Self Improving Agent

207 installs

Summary

This logs every failure, user correction, and capability gap to help Claude learn from mistakes across sessions. When a bash command times out, you say "actually, try it this way," or you ask for something it can't do, it writes structured logs that get reviewed at session start. The hook system catches corrections like "No, that's wrong" and integrates with QAVR to boost memory rankings for successful patterns. It's basically a feedback loop that persists beyond individual conversations. The capability tracker is especially useful because it shows you what features users keep requesting that don't exist yet. Fair warning, you'll need to maintain these log files or they'll grow indefinitely.

Install to Claude Code

npx -y skills add kimasplund/clawdbot-skills-pack --skill self-improving-agent --agent claude-code

Installs into .claude/skills of the current project.

CodeRabbit

AI writes the code. CodeRabbit catches the slop.

Try For Free →

MCP-ready Email Sending

Plug Mailtrap into your AI workflow and let it handle the email.

Connect Mailtrap MCP →

Make your agent a DeFi expert

Agent, run crypto. Access onchain data & trade routes via 1inch.

Install now →

Capacitor - Shared memory for your team’s coding agents.

Make coding agent sessions - Searchable, Shareable, Vendor-neutral & Scored.

Try For Free →

CodeScene MCP Server

Your agent targets a perfect 10 Code Health score. Deterministic. Every commit.

Try For Free →

Give your AI the whole web as clean markdown

Integrate web data into your AI product. One API to scrape website & brand data.

Get API Key Now →

belt - the only tool your agent needs

belt cli automatically finds the best tools and skills for your agent. image, video, music, tts...

one prompt install →

inference shell

create and run specialised agents in minutes

build now →

CodeRabbit

AI writes the code. CodeRabbit catches the slop.

Try For Free →

MCP-ready Email Sending

Plug Mailtrap into your AI workflow and let it handle the email.

Connect Mailtrap MCP →

Make your agent a DeFi expert

Agent, run crypto. Access onchain data & trade routes via 1inch.

Install now →

Capacitor - Shared memory for your team’s coding agents.

Make coding agent sessions - Searchable, Shareable, Vendor-neutral & Scored.

Try For Free →

CodeScene MCP Server

Your agent targets a perfect 10 Code Health score. Deterministic. Every commit.

Try For Free →

Give your AI the whole web as clean markdown

Integrate web data into your AI product. One API to scrape website & brand data.

Get API Key Now →

belt - the only tool your agent needs

belt cli automatically finds the best tools and skills for your agent. image, video, music, tts...

one prompt install →

inference shell

create and run specialised agents in minutes

build now →

Files

SKILL.mdView on GitHub

Self-Improving Agent

Captures learnings, errors, and corrections to enable continuous improvement across sessions.

Activation Triggers

This skill automatically activates when:

Trigger	What Gets Logged	Location
Command fails	Error type, context, recovery suggestion	`logs/failures_detailed.jsonl`
User corrects	"No, that's wrong...", "Actually..."	`logs/corrections.jsonl`
Missing capability	"Can you X?" where X isn't available	`logs/missing_capabilities.jsonl`
API/tool fails	Failure pattern, suggested fix	`logs/failures_detailed.jsonl`
Better approach found	Optimization learned	`logs/learnings.jsonl`

How It Works

1. Automatic Logging (via hooks)

PostToolUse → enhanced-failure-logger.js → logs failures with context
UserMessage → correction-detector.js → detects "wrong/actually/try again"
Response → capability-tracker.js → detects unfulfilled requests

2. Learning Aggregation

All learnings flow to logs/learnings.jsonl:

{
  "timestamp": "2026-01-26T12:00:00Z",
  "type": "user_correction|tool_failure|missing_capability",
  "category": "factual_error|command_failed|web_browsing",
  "description": "what was learned",
  "source": "which detector"
}

3. Session Start Review

On each session start, recent learnings are shown:

=== Learning Review ===
[Learnings] 47 total entries
[Recent]
  • [tool_failure] Bash failed: timeout - WebFetch to external API...
  • [user_correction] User corrected: "No, use the other file..."
[Corrections] 12 user corrections logged
[Capability Gaps] Top requested:
  • send emails (5x)
  • browse web (3x)

4. QAVR Integration

Successful learnings boost Q-values for related memories, improving future retrieval.

Manual Commands

Review Learnings

# Show all learnings
cat ~/.claude/logs/learnings.jsonl | tail -20

# Show corrections only
cat ~/.claude/logs/corrections.jsonl | jq -s 'group_by(.correction_type) | map({type: .[0].correction_type, count: length})'

# Show capability gaps report
node ~/.claude/scripts/hooks/capability-tracker.js --report

Test Detection

# Test correction detector
node ~/.claude/scripts/hooks/correction-detector.js

# Test failure logger
node ~/.claude/scripts/hooks/enhanced-failure-logger.js

# Test capability tracker
node ~/.claude/scripts/hooks/capability-tracker.js

Learning Categories

Correction Types

factual_error - Wrong information provided
retry_request - User asked to try again
misunderstanding - Misinterpreted the request
failed_solution - Solution didn't work

Failure Types

permission_error - Access denied
not_found - File/resource missing
timeout - Operation timed out
network_error - Connection issues
syntax_error - Invalid syntax
api_error - External API failed
command_failed - Shell command failed
agent_failed - Subagent failed

Capability Categories

web_browsing - Internet access requests
image_processing - Image/photo handling
communication - Email/messaging
database_access - SQL/database queries
external_api - Third-party services
memory_persistence - Long-term memory

Configuration

In settings.json, these hooks enable self-improvement:

{
  "hooks": {
    "SessionStart": [...],  // Reviews learnings
    "PostToolUse": [
      {"matcher": "Bash", "hooks": [{"command": "enhanced-failure-logger.js"}]},
      {"matcher": "Task", "hooks": [{"command": "enhanced-failure-logger.js"}]}
    ]
  }
}

Benefits

Learn from mistakes - Don't repeat the same errors
Understand user preferences - Track what corrections mean
Identify skill gaps - Know what features to build
Improve over time - QAVR ranking gets better with feedback
Context persistence - Learnings survive session restarts

Integration with Other Skills

Skill	Integration
QAVR	Successful learnings boost memory Q-values
Memory Consolidation	Periodic cleanup of old learnings
Confidence Check	Review learnings before major tasks
IR-v2	Use learnings to inform pattern selection

Files

~/.claude/
├── logs/
│   ├── learnings.jsonl           # All learnings
│   ├── corrections.jsonl         # User corrections
│   ├── failures_detailed.jsonl   # Enhanced failure logs
│   ├── missing_capabilities.jsonl # Capability requests
│   └── capability_gaps.json      # Aggregated gaps
└── scripts/hooks/
    ├── correction-detector.js
    ├── enhanced-failure-logger.js
    ├── capability-tracker.js
    └── session-start.js (reviews learnings)

Featured

CodeRabbit

AI writes the code. CodeRabbit catches the slop.

Try For Free →

MCP-ready Email Sending

Plug Mailtrap into your AI workflow and let it handle the email.

Connect Mailtrap MCP →

Make your agent a DeFi expert

Agent, run crypto. Access onchain data & trade routes via 1inch.

Install now →

Capacitor - Shared memory for your team’s coding agents.

Make coding agent sessions - Searchable, Shareable, Vendor-neutral & Scored.

Try For Free →

CodeScene MCP Server

Your agent targets a perfect 10 Code Health score. Deterministic. Every commit.

Try For Free →

Give your AI the whole web as clean markdown

Integrate web data into your AI product. One API to scrape website & brand data.

Get API Key Now →

belt - the only tool your agent needs

belt cli automatically finds the best tools and skills for your agent. image, video, music, tts...

one prompt install →

inference shell

create and run specialised agents in minutes

build now →

Self-Improving Agent

Captures learnings, errors, and corrections to enable continuous improvement across sessions.

Activation Triggers

This skill automatically activates when:

Trigger	What Gets Logged	Location
Command fails	Error type, context, recovery suggestion	`logs/failures_detailed.jsonl`
User corrects	"No, that's wrong...", "Actually..."	`logs/corrections.jsonl`
Missing capability	"Can you X?" where X isn't available	`logs/missing_capabilities.jsonl`
API/tool fails	Failure pattern, suggested fix	`logs/failures_detailed.jsonl`
Better approach found	Optimization learned	`logs/learnings.jsonl`

How It Works

1. Automatic Logging (via hooks)

PostToolUse → enhanced-failure-logger.js → logs failures with context
UserMessage → correction-detector.js → detects "wrong/actually/try again"
Response → capability-tracker.js → detects unfulfilled requests

2. Learning Aggregation

All learnings flow to logs/learnings.jsonl:

{
  "timestamp": "2026-01-26T12:00:00Z",
  "type": "user_correction|tool_failure|missing_capability",
  "category": "factual_error|command_failed|web_browsing",
  "description": "what was learned",
  "source": "which detector"
}

3. Session Start Review

On each session start, recent learnings are shown:

=== Learning Review ===
[Learnings] 47 total entries
[Recent]
  • [tool_failure] Bash failed: timeout - WebFetch to external API...
  • [user_correction] User corrected: "No, use the other file..."
[Corrections] 12 user corrections logged
[Capability Gaps] Top requested:
  • send emails (5x)
  • browse web (3x)

4. QAVR Integration

Successful learnings boost Q-values for related memories, improving future retrieval.

Manual Commands

Review Learnings

# Show all learnings
cat ~/.claude/logs/learnings.jsonl | tail -20

# Show corrections only
cat ~/.claude/logs/corrections.jsonl | jq -s 'group_by(.correction_type) | map({type: .[0].correction_type, count: length})'

# Show capability gaps report
node ~/.claude/scripts/hooks/capability-tracker.js --report

Test Detection

# Test correction detector
node ~/.claude/scripts/hooks/correction-detector.js

# Test failure logger
node ~/.claude/scripts/hooks/enhanced-failure-logger.js

# Test capability tracker
node ~/.claude/scripts/hooks/capability-tracker.js

Learning Categories

Correction Types

factual_error - Wrong information provided
retry_request - User asked to try again
misunderstanding - Misinterpreted the request
failed_solution - Solution didn't work

Failure Types

permission_error - Access denied
not_found - File/resource missing
timeout - Operation timed out
network_error - Connection issues
syntax_error - Invalid syntax
api_error - External API failed
command_failed - Shell command failed
agent_failed - Subagent failed

Capability Categories

web_browsing - Internet access requests
image_processing - Image/photo handling
communication - Email/messaging
database_access - SQL/database queries
external_api - Third-party services
memory_persistence - Long-term memory

Configuration

In settings.json, these hooks enable self-improvement:

{
  "hooks": {
    "SessionStart": [...],  // Reviews learnings
    "PostToolUse": [
      {"matcher": "Bash", "hooks": [{"command": "enhanced-failure-logger.js"}]},
      {"matcher": "Task", "hooks": [{"command": "enhanced-failure-logger.js"}]}
    ]
  }
}

Benefits

Learn from mistakes - Don't repeat the same errors
Understand user preferences - Track what corrections mean
Identify skill gaps - Know what features to build
Improve over time - QAVR ranking gets better with feedback
Context persistence - Learnings survive session restarts

Integration with Other Skills

Skill	Integration
QAVR	Successful learnings boost memory Q-values
Memory Consolidation	Periodic cleanup of old learnings
Confidence Check	Review learnings before major tasks
IR-v2	Use learnings to inform pattern selection

Files

~/.claude/
├── logs/
│   ├── learnings.jsonl           # All learnings
│   ├── corrections.jsonl         # User corrections
│   ├── failures_detailed.jsonl   # Enhanced failure logs
│   ├── missing_capabilities.jsonl # Capability requests
│   └── capability_gaps.json      # Aggregated gaps
└── scripts/hooks/
    ├── correction-detector.js
    ├── enhanced-failure-logger.js
    ├── capability-tracker.js
    └── session-start.js (reviews learnings)

Self Improving Agent

Install to Claude Code

Self-Improving Agent

Activation Triggers

How It Works

1. Automatic Logging (via hooks)

2. Learning Aggregation

3. Session Start Review

4. QAVR Integration

Manual Commands

Review Learnings

Test Detection

Learning Categories

Correction Types

Failure Types

Capability Categories

Configuration

Benefits

Integration with Other Skills

Files

Self Improving Agent

Install to Claude Code

Self-Improving Agent

Activation Triggers

How It Works

1. Automatic Logging (via hooks)

2. Learning Aggregation

3. Session Start Review

4. QAVR Integration

Manual Commands

Review Learnings

Test Detection

Learning Categories

Correction Types

Failure Types

Capability Categories

Configuration

Benefits

Integration with Other Skills

Files

Recommended

Recommended