Verified from AWS docs — Opus 4.7 only supports Adaptive Thinking. The legacy thinking.type: "enabled" + budget_tokens API returns a 400 error on Opus 4.7. Cowork's Effort slider uses the new thinking.type: "adaptive" + output_config: {effort: "low|medium|high|max"} API. Most users won't notice — but if your tech team is migrating production code from Opus 4.6, they'll hit this.

💰 Worked example for the country GM: 1,000 invoices/month processed with the invoice-processor Skill. On Sonnet 4.6 at Low effort: ~$15/month. Same workload at Max effort: ~$1,200/month — 80× the cost for marginal quality gain on a routine task. The Effort slider matters most when batch-running many tasks. For one-off complex analyses, Max can be worth it.

🚨 The 5× Quota Burndown — A Bedrock Quirk Worth Knowing

Verified from AWS Bedrock docs: for all Claude 3.7 and later models (which means all Claude 4.x — Opus 4.5/4.6/4.7, Sonnet 4/4.5/4.6, Haiku 4.5), every output token deducts 5 tokens from your TPM/TPD throttle quota. Your bill is unchanged — but you'll hit ThrottlingException 5× sooner than naive math suggests.

      # AWS official worked example (Sonnet 4):

      Request: 1,000 input tokens + 100 output tokens

      Quota deducted: 1,000 + (100 × 5) = 1,500 tokens

      You're billed for: 1,100 tokens (no 5× on the bill)

💡 Two implications participants should know: (1) Set max_tokens deliberately — it's reserved upfront against the quota, so leaving it at the default cap reduces concurrency. (2) Cache reads do NOT count against TPM/TPD quotas. Combined with the 5× output burndown, prompt caching is disproportionately valuable on Claude 4.x — both saves money AND saves throttle headroom.

🗂️ Prompt Caching — Cost Savings the IT Team Should Configure

If your team repeatedly sends the same context (a long Project Instructions block, a frequently-referenced policy document, a Skill definition), prompt caching slashes the per-message cost dramatically. Configured by your tech team via the Bedrock API; you benefit transparently.

Model	Min cache size	5-min TTL	1-hour TTL
Claude Opus 4.7	4,096 tokens	✅	✅
Claude Opus 4.6 / 4.5	4,096 tokens	✅	✅
Claude Sonnet 4.6 / 4.5	1,024 tokens (4× lower threshold)	✅	✅
Claude Haiku 4.5	4,096 tokens	✅	✅
Verified all-models support both TTLs. Cache pricing: write 5-min = 1.25× input, write 1-hour = 2× input, read = 0.1× input (90% off). Max 4 cache breakpoints per request. Caching is on-demand only — cannot stack with the 50% Batch discount.

🎯 Real cost math: Caching your 5,000-token Project Instructions on Opus 4.7 — first request costs $0.03125 (cache write at 5-min); next 100 requests within 5 minutes cost $0.0025 each instead of the normal $0.025 each. That's a 90% discount on input tokens for repeated context. For 1,000 case summarisation runs/month, this turns a SGD 30 input-cost line into SGD 3.

Pricing data verified 2026-06-16. Source: bedrock-pricing-research.md (project root) — full sources include AWS Bedrock pricing page, per-model docs, Anthropic SDK, and live aws bedrock list-foundation-models output for ap-southeast-1.

🔄 When to Start a Fresh Conversation

Starting fresh resets the context window. Project Instructions and auto-memory still load — they're persistent. Two reasons to do it even on Opus 4.7: cost (every token is resent on every message) and task hygiene (mixed threads hurt quality).

Signal	Action
Switching to a different task	Start fresh. Old data wastes tokens.
Beginning a new day	Start fresh. Yesterday's context costs and rarely helps.
Just pasted a large PDF or CSV	Start fresh when done with it. Don't carry its tokens forward.
Conversation is 2+ hours old	Consider fresh start. First: "Summarise key decisions from this conversation."
Claude gives inconsistent answers	Start fresh immediately. Context is overflowing — continuing wastes money.

💡 Circuit breaker technique: Before starting fresh, ask Claude: "Summarise the key decisions and context from this conversation I should carry forward." Save that summary. Paste it into the new conversation if needed. Continuity without the cost.

🧹 Memory Hygiene — What Goes Where

Type of knowledge	Where it belongs	Why
Rules that never change (currency, PII, thresholds)	Project Instructions	Stable, you control it, always active
Facts about you (preferences, working style)	Auto-memory	Claude manages it, evolves over sessions
How to do a recurring task (case-summary, SOP-lookup)	Skill	Reusable template — not a memory fact
Data for today's work (case data, transcripts)	Inline — start fresh when done	Temporary — don't carry it forward
PAX / DAX / MEX names, phone numbers, booking IDs	Nowhere in Cowork	PII should never enter the AI context

Monthly memory habit: Ask Claude: "Summarise what you know about me and my projects from memory. Is anything outdated?" Fix stale entries. Ask Claude to "consolidate my memory — merge duplicates, remove outdated entries."

👥 Team Collaboration — The 3-File Pattern

Auto-memory is per account. Your colleague's Claude won't know what you decided — unless it's in a shared file in the mounted folder. Create these three files so both Claude instances work from the same picture:

File	What goes in it	Update when
`CONTEXT.md`	Current state only — active rules, thresholds, what's in progress, what's blocked	Any decision changes the project
`SESSION_LOG.md`	Append-only handoff — date, what changed, what's pending	End of every session: "Append today's session to SESSION_LOG.md"
`DECISIONS.md`	Decision registry — threshold changes, policy updates, rule changes with reason and owner	Whenever a rule or threshold changes

⚠️ Add this to every collaborator's Project Instructions: "At the start of each session, read CONTEXT.md from the connected folder. Treat it as authoritative if it conflicts with your memory." Claude then loads shared context automatically — no one has to remember to ask.