3.8 · Permissions, Usage & Chọn Model

Mở đầu: Với power comes responsibility

• Read/write real files on your disk
• Connect to real systems — email, CRM, bank data, PHI (healthcare), PII (personal info)
• Run code in isolated environments
• Send real emails (nếu bạn grant)
• Delete files (với approval gate)

Đây là order of magnitude mạnh hơn chat. Cũng nghĩa là: attention to boundaries matter more.

Bài này cover 3 chủ đề trust-related:

┌────────────────────────────────────────────────────┐
│                                                    │
│   1. 🛡  PERMISSIONS & SAFETY                       │
│      Boundaries Cowork operates within             │
│                                                    │
│   2. 📊 USAGE ALLOCATION                            │
│      Where your compute budget goes                │
│                                                    │
│   3. 🧠 MODEL SELECTION                             │
│      Opus/Sonnet/Haiku — match task to tier        │
│                                                    │
└────────────────────────────────────────────────────┘

Plus: review habit that ties them all together.

Phần 1: Permissions & Safety

4 safety boundaries

┌────────────────────────────────────────────────────────┐
│                                                        │
│   🔒 1. ISOLATED EXECUTION                              │
│       Work runs trong isolated environment             │
│       Separate from your OS                            │
│       Không reach thứ chưa granted                     │
│                                                        │
│   📁 2. CONTROLLED FILE ACCESS                          │
│       You decide which folders Cowork sees             │
│       No grant = no access                             │
│                                                        │
│   🌐 3. NETWORK POLICIES RESPECTED                      │
│       Follows your organization's network rules        │
│       Restricted environments stay restricted          │
│                                                        │
│   🗑 4. DELETION IS GATED                               │
│       Permanent delete requires your approval          │
│       Always prompt first                              │
│                                                        │
└────────────────────────────────────────────────────────┘

Isolated execution — What it means concretely

Cowork's work happens trong a sandbox:

┌──────────────────────────────────────────────────┐
│  YOUR OS (main system — protected)               │
│                                                  │
│  ┌────────────────────────────────────────┐     │
│  │  COWORK SANDBOX                         │     │
│  │                                         │     │
│  │  ├── Approved folder (read/write)      │     │
│  │  ├── Connected tools (via OAuth scopes)│     │
│  │  ├── Isolated compute (code runs here) │     │
│  │  └── Network (subject to your policies)│     │
│  └────────────────────────────────────────┘     │
│                                                  │
│  ❌ Cowork cannot touch:                          │
│  • System files (OS internals)                   │
│  • Other folders not granted                     │
│  • Apps not via connector                        │
│  • Network endpoints outside your policies       │
└──────────────────────────────────────────────────┘

Practically: Cowork chạy Python code để process data → code chạy isolated, không impact Python environment của bạn. Installed packages riêng.

File access — Grant granularity

Folder access granted TO Cowork:
  ✅ That folder's files (read + write)
  ✅ That folder's subfolders
  ❌ Parent folder
  ❌ Sibling folders
  ❌ Root filesystem

Best practice: Never grant root (/, C:\). Always grant specific project folders.

Network policies

Nếu công ty bạn có corporate firewall, VPN, Zero Trust Network Access — Cowork respects all of it. Không bypass. Network request Cowork makes goes through same egress rules như browser của bạn.

Implication: Nếu IT whitelist *.anthropic.com, Cowork work. Nếu không, Cowork connector setup fails với clear error.

Deletion gating

Cowork không tự delete files permanently. Khi delete needed:

⚠️  Cowork đề xuất delete:

   - old-draft-v1.docx  (replaced by v2)
   - temp-export.csv    (processed, safe to remove)
   - screenshots/       (unused, 15 files)

   Proceed with delete? [Yes, delete all] [Review each] [Cancel]

Bạn luôn có final approval.

Conversation history storage

Implication:

• Conversations không leaked to cloud (trong default config)
• Easy backup, restore (copy Claude Desktop data folder)
• If laptop compromised, conversations accessible locally
• Disk encryption matters (FileVault / BitLocker) cho sensitive work

Audit logging & compliance

Nếu bạn có workloads với regulatory requirements (HIPAA, GDPR, SOC2, PCI-DSS...):

• Check plan's documentation cho current audit logging features
• Team & Enterprise plans có enhanced compliance controls
• Ask admin về audit trail config nếu cần

Phần 2: Usage Allocation

Why Cowork uses more hơn chat

Cowork inherently more compute-intensive vì:

• Multi-step tasks (mỗi step = API call)
• Long-running work (10-30 phút runtime typical)
• Subagents parallelize (N times compute)
• Reading nhiều files (tokens)
• Generating structured outputs (more tokens than chat reply)

Quantifier: Single Cowork task có thể tốn 5-50× compute của chat turn. Not linear to length — depends trên structure.

Good habits cho allocation management

┌────────────────────────────────────────────────────┐
│                                                    │
│   1. 📦 BATCH RELATED WORK                          │
│      Starting fresh session có overhead            │
│      Several related tasks → 1 session             │
│                                                    │
│   2. 💬 USE CHAT FOR TASKS THAT FIT                 │
│      Task không cần files/tools/real output?       │
│      Chat faster + cheaper                         │
│                                                    │
│   3. 👀 MONITOR WHERE YOU STAND                     │
│      Cowork settings include usage visibility      │
│      Check periodically, especially building habits│
│                                                    │
└────────────────────────────────────────────────────┘

Habit #1: Batch related work

Not ideal:

Task 1 trong new session → finish
Close, reopen
Task 2 trong new session → finish  
Close, reopen
Task 3 trong new session → finish

Mỗi new session có cold-start cost: load context, parse prompts, warm up.

Better:

New session:
  Task 1 → wait output
  Task 2 (follow-up) → wait output
  Task 3 (related) → wait output
End session.

Warm context thread saves overhead.

Habit #2: The right tool for each shape

Recall decision rule từ Bài 3.0:

"Does this task need to touch your files, your connected tools, or produce a real output file?"

• Yes → Cowork
• No → Chat

Common misuse: Using Cowork cho:

• Brainstorm session ("help me think about strategy")
• Quick Q&A ("what's the syntax for X?")
• Writing refinement ("make this paragraph punchier")
• Explain concept ("how does vectorization work?")

All of above are chat-shaped. Cowork does them but wastefully.

Habit #3: Monitor usage

Where to find:

Settings → Cowork → Usage

Weekly allocation: 100 units
Used: 67 units (67%)
Reset: Friday 23:59 UTC

Track trend:

• Spike weeks → identify which tasks consumed most
• Steady weeks → comfortable, can take on more ambitious work

Common allocation sinks

┌────────────────────────────────────────────────────┐
│                                                    │
│   🚨 HIGH-COST PATTERNS (watch out for)             │
│                                                    │
│   • Running Opus cho simple tasks                  │
│   • Scheduled tasks cadence quá tight (every 15m)  │
│   • Huge file/folder ingestion with vague prompt   │
│   • Re-running failed tasks without fix           │
│   • Subagent spawn cho sequential task             │
│                                                    │
└────────────────────────────────────────────────────┘

Low-cost alternatives

Phần 3: Model Selection

Three tiers of Claude

Claude models come in three capability tiers, trade-off capability vs cost:

┌──────────────────────────────────────────────────────┐
│                                                      │
│   🏔  OPUS                                            │
│      Most capable cho complex multi-step work        │
│      Uses most allocation                            │
│      Best for: deep reasoning, architectural tasks,  │
│                 research synthesis, complex coding   │
│                                                      │
│   🏃 SONNET                                           │
│      Middle ground — sensible default                │
│      Good balance capability/cost                    │
│      Best for: most everyday tasks, file ops,        │
│                 standard prompting                   │
│                                                      │
│   ⚡ HAIKU                                            │
│      Fastest and lightest                            │
│      Lowest compute cost                             │
│      Best for: simple extraction, formatting,        │
│                 quick Q&A, bulk repetitive work      │
│                                                      │
└──────────────────────────────────────────────────────┘

Decision framework: Match model to task

┌─────────────────────────────────────────────────────┐
│                                                     │
│   TASK SHAPE              →    RECOMMENDED MODEL    │
│                                                     │
│   "Quick extract/format"  →    Haiku                │
│   "Standard analysis"     →    Sonnet               │
│   "Complex reasoning"     →    Opus                 │
│   "Multi-step research"   →    Opus                 │
│   "Creative writing"      →    Sonnet/Opus          │
│   "Bulk processing"       →    Haiku (parallel)     │
│                                                     │
└─────────────────────────────────────────────────────┘

Concrete examples per tier

✅ Haiku fits:

• "Extract invoice # from 100 PDFs"
• "Format these 50 emails consistently"
• "What's the syntax for SQL inner join?"
• "Summary of 1-page document"
• "Classify 500 tickets by category"

✅ Sonnet fits:

• "Review this contract against template, flag differences"
• "Draft weekly report from data + Slack + emails"
• "Create project tracker from meeting notes"
• "Explain this regulation and implications"
• "Refactor this code module"

✅ Opus fits:

• "Synthesize literature review from 40 papers with contradictions + research gaps"
• "Architect a new data pipeline, considering 5 constraints + 3 trade-offs"
• "Build comprehensive financial model với sensitivity analysis"
• "Legal brief with case law research and argument structure"
• "Debug this subtle concurrency issue"

Switch model trong Cowork

Top of Cowork window hoặc settings:

┌─────────────────────────────────────┐
│   Model: [Claude Sonnet 4.6   ▼]    │
│   ✓      Claude Sonnet 4.6 (default)│
│          Claude Opus 4.7            │
│          Claude Haiku 4.5           │
└─────────────────────────────────────┘

Change per task hoặc set default trong Settings.

Model comparison — What actually differs

"Default to highest" anti-pattern

Many users fall into: "I'll use Opus cho everything — why settle for less?"

Wrong because:

• Opus 5-10× cost of Haiku per operation
• Simple tasks don't benefit — Opus output similar
• Faster feedback loop với Haiku cho iteration
• Run out of allocation sooner

Rule of thumb:

┌─────────────────────────────────────────────────┐
│                                                 │
│   Start with SONNET (default)                   │
│                                                 │
│   Downgrade to HAIKU if:                        │
│   - Task simple (extract/format/classify)       │
│   - Need speed (bulk processing)                │
│   - Iterate cheaply                             │
│                                                 │
│   Upgrade to OPUS if:                           │
│   - Multi-step reasoning needed                 │
│   - High-stakes output (legal, financial)       │
│   - Complex architecture/design                 │
│                                                 │
└─────────────────────────────────────────────────┘

Task workflow: Haiku draft → Sonnet polish → Opus tough cases

Sophisticated pattern:

Step 1: Haiku - bulk draft 50 responses   (fast, cheap)
Step 2: Sonnet - polish top 20            (quality)
Step 3: Opus - handle 3 tricky cases      (depth)

Total cost << all Opus. Total quality similar.

Phần 4: Review habit

Why review matter disproportionately

Key insight:

The more polished an output looks, the more a second look is worth. Cowork's confidence is the same whether the work is right or wrong. Your review catches the difference.

Polished PDF với charts và exec summary có thể có:

• Wrong number in chart
• Misattributed quote
• Logical gap in argument
• Regulatory claim unverified

Look same như correct version. Only your review catches it.

The 3-step review habit

┌────────────────────────────────────────────────────┐
│                                                    │
│   Before acting on output, do 3 things:            │
│                                                    │
│   1. 📂 OPEN THE FILE                               │
│      Don't rely on Cowork's summary text.          │
│      Actually open the artifact.                   │
│                                                    │
│   2. 🔢 CHECK A NUMBER                              │
│      Pick 1 specific number.                       │
│      Trace back to source file.                    │
│      Verify.                                       │
│                                                    │
│   3. 🧵 FOLLOW ONE THREAD OF REASONING              │
│      Pick 1 claim → evidence → source.             │
│      Does it hold up?                              │
│                                                    │
│   ~5 phút. Catches 90%+ of errors.                 │
└────────────────────────────────────────────────────┘

Review checklist cho high-stakes outputs

Cho outputs bạn sẽ act on (send to exec, submit to regulator, pay customers):

• Opened file (not just read Cowork's summary)
• Spot-checked 3 random numbers against source
• Verified 1 quote/citation
• Followed logic of 1 major recommendation
• Checked formatting consistency
• Flagged anything unclear back to Cowork
• Had peer review if >$X impact / sensitive

When review fails

Typical errors Cowork can make:

📊 Numerical errors

• Wrong aggregation (sum vs average)
• Unit mismatches (thousands vs millions)
• Rounding where precision matters

📝 Citation errors

• Attribute quote to wrong source
• Mix up authors
• Date/year wrong

🧠 Reasoning errors

• Assume correlation = causation
• Sample selection bias unnoticed
• Missing counter-evidence

🎨 Format errors

• Charts titled wrong
• Chart axis inverted
• Data cut off

All look polished in output. Review is the catch.

Context: AI Fluency's 4D framework

Anthropic's AI Fluency course (anthropic.com/ai-fluency) identifies 4 competencies:

1. Delegationwhat to hand to AI
2. Descriptionhow to describe tasks
3. Discernmentreview outputs critically ← this lesson
4. Diligenceverify, audit, iterate

Cowork leverage = Delegation (giao việc). But Discernment + Diligence là cái catch errors.

Build all 4 competencies, not just delegation.

Bảng tổng hợp: Decision matrix

Use this cheat sheet cho mỗi task:

┌────────────────────────────────────────────────────────┐
│                                                        │
│   Task incoming                                        │
│           │                                            │
│           ▼                                            │
│    ┌─────────────┐                                     │
│    │ Need files/ │  No  ┌──────────────────┐           │
│    │ tools/real  ├─────▶│   Use CHAT       │           │
│    │ output?     │      └──────────────────┘           │
│    └──────┬──────┘                                     │
│           │Yes                                         │
│           ▼                                            │
│    ┌─────────────┐                                     │
│    │ Task        │                                     │
│    │ complexity? │                                     │
│    └──────┬──────┘                                     │
│           │                                            │
│    ┌──────┼──────┬──────────┐                         │
│    ▼      ▼      ▼          ▼                         │
│  Simple  Standard Complex   Regulated                  │
│  Haiku   Sonnet   Opus      Opus + audit              │
│                                                        │
│    │      │      │          │                         │
│    └──────┴──────┴──────────┘                         │
│           ▼                                            │
│    ┌─────────────┐                                     │
│    │   Execute   │                                     │
│    └──────┬──────┘                                     │
│           ▼                                            │
│    ┌─────────────┐                                     │
│    │   Review    │  Always                             │
│    │  (3-step)   │                                     │
│    └─────────────┘                                     │
└────────────────────────────────────────────────────────┘

Ví dụ theo ngành

Legal Counsel — High stakes, audit needed

Task: Contract review cho partnership agreement $5M.

Setup:

• Model: Opus (complex reasoning, high stakes)
• Folder: /contracts/partnerships/active/
• Connectors: Microsoft 365 (Word, SharePoint)
• Audit: enable Enterprise audit logging

Review habit: Open file, check 5 random clauses against template, verify all "red flag" flagged sections make sense, partner sign-off trước send.

Finance — Recurring with trust built

Task: Monthly reconciliation (done 12+ times).

Setup:

• Model: Sonnet (trusted workflow, no need Opus)
• Scheduled: 1st of month 6am
• Folder: /monthly-close/[YYYY-MM]/

Review habit: Spot-check 3 random matched transactions, review exception report carefully, CFO sign-off before books close.

Marketing — Bulk content, low stakes

Task: Generate 50 product descriptions từ spec.

Setup:

• Model: Haiku (bulk, simple transformation)
• Folder: /product-launches/spring-2026/
• No scheduling

Review habit: Sample 10 random descriptions, check against brand voice, fix outliers — rest trust.

Research — Deep synthesis, high cost OK

Task: Literature review synthesis cho grant.

Setup:

• Model: Opus (deep reasoning needed)
• Multi-phase prompt
• Allocation budget: entire week's quota acceptable

Review habit: Verify 10 random citations, check for made-up references, PhD colleague peer-review.

Anti-patterns

Opus-by-default mindset

Cost

5-10× more than needed cho simple tasks.

Fix

Start Sonnet. Upgrade only when task demands it.

Cowork cho all conversation shapes

Cost

Chat-shaped tasks in Cowork waste allocation.

Fix

Internalize decision rule — "does it need files/tools/output file?"

Skip review vì "Cowork usually gets it right"

Cost

When it's wrong, you propagate error.

Fix

3-step review là minimum. No shortcut.

Connect tools "just in case"

Cost

Security surface area. Token management overhead.

Fix

Connect essential 5, add incrementally as needed.

Grant write scopes cho everything

Cost

Missed errors get sent for real.

Fix

Default read-only. Write for trusted workflows only.

Ignore usage monitoring until cap hit

Cost

Surprise mid-week.

Fix

Weekly glance at allocation dashboard. Adjust model/batching.

Schedule everything weekly, even tasks running every day

Cost

Spike at schedule time.

Fix

Spread schedules throughout week. Mix daily/weekly/monthly.

Mẹo nâng cao

Mẹo 1: "Prompt 1 lần với Opus, rerun với Haiku pattern"

For exploratory:

• Iteration 1-3 với Opus (figure out what works)
• Once stable, save as skill hoặc template
• Future runs với Haiku (execute proven pattern)

Mẹo 2: Model switching mid-task

Complex task có phases:

• Phase 1 (research): Opus (needs reasoning)
• Phase 2 (format): Sonnet (standard formatting)
• Phase 3 (bulk output): Haiku (repetitive)

Explicitly prompt:

For phase 1, use Opus. For phase 2 & 3, switch to Haiku.

Mẹo 3: Track "allocation per hour saved"

Metric: allocation used / time_saved_vs_manual = efficiency ratio.

Reconcile task: 5 units allocation, saves 6 hours = 1.2 hrs/unit. Generic brainstorm: 3 units, saves 30 min = 0.17 hrs/unit.

Second task is low efficiency → move to chat.

Mẹo 4: Pre-flight prompts for sensitive tasks

Before any "send email" / "delete files" / "submit" task:

Show me exactly what you will do. Don't execute yet. Wait for my approve.

Forces dry-run mode.

Mẹo 5: Use Haiku cho "decide which model" tasks

Meta:

Here's a task: [describe]. Which model should I use

Haiku, Sonnet, or Opus? Give reason."

Haiku answers instantly. Route decision.

Áp dụng ngay

Bài tập 1: Review habit practice (~5 phút)

Take output Cowork gần đây của bạn. Run 3-step review:

• Opened file (actually)
• Checked 1 number against source
• Followed 1 reasoning thread

Found any error? ☐ Yes ☐ No

If yes → what would have been impact nếu không review?

Bài tập 2: Model audit (~5 phút)

Look at 5 recent Cowork tasks. For each:

Adjust going forward.

Bài tập 3 (optional): Usage monitoring (~2 phút)

Open Settings → Cowork → Usage.

Note:

• Current usage: ___%
• Days left in period: ___
• Pace: on track / spiking / under

Tóm tắt bài học

Tài liệu tham khảo

• AI Fluency course — 4 competencies: Delegation, Description, Discernment, Diligence
• Choosing the right Claude model
• Claude plans & pricing
• Help Center: Cowork usage
• Enterprise compliance & audit