Every time you use an AI coding agent, it starts from zero. You spend an hour debugging some obscure error, the agent figures it out, session ends. Next time you hit the same issue? Another hour.

This skill fixes that. When Claude Code discovers something non-obvious (a debugging technique, a workaround, some project-specific pattern), it saves that knowledge as a new skill. Next time a similar problem comes up, the skill gets loaded automatically.

User-level (recommended)

git clone https://github.com/blader/Claudeception.git ~/.claude/skills/claudeception

Project-level

git clone https://github.com/blader/Claudeception.git .claude/skills/claudeception

The skill can activate via semantic matching, but a hook ensures it evaluates every session for extractable knowledge.

1. Create the hooks directory and copy the script:

mkdir -p ~/.claude/hooks
cp ~/.claude/skills/claudeception/scripts/claudeception-activator.sh ~/.claude/hooks/
chmod +x ~/.claude/hooks/claudeception-activator.sh

2. Add the hook to your global Claude settings (~/.claude/settings.json):

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "~/.claude/hooks/claudeception-activator.sh"
          }
        ]
      }
    ]
  }
}

1. Create the hooks directory inside your project and copy the script:

mkdir -p .claude/hooks
cp .claude/skills/claudeception/scripts/claudeception-activator.sh .claude/hooks/
chmod +x .claude/hooks/claudeception-activator.sh

2. Add the hook to your project settings (.claude/settings.json in the repo):

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          {
            "type": "command",
            "command": ".claude/hooks/claudeception-activator.sh"
          }
        ]
      }
    ]
  }
}

If you already have a settings.json, merge the hooks configuration into it.

The hook injects a reminder on every prompt that tells Claude to evaluate whether the current task produced extractable knowledge. This achieves higher activation rates than relying on semantic description matching alone.

The skill activates automatically when Claude Code:

• Just completed debugging and discovered a non-obvious solution
• Found a workaround through investigation or trial-and-error
• Resolved an error where the root cause wasn't immediately apparent
• Learned project-specific patterns or configurations through investigation
• Completed any task where the solution required meaningful discovery

Trigger a learning retrospective:

/claudeception

Or explicitly request skill extraction:

Save what we just learned as a skill

Not every task produces a skill. It only extracts knowledge that required actual discovery (not just reading docs), will help with future tasks, has clear trigger conditions, and has been verified to work.

The idea comes from academic work on skill libraries for AI agents.

Voyager (Wang et al., 2023) showed that game-playing agents can build up libraries of reusable skills over time, and that this helps them avoid re-learning things they already figured out. CASCADE (2024) introduced "meta-skills" (skills for acquiring skills), which is what this is. SEAgent (2025) showed agents can learn new software environments through trial and error, which inspired the retrospective feature. Reflexion (Shinn et al., 2023) showed that self-reflection helps.

Agents that persist what they learn do better than agents that start fresh.

Claude Code has a native skills system. At startup, it loads skill names and descriptions (about 100 tokens each). When you're working, it matches your current context against those descriptions and pulls in relevant skills.

But this retrieval system can be written to, not just read from. So when this skill notices extractable knowledge, it writes a new skill with a description optimized for future retrieval.

The description matters a lot. "Helps with database problems" won't match anything useful. "Fix for PrismaClientKnownRequestError in serverless" will match when someone hits that error.

More on the skills architecture here.

Extracted skills are markdown files with YAML frontmatter:

---
name: prisma-connection-pool-exhaustion
description: |
  Fix for PrismaClientKnownRequestError: Too many database connections 
  in serverless environments (Vercel, AWS Lambda). Use when connection 
  count errors appear after ~5 concurrent requests.
author: Claude Code
version: 1.0.0
date: 2024-01-15
---

# Prisma Connection Pool Exhaustion

## Problem
[What this skill solves]

## Context / Trigger Conditions
[Exact error messages, symptoms, scenarios]

## Solution
[Step-by-step fix]

## Verification
[How to confirm it worked]

See resources/skill-template.md for the full template.

The skill is picky about what it extracts. If something is just a documentation lookup, or only useful for this one case, or hasn't actually been tested, it won't create a skill. Would this actually help someone who hits this problem in six months? If not, no skill.

See examples/ for sample skills:

• nextjs-server-side-error-debugging/: errors that don't show in browser console
• prisma-connection-pool-exhaustion/: the "too many connections" serverless problem
• typescript-circular-dependency/: detecting and fixing import cycles

Contributions welcome. Fork, make changes, submit a PR.

MIT