our-claude-skills/custom-skills/00-our-settings-audit/desktop/SKILL.md

name, description

name	description
our-settings-audit	Diagnose all Claude Desktop errors and optimize context usage. Covers 8 error categories: context limits, output interruption, length/file errors, usage limits, server capacity, MCP connection, account issues, and output quality problems. Triggers: settings audit, exceed response limit, MCP error, token error.

Claude Desktop Settings Checker

Diagnose Claude Desktop errors and optimize context usage.

Error Quick Reference

Error Type	Message Pattern	Category
Context too large	"Exceed response limit"	Input
Output interrupted	"Response could not be fully generated"	Output
Length exceeded	"Message will exceed the length limit"	Input
Conversation limit	"This conversation reached its maximum length"	Input
File too large	"Files larger than 10mb" / "X% over the length limit"	File
Rate limited	"5-hour limit reached" / "X messages left"	Usage
Server overload	"Due to unexpected capacity constraints"	Server
MCP failure	"Error connecting to [ServerName]"	MCP
Account issue	"Account has been disabled"	Account

Quick Diagnosis Tree

Error occurred?
│
├─ BEFORE response started?
│   ├─ "Exceed response limit" → Reduce input context
│   ├─ "Length limit" → Break into smaller messages
│   ├─ "File too large" → Reduce file size or excerpt
│   ├─ "Capacity constraints" → Wait and retry
│   └─ "MCP error" → Check MCP configuration
│
├─ DURING response?
│   ├─ "Could not be fully generated" → Request smaller output
│   └─ Output suddenly stopped → Retry or chunk request
│
└─ BEFORE even sending?
    ├─ "5-hour limit reached" → Wait for reset
    ├─ "Messages left" → Near usage limit
    └─ "Account disabled" → Contact support

Understanding Token Limits

Model	Context Window	Output Limit	Practical Available
Claude 3.5 Sonnet	200K tokens	~8K tokens	~150K after system
Claude 3 Opus	200K tokens	~4K tokens	~150K after system
Claude 3 Haiku	200K tokens	~4K tokens	~160K after system

---

CATEGORY 1: Input/Context Errors

Error: "Exceed response limit"

Causes:

1. Loaded context too large (MCP tools + skills + files)
2. Conversation history accumulated
3. Multiple large files attached

Quick Diagnosis

When user reports "Exceed response limit":

Step 1: Check Loaded Context

Ask user to share their claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Analyze:

• Number of MCP servers configured
• Servers with many tools (>20 tools each)
• Missing or poor serverInstructions

Step 2: Estimate Token Usage

Component	Tokens (approx)	Notes
Per MCP server	500-2,000	Base overhead
Per tool definition	100-500	Depends on schema complexity
Playwright MCP	~13,500	20+ tools, detailed schemas
Notion MCP	~5,000	Moderate
GitHub MCP	~18,000	Many tools
Custom SKILL.md	200-2,000	Depends on length
Attached file	Variable	~4 tokens per word

Red flags:

• More than 5 MCP servers active
• Any server without serverInstructions
• Total MCP tools > 100

Step 3: Review Loaded Skills

Check for:

• Skills with long SKILL.md files (>1,000 words)
• Multiple skills loaded simultaneously
• Skills with embedded reference data

Optimization Recommendations

MCP Server Optimization

Priority 1: Add serverInstructions

Every MCP server needs concise instructions:

{
  "mcpServers": {
    "playwright": {
      "command": "...",
      "serverInstructions": "Browser automation for web testing. Use for: screenshots, page analysis, form testing. Keywords: browser, DOM, screenshot"
    }
  }
}

Pattern: [Purpose]. Use for: [case1], [case2]. Keywords: [kw1], [kw2]

Priority 2: Reduce Active Servers

Keep only essential servers active. Disable or remove:

• Servers used less than weekly
• Servers with overlapping functionality
• Servers with 30+ tools (like Zapier)

Conversation Management

Before complex tasks:

1. Start fresh conversation for unrelated work
2. Avoid attaching multiple large files
3. Request concise outputs

During long sessions:

1. Start new conversation at natural breakpoints
2. Summarize important context before continuing
3. Remove completed file attachments

Project File Optimization

For project skills and configurations:

File Type	Target Length	Max
SKILL.md	500 words	1,000 words
Project instructions	200 words	500 words
Embedded data	0	Move to external

Optimization techniques:

1. Move reference data to separate files (load on demand)
2. Use tables instead of prose
3. Remove redundant information
4. Link to documentation instead of embedding

Token Budget Calculator

Help user calculate their baseline:

System overhead:        ~10,000 tokens (fixed)
MCP servers:            [count] × ~3,000 = [estimate]
Loaded skills:          [count] × ~500 = [estimate]
Project instructions:   ~[estimate based on length]
────────────────────────────────────────────────
Baseline total:         [sum]
Available for work:     200,000 - [sum] = [remaining]
Available percentage:   [remaining/200,000 × 100]%

Targets:

• Baseline: < 40,000 tokens (20%)
• Available: > 160,000 tokens (80%)

Warning zone:

• Baseline > 60,000 tokens: Likely to hit limits
• Available < 140,000 tokens: Reduce configuration

Output Format

After diagnosis, provide:

## Settings Check Report

### Configuration Summary
- MCP Servers: [count] ([tokens estimate])
- Active Skills: [count] ([tokens estimate])
- Baseline Total: ~[X] tokens ([Y]%)
- Available: ~[Z] tokens ([W]%)

### Issues Found
1. [Critical]: [issue description]
2. [Warning]: [issue description]

### Recommended Actions
1. [Action with specific steps]
2. [Action with specific steps]

### Quick Wins (immediate impact)
- [ ] [Specific change]
- [ ] [Specific change]

Common Scenarios

Scenario: Too Many MCP Servers

Symptoms: Errors on first message of conversation Solution: Disable non-essential servers, add serverInstructions

Scenario: Long Conversation History

Symptoms: Errors after many exchanges Solution: Start new conversation, summarize context

Scenario: Large File Attachments

Symptoms: Errors when attaching files Solution: Attach smaller excerpts, use file references

Scenario: Complex Skill Loaded

Symptoms: Errors when using specific skill Solution: Refactor skill, move data to references

---

Error 2: "Response could not be fully generated"

This error occurs when output generation is interrupted mid-stream.

Causes

Cause	Description	Frequency
Output token limit	Response exceeded ~4K-8K output limit	Common
Safety filter	Content flagged during generation	Occasional
Network interruption	Connection dropped mid-stream	Occasional
Server timeout	Generation took too long	Rare
Rate limiting	Too many requests in short time	Rare

Common Triggers

1. Requesting too much output:

❌ "Write a complete 50-page report on..."
❌ "Generate the entire application code"
❌ "List all 500 items with full descriptions"

✅ "Write section 1: introduction and overview"
✅ "Generate the authentication module first"
✅ "List items 1-50 with descriptions"

2. Vague prompts (produce longer responses):

❌ "Explain everything about machine learning"
✅ "Explain gradient descent in 3 paragraphs"

3. Code generation without scope:

❌ "Build me a full-stack e-commerce site"
✅ "Create the product listing component with pagination"

Solutions

Situation	Fix
Long report/document	Request section by section
Large codebase	Generate module by module
Comprehensive list	Paginate (1-50, 51-100, etc.)
Detailed explanation	Ask for "brief" or "concise" version
Transient failure	Simply retry the request

Prompting Techniques

Add output constraints:

• "In under 500 words..."
• "Provide a concise summary..."
• "List the top 10..."
• "Focus only on..."

Chunk large tasks:

• "First, outline the structure"
• "Now write section 1"
• "Continue with section 2"

Request format control:

• "Use bullet points, not paragraphs"
• "Provide a table summary"
• "Give me just the key points"

When to Retry vs. Rephrase

Scenario	Action
Error on first attempt	Retry once
Error persists after retry	Rephrase with smaller scope
Error always at same point	Content may be triggering filter
Error after long generation	Request was too ambitious

---

CATEGORY 3: Length & File Errors

Error: "Your message will exceed the length limit"

Full message: "Your message will exceed the length limit for this chat..."

Causes:

• Single message too long (approaching context limit)
• Too much text pasted at once

Solutions:

1. Break content into smaller chunks
2. Summarize key sections before sending
3. Ask Claude to identify relevant portions first
4. Start a new conversation

Error: "This conversation reached its maximum length"

Cause: Accumulated conversation history filled context window

Solutions:

1. Start a new conversation
2. Copy essential context to new chat
3. Use shorter exchanges going forward

Error: "Message is X% over the length limit"

Full message: "Message is X% over the length limit. Try replacing the attached file with smaller excerpts."

Cause: Attached file(s) too large for remaining context

Solutions:

1. Extract relevant excerpts only
2. Summarize large documents before attaching
3. Split across multiple messages
4. Use file references instead of full content

Error: "You may not upload files larger than 10MB"

Cause: Single file exceeds 10MB limit

Solutions:

1. Compress file if possible
2. Split into smaller files
3. Extract text content only
4. Use external link and describe content

---

CATEGORY 4: Usage & Rate Limit Errors

Error: "5-hour limit reached - resets [time]"

Cause: Hit plan's usage cap within 5-hour window

Solutions:

1. Wait until reset time shown
2. Upgrade plan for higher limits
3. Optimize prompts to use fewer tokens

Error: "Approaching 5-hour limit"

Cause: Warning before hitting limit

Solutions:

1. Prioritize remaining important tasks
2. Use concise prompts
3. Avoid exploratory conversations

Error: "X messages left until [time]"

Cause: Near message limit (free tier)

Solutions:

1. Combine multiple questions into one
2. Wait for reset
3. Consider paid plan

Error: "5-hour limit resets [time] - continuing with extra usage"

Note: This is informational for paid plans with extra usage enabled Action: Extra usage automatically continues; monitor costs

---

CATEGORY 5: Server & Capacity Errors

Error: "Due to unexpected capacity constraints, Claude is unable to respond"

Cause: High system-wide demand

Solutions:

1. Wait 2-5 minutes and retry
2. Try during off-peak hours
3. Check status.claude.com for incidents

HTTP Error Codes

Code	Meaning	Solution
500	Internal Server Error	Wait and retry
429	Too Many Requests	Slow down, wait
403	Forbidden	Check account status
400	Bad Request	Check input format

Error: "Claude AI Saving Chat Failed"

Cause: Issue saving conversation to servers

Solutions:

1. Refresh the page
2. Check internet connection
3. Copy important content locally
4. Try again later

---

CATEGORY 6: MCP Connection Errors

Error: "There was an error connecting to [ServerName]"

Causes:

1. Server not running
2. Incorrect configuration
3. Transport type mismatch (stdio vs HTTP)
4. Authentication failure

Diagnosis Steps

Step 1: Check server logs

• macOS: ~/Library/Logs/Claude/mcp-server-SERVERNAME.log
• Windows: %APPDATA%\Claude\logs\mcp-server-SERVERNAME.log

Step 2: Verify configuration

# macOS
cat ~/Library/Application\ Support/Claude/claude_desktop_config.json

# Windows PowerShell
Get-Content "$env:APPDATA\Claude\claude_desktop_config.json"

Step 3: Test server independently

# Test if server starts
npx @anthropic-ai/mcp-server-name

Common MCP Issues

Issue	Symptom	Fix
Transport mismatch	HTTP server with stdio config	Add "transport": "http"
Path with spaces	Server fails on Windows	Use escaped paths or %APPDATA%
Missing env vars	Auth failures	Add to "env" in config
Server not installed	"command not found"	Run npm install -g first
Port conflict	Server won't start	Change port or kill existing

After Config Changes

Critical: Completely quit and restart Claude Desktop after any config changes.

Error: "Required parameter" validation errors

Cause: MCP tool schema mismatch (regression bug)

Solutions:

1. Update Claude Desktop to latest version
2. Check MCP server for updates
3. Temporarily disable problematic server

---

CATEGORY 7: Account Errors

Error: "Your account has been disabled"

Full message: "Your account has been disabled after an automatic review of your recent activities."

Cause: Automated system flagged account

Solutions:

1. Contact Anthropic support
2. Review Terms of Service
3. Do not create new accounts (may worsen situation)

Error: "This organization has been disabled"

Cause: Organization-level suspension

Solution: Organization admin must contact Anthropic

Error: "There was an error logging you in"

Causes:

• VPN interference
• Browser extensions blocking
• Cached credentials issue

Solutions:

1. Disable VPN
2. Disable browser extensions temporarily
3. Clear browser cache and cookies
4. Try incognito/private window
5. Check status.claude.com

Error: "Error Sending Codes - Check Your Phone Number"

Cause: Phone verification failed during signup

Solutions:

1. Verify correct country code
2. Check all digits entered correctly
3. Try different phone number
4. Wait and retry later

---

CATEGORY 8: Output Quality Issues

Issue: Output Suddenly Stops (Cutoff)

Symptoms: Response ends mid-sentence or mid-code

Causes:

1. Hit output token limit (~4K-8K)
2. Safety filter triggered
3. Server timeout

Solutions:

1. Ask "Please continue from where you stopped"
2. Request smaller chunks upfront
3. Be more specific in scope

Issue: Gibberish or Nonsensical Output

Symptoms: Incoherent text, random characters

Causes:

1. Model confusion from complex prompt
2. Conflicting instructions
3. Edge case in training

Solutions:

1. Simplify prompt
2. Start new conversation
3. Rephrase request differently

Issue: Mixed Languages or Character Issues

Symptoms: Unexpected language switches, formatting problems

Causes:

1. Ambiguous language context
2. Copy-pasted special characters

Solutions:

1. Explicitly state desired language
2. Clean input text of special characters
3. Add "Please respond in [language]"

Issue: Claude Refuses to Respond

Symptoms: Declines request citing policy

Causes:

1. Request triggered safety guidelines
2. Ambiguous phrasing interpreted as harmful

Solutions:

1. Rephrase more clearly
2. Explain legitimate use case
3. Break into smaller, clearer requests

---

Prevention Checklist

Context/Input Errors

• Maximum 5 MCP servers active
• All servers have serverInstructions
• Skills under 1,000 words
• Start fresh conversations for new topics
• Avoid attaching files over 50KB (prefer excerpts)
• Individual files under 10MB

Output Errors

• Break large requests into chunks
• Use specific scope in prompts
• Request concise/brief outputs when possible
• Paginate lists and tables (50 items max)
• Generate code module by module
• Retry once before rephrasing

Usage Limits

• Monitor "messages left" warnings
• Combine related questions into single prompts
• Use concise prompts to conserve tokens
• Know your plan's reset schedule

MCP Stability

• Test MCP servers independently before adding
• Keep server logs accessible for debugging
• Restart Claude Desktop after config changes
• Update MCP servers regularly
• Document working configurations

General Best Practices

• Check status.claude.com when experiencing issues
• Keep Claude Desktop updated
• Clear browser cache periodically (web version)
• Disable VPN if experiencing login issues
• Back up important conversations locally