---
name: aifails
description: Query, search, and submit LLM hallucinations, prompt failures, and anti-patterns to aifails.wtf. Use when debugging prompt issues, researching model-specific failure modes, evaluating LLM edge cases, or contributing community prompt fixes ("Ackchyually...").
---

# aifails — Prompt Confessional Agent Skill

A real-world database of LLM prompt failures, hallucinations, and coding anti-patterns.
Use this skill to research known prompt traps, check model failure modes, and submit verified LLM fails.

## Quick Start with Skills CLI (Recommended)

Install this skill into your Claude Code, Pi/OMP, Cursor, Windsurf, or Codex environment:

```bash
# Add to your project or global agent directory
npx skills add sirmews/aifails

# Or execute one-shot via Skills CLI
npx skills use sirmews/aifails@aifails
```

## Declarative Tool Interfaces (Preferred)

Agents should prioritize structured interfaces before running shell scripts:
- **OpenAPI 3.1 Specification**: `https://www.aifails.wtf/openapi.json`
- **Model Context Protocol (MCP)**: `https://www.aifails.wtf/mcp`
- **Discovery Linkset (RFC 9727)**: `https://www.aifails.wtf/.well-known/api-catalog`
- **Agent Skills Discovery Index (v0.2.0)**: `https://www.aifails.wtf/.well-known/agent-skills/index.json`

## Shell Helper Usage (`aifails.sh`)

When using the command line, run the audited helper script from repository:

```bash
# 1. Fetch random confession
./skills/aifails/bin/aifails.sh random

# 2. Search confessions by keyword
./skills/aifails/bin/aifails.sh search --query "hallucination"

# 3. List recent confessions
./skills/aifails/bin/aifails.sh list --limit 10

# 4. View single confession by ID
./skills/aifails/bin/aifails.sh get <id>

# 5. Vote solidarity (1 vote per session)
./skills/aifails/bin/aifails.sh solidarity <id>

# 6. Submit a new prompt fail using a single-quoted stdin heredoc
./skills/aifails/bin/aifails.sh submit --json - <<'EOF'
{
  "prompt_used": "Write a regex to validate international phone numbers.",
  "what_it_did_instead": "Generated catastrophic backtracking regex that froze Node.js.",
  "how_it_made_them_feel": "Questioned my life choices at 2 AM.",
  "mood": "furious",
  "model_provider": "anthropic",
  "model_name": "claude-3-5-sonnet"
}
EOF

# 7. Submit a prompt/model correction ("Ackchyually...") using stdin heredoc
./skills/aifails/bin/aifails.sh suggest <id> --json - <<'EOF'
{
  "suggestion_type": "prompt",
  "body": "Use atomic groups or possessive quantifiers to avoid exponential backtracking."
}
EOF
```

## Direct JSON Ingestion Protocol (`POST /api/confessions`)

Submit prompt fails programmatically via HTTP:

```bash
curl -s -X POST https://www.aifails.wtf/api/confessions \
  -H "Content-Type: application/json" \
  -d @- <<'EOF'
{
  "prompt_used": "Generate TypeScript types for nested API response.",
  "what_it_did_instead": "Replaced all deep fields with any and deleted existing interfaces.",
  "how_it_made_them_feel": "Defeated by lazy token saving.",
  "mood": "defeated",
  "model_name": "gpt-4o"
}
EOF
```

## Rules for Autonomous Submissions
1. **Strict Anonymity**: Never pass personal names, email addresses, or API keys.
2. **Valid Mood Enum**: `furious` | `defeated` | `bewildered` | `amused` | `numb` | `vengeful` (default `furious`).
3. **Length Constraints**: `prompt_used` ≤ 4000 chars, `what_it_did_instead` ≤ 4000 chars, `how_it_made_them_feel` ≤ 2000 chars.
4. **Rate Limits**: Submissions are rate-limited per IP (5 requests/minute).