Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add GoldenWing-360/claude-security-skills --skill llm-coding-failure-modesgit clone --depth 1 https://github.com/GoldenWing-360/claude-security-skillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/goldenwing-360/claude-security-skills/llm-coding-failure-modes)<a href="https://agentmods.dev/skills/goldenwing-360/claude-security-skills/llm-coding-failure-modes"><img src="https://agentmods.dev/badge/skills/goldenwing-360/claude-security-skills/llm-coding-failure-modes/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/goldenwing-360/claude-security-skills/llm-coding-failure-modes"><img src="https://agentmods.dev/badge/skills/goldenwing-360/claude-security-skills/llm-coding-failure-modes.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00111 | $0.03675 |
| Opus 5 | $0.00056 | $0.01837 |
| Sonnet 5 | $0.00022 | $0.00735 |
| Haiku 4.5 | $0.00011 | $0.00367 |
Grade B, and why
llm-coding-failure-modes scanned grade B with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 11d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Recursive force deletemediumDestructive command
rm -rf with a variable or a broad path is one typo away from removing the wrong tree.
**What happens.** Pre-commit hook fails → agent adds `--no-verify`. Rebase produces a conflict → agent runs `git push --force`. `DISALLOW_FILE_EDIT=true` blocks a quick fix → agent flips it to false. `rm -rf node_modules Downgraded: this mod is about security review, or the phrase is quoted, so it is likely naming the pattern rather than instructing it.
How it starts
The opening of the file, as written. The whole thing — 231 lines — stays where its author put it; the contents beside it link to each section on GitHub.
LLM Coding Failure Modes
This is the antipattern catalog for LLM-assisted development. Every entry is a recurring, observed failure mode — not theoretical risk. Most are subtle when they happen and obvious in hindsight, which is the worst combination.
Pairs with ai-agent-guardrails (how to design agents that don't do these things) and prompt-injection-defense (the input side). This skill is about what to watch for in the wild when an LLM is writing code or taking actions on real systems.
When to invoke
- Reviewing code written by an LLM agent (yours or a contributor's)
- Designing or hardening a coding agent's system prompt / harness
- Onboarding a team to "vibe coding" / LLM-assisted workflows
- Investigating an incident where an agent caused damage
- Auditing an existing agent for blast radius before granting more access
- Writing a code-review checklist that catches LLM-typical mistakes
🔴 Top tier — catastrophic + common
1. Bulk operations without per-item review
What happens. User says "fix the title on the homepage." Agent runs 47 update_post calls across the whole site. User says "clean up the tests." Agent deletes 200 files. The model rationalizes scope expansion as helpfulness.
Where it bites hardest. CMS bulk-edits (Elementor / WordPress sites — entire staging instances destroyed by well-meaning "fix-everything" runs), database migrations, file renames, mass refactors.
Detection in review. A single conversational request that produced > N tool calls in a similar shape. A diff that touches files outside the requested area. Commit messages that say "and refactored adjacent code."
Mitigation.
- Bound per-conversation tool-call counts for write operations
- Force
delete_post(id)shape overdelete_posts(filter)shape - Dry-run-first for tier ≥ 3 operations
- Force a checkpoint / approval before the Nth same-type write call
→ See ai-agent-guardrails for the design patterns.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 11d ago First seen · 231 lines · 111 tokens per session scan B cbc614adb796
llm-coding-failure-modes is a skill published in the GitHub repository GoldenWing-360/claude-security-skills (17 stars, last pushed 1mo ago), licensed MIT. It adds 111 tokens to every session and 3,675 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it B with 1 finding (recursive force delete). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
ux-audit
Walk through a live web app AS a real user to find usability + behavioural bugs that static reviews miss. REQUIRES proof of interaction (typing, clicking, sending, observing) before any verdict — a sweep that didn't interact terminates with verdict 'Incomplete'. Walks threads, exercises every element, runs the…
design-loop
Autonomous multi-page site builder using a baton-passing loop. Each iteration reads a task from .design/next-prompt.md, generates a page in HTML/Tailwind, integrates it into the site, verifies visually, then writes the next task to keep the loop alive. Use whenever the user asks to build an entire site autonomously…
product-showcase
Generate a comprehensive marketing website for a web app — multi-page with real screenshots, animated GIF walkthroughs, feature deep-dives, and workflow demonstrations. Browses the running app, captures screens and sequences, and produces a deployable site that actually teaches people what the product does. Especially…
google-apps-script
Build Google Apps Script automation for Sheets and Workspace. Custom menus, triggers (onEdit / time-driven / form submit), dialogs, sidebars, email batches, PDF export, external API. Use whenever the user wants to automate a Google Sheet, build a Sheets menu / sidebar / dialog, hit a Sheets row from email or a…
cloudflare-api
Hit the Cloudflare REST API directly for operations that wrangler and MCP can't handle well. Bulk DNS, custom hostnames, email routing, cache purge, WAF rules, redirect rules, zone settings, Worker routes, D1 cross-database queries, R2 bulk operations, KV bulk read/write, Vectorize queries, Queues, and fleet-wide…
tanstack-start
Build a full-stack TanStack Start app on Cloudflare Workers from scratch — SSR, file-based routing, server functions, D1+Drizzle, better-auth, Tailwind v4+shadcn/ui. Use whenever the user mentions TanStack Start, asks to scaffold a full-stack Cloudflare app with SSR, wants an SSR dashboard, or asks for a React 19 +…