Agentic AI Skills

Six reusable patterns for building agentic AI pipelines that don't fail silently. Extracted from production code after burning a week on bugs that turn out to be universal — every team that builds something like this rediscovers them in the same order.

Each file is a self-contained Claude Code skill: YAML frontmatter declaring when it applies, then the operational pattern, code, and the mistakes to avoid.

The bugs these skills prevent

If you're nodding at any of these, you're the target audience:

• Your scan / pipeline runs to completion. The report is empty. No error, no warning, just three section headers with whitespace under them.
• You added an example bash command in a prompt as guidance. The 7B model executed it verbatim and wrote [fill in the value] into your knowledge file.
• You wired up /model groq → ollama to switch providers. It only changed the CLI label. Execution still hit Groq because the agent's provider field was baked in at module load.
• Your model assigns CVSS scores by vibes. 7.5 one run, 9.0 the next, on the same finding.
• Your model classifies a Laravel site as WordPress despite four unambiguous signals because a 100-line prompt diluted its attention.
• A safety-tuned hosted model refuses on a benign offensive-security prompt. No retry. Pipeline dead.

Six skills, one for each. None of them are clever — they're just the things you wish someone had told you before the third week of debugging.

The six

Skill	When it applies
agentic-pipeline-prompt-engineering	Designing multi-step LLM pipelines where each step calls tools and writes intermediate state. Six anti-patterns + a regression-test approach that catches all of them.
llm-failover-adapter	Primary LLM provider intermittently refuses or rate-limits; you want silent fallback to a local/secondary provider with refusal-pattern detection. 10-pattern regex + the tool-use short-circuit.
multi-provider-ai-cli	Building an interactive AI CLI with --provider / --model flags and a /model slash-command picker. Includes the runtime-config-threading bug that costs every team a day.
cvss-scoring	Drop-in pure-function CVSS 3.1 scorer. ~150 lines. Take the score out of the LLM's hands; let it pick the vector string only.
deterministic-over-llm-detection	When a categorical decision is bounded and signals are observable, replace the LLM call with regex+bash. The framework-detector worked example.
domain-knowledge-base-injection	Per-domain markdown KB files loaded conditionally with size budgets. The pattern that gives a 7B model the recall of a 70B model on a specific domain.

Reading order

If you're building an agentic pipeline from zero:

1. agentic-pipeline-prompt-engineering — the foundation. Six bugs you'd otherwise hit one at a time.
2. deterministic-over-llm-detection — fewer LLM calls, more correctness.
3. domain-knowledge-base-injection — give the model the recall it doesn't have.
4. cvss-scoring (if you do security work) — take scoring out of the LLM.
5. multi-provider-ai-cli — once you have a tool, ship it as a usable CLI.
6. llm-failover-adapter — operability layer for production.

Use as Claude Code skills

Copy any *.md file to:

• ~/.claude/skills/ for user-global use
• <project>/.claude/skills/ for project-local use

Claude Code reads the name and description from each file's frontmatter and surfaces the relevant skill when its description matches your task.

Use as patterns in any codebase

These aren't a runtime dependency — they're patterns. Copy the code snippets into your own project. The skills cite reference implementations (file + line ranges) so you can see the patterns in production code, but you don't need to clone or import anything.

What's deliberately NOT here

• A specific pipeline structure (e.g. recon → enumerate → attack → analyze). Domain-specific. Useful as inspiration; the reusable insight is in the prompt-engineering skill.
• A 100-scenario attack KB. Reusable for offensive-security work but too bug-bounty-specific to belong in a generic skills repo.
• Cheat sheets for individual attack classes (subdomain takeover, JWT, SSRF, GraphQL). Already exist elsewhere; the value-add of the source project was the pipeline wiring, not the cheat-sheet content.

The point is: be honest about what generalizes. Most of what a single project does isn't generalizable. The parts that are, are these.

Provenance

Extracted from production code that survives a npm run build clean and 87 unit tests passing. Each skill includes references to its source file and line ranges so you can see the pattern in a working system, not just on paper.

Contributing

Pull requests welcome — especially worked examples in non-security domains (content moderation, code review, customer-support agents) showing the same patterns generalize. Open an issue first if you're proposing a new skill rather than improving an existing one.

License

MIT. Copy freely. If a skill saved you time, a star is appreciated but not expected — the goal is to delete a class of bug from the world, not to collect numbers.