AI Skills That Actually Do Things

正體中文

A collection of AI agent skills that solve real problems — not "summarize this PDF" kind of skills, but "scan my repo for leaked API keys before I push" kind of skills. Works with any LLM client that supports skill/prompt loading, cloud or local.

Skills follow the Claude Code skill convention (SKILL.md + scripts/), but the concepts are framework-agnostic. Think of them as reusable checklists your AI actually follows.

Security Notice: These skills are designed for local development and trusted LAN environments. Skills that interact with external services (e.g., searxng) default to secure settings (TLS verification enabled), but do not implement additional authentication layers. Review each skill's configuration before deploying in sensitive environments.

Skills

Grouped by what you're trying to get done — every skill is still a self-contained top-level folder, the grouping is just a map.

Generate & create

Skill	What It Actually Does
gpt-image-gen	Talk Codex into drawing it for you. Drafts a tight bilingual prompt, iterates until you give an explicit go, then fires Codex CLI for text-to-image — the hard confirmation gate exists because "looks good" is not "yes" and every generation costs real quota. Does text-to-image and img2img (drop an on-disk reference image to lock a face/character via Codex -i), hands you a jpg, and pointedly refuses to shove a preview window in your face
character-lora	Take an original character from "an idea + one reference look" to a trained LoRA that holds its identity across every angle and scene. Orchestrates the whole pipeline — define → multi-angle dataset (delegating image-gen to gpt-image-gen) → caption → base-specific homework → train on a local GPU → verify — and gates every expensive step. The hard rule: read the base model's own training docs before you train, because improvising the settings is how characters come out "rolling the dice" on every generation
rewrite-tone	Turn your dry technical docs into something people actually want to read. War stories > whitepapers

Docs & decks

Skill	What It Actually Does
md2pdf	Turn your Markdown into a PDF that doesn't look like it was generated by a computer from 2003. Handles Mermaid diagrams, CJK fonts, and ASCII art conversion — because we already mass-debugged all the cursed edge cases so you don't have to
md2ppt	md2pdf's louder sibling. Turn your Markdown report into a presentation-quality .pptx via interactive design dialogue + a reusable Python build script. Generic markdown→pptx tools (Marp, pandoc) produce slides that are syntactically correct but visually broken — md2ppt walks per-slide layout decisions with you, then emits a hand-coded script you can re-run in 5 seconds after content edits. Optional self-check via LibreOffice. Brand template integration via ad-hoc helpers — we tried prescribing a workflow, pulled it back after 5 rounds of "wait that's not the cover layout"
conference-report	You went to a conference, recorded the talks, photographed the slides, and came home with a pile of audio, blurry photos, and a vague promise to "write it up someday". This rebuilds faithful per-session notes (slide visuals + speaker transcript, with Whisper hallucinations flagged so you do not quote a robot daydream), then actually asks what report you need -- one talk, a full day, or a multi-day synthesis -- before writing a single word

Spec & project

Skill	What It Actually Does
prd-create	Turn a pile of meeting notes, scattered chat messages, and hand-waved feature requests into a structured PRD that follows your org's guideline — it quizzes you for the gaps instead of fabricating them, surfaces stakeholder conflicts instead of silently picking a side, then sanitizes and publishes to ADO Wiki. Pure prompt-driven, no Python. First half of the prd-create → prd-breakdown → ADO chain
prd-breakdown	Take a finished PRD and slice it into Azure DevOps work items along vertical slices (HITL/AFK tags + blocked_by dependencies), then push them via the az CLI — idempotent re-runs via fingerprint markers, so re-running won't duplicate items. Second half of the prd-create → prd-breakdown → ADO chain
spec	Spec-driven development workflow — from fuzzy idea to verified deliverable. One command, auto-detects project state, walks you through: requirements → review → implement → verify → report. Because "just start coding" is how you end up rewriting everything
goal-engineer	For when you want an agent to grind on something overnight without you hovering over it. Interview-style, it pins down a goal-driven evaluator-optimizer loop of the generate-and-select kind (generate candidates -> grade against a rubric -> iterate by reason-code -> you pick the winner), then emits a self-contained dispatch doc a fresh session runs blind while you just watch the green/yellow/red pings roll in. It is the upstream spec author, not the engine -- hand the dispatch to Claude Code's built-in /goal, a headless claude -p, or any unattended agent. NOT /goal itself, NOT a build-to-spec PRD writer (that's prd-create), NOT a cron timer. One narrow exception: if your build spec is already frozen (approved ADR, locked design, machine-checkable AC) and all you're missing is the unattended-run wrapper, it packages a lean build dispatch instead of making you write a full PRD for a decision you already made. Channel-agnostic notifications (Telegram/Discord/Slack/iMessage), and it bakes in a "want an adversarial review before we ship?" gate -- because we got tired of remembering to ask ourselves

Engineering discipline

This group's discipline mechanics are adapted from mattpocock/skills (MIT) — grilling / diagnosing-bugs / domain-modeling — rewritten in Chinese and fitted to this repo's conventions.

Skill	What It Actually Does
grill	"Let's discuss first," institutionalized. You drop a fuzzy idea, it refuses to touch code, asks you one question at a time — each with a suggested answer — until shared understanding is confirmed. Its sharpest rule: anything answerable by grep may not be asked. It interrogates the filesystem before it interrogates you. When vocabulary drifts, it also grows a CONTEXT.md glossary for the repo
diagnose	Build the lie detector before the interrogation. Until there's one command that reproduces the symptom, no root-cause theory is allowed; and even then you list 3-5 ranked hypotheses with falsifiable predictions before testing the first one. Cures the age-old habit — in AIs and humans — of reading code for two minutes and declaring "found it"
adr	Records architecture decisions, but its first job is talking you out of it. Three gates (hard to reverse / confusing without context / real trade-off) must all pass before it writes anything, and the default format is a title plus three sentences — an ADR's value is in recording why, not in filling out a template. It watches for two things especially: deliberate departures from the obvious path, and explicit no's

Research & security

Skill	What It Actually Does
searxng	Give your local LLM the ability to search the web without sending your queries to Google
repo-scan	Security scan a GitHub repo before you install it. Static analysis, dependency audit, supply chain risks, issue-reported vulnerabilities, maintainer health — because npm install random-package shouldn't require a leap of faith
ctf-kit	Battle-tested playbook for bypassing Windows app authentication — VMProtect, Themida, network verification, you name it. Born from 67+ failed attempts so you don't have to repeat them. Includes ready-to-use Frida recon scripts and a zero-dependency PE analyzer. Pairs well with ljagiello/ctf-skills for broader CTF coverage
job-scout	Research a company before you waste time applying. Salary, reviews, red flags, financials — the due diligence you should've done before that last interview

AI knowledge hygiene

Skill	What It Actually Does
memory-lint	Your AI's memory directory accumulates duplicate rules, stale "active" projects, and orphan files over time. This skill scans for all of that and reports before Claude confidently quotes the wrong rule back at you. Read-only — finds problems, you decide what to fix
llm-wiki-lint	Karpathy's LLM Wiki pattern has a blind spot — past ~15 pages, stale claims, orphan cross-refs, and missing topics silently rot your knowledge base. This skill is the lint pass: contradictions, source traceability, data gaps, frontmatter completeness, index drift. For three-tier raw/ + wiki/ + schema repos. Read-only. Pair with memory-lint for full-stack AI knowledge hygiene
llm-benchmark	Find out which Ollama model actually fits in your GPU — before you waste 30 minutes downloading one that doesn't

Automation & watch

Skill	What It Actually Does
skill-cron	One manager to schedule them all. Register any skill for crontab execution + Telegram push — because claude -p doesn't support /skill syntax, so somebody had to build the bridge. Config in ~/.claude/configs/, logs auto-rotate, crontab entries self-managed
prep-repo	The "did I forget anything?" checklist before pushing to GitHub. README, commits, secrets, broken links — the stuff you always forget at 2 AM

Installation

Grab what you need, leave what you don't:

git clone https://github.com/KerberosClaw/kc_ai_skills.git

# Example: install for Claude Code (user-level)
cp -r kc_ai_skills/prep-repo ~/.claude/skills/

# Example: install for OpenClaw (workspace-level)
cp -r kc_ai_skills/searxng ~/.openclaw/workspace/skills/

Naming tip: Feel free to rename the skill folder with your own prefix when copying (e.g. my_prep-repo). It won't break anything. Probably.

Other clients: Each SKILL.md is a self-contained markdown instruction file. You can paste its content into any AI chat, system prompt, or custom instruction field. No SDK required, no API key needed — just copy and paste.

Skill Structure

Every skill follows a dead-simple convention. If you can write markdown, you can write a skill:

skill-name/
├── SKILL.md          # Frontmatter (name, description, version) + instructions
└── scripts/          # Executable scripts (optional)
    └── script.py

Related Projects

• kc_tradfri_mcp — "Turn on the living room lights" — yes, we made an AI do that
• kc_openclaw_local_llm — We tested 13 local LLMs. Only 2 could reliably call tools. Here's the full report.