A small, MIT-licensed collection of cross-agent Agent Skills by buddypia. Most skills make two or three different-vendor LLMs (Gemini, Claude, GPT) check and build on each other's work; others are single-agent procedures that run entirely inside your agent — so you get more than a single model's one-shot answer on the questions that matter. Every skill follows the open SKILL.md standard and works across Claude Code, OpenAI Codex, Cursor, Gemini CLI, and other compatible agents.
A single LLM has predictable blind spots: it rarely catches its own mistakes, it inherits the biases of its training data, and it tends to agree with the framing it is given (sycophancy). Asking the same model to "double-check itself" mostly repeats those blind spots.
The multi-LLM skills take a different approach — they assign each role to a different vendor's model. One model proposes; another, from a different lab with different training, critiques it or argues the opposite; then the results are reconciled. Independent models make less-correlated errors, so genuine disagreement surfaces real problems instead of echoing them. The aim is a more robust result on the decisions and deliverables where a single pass is not enough.
| Skill | Pattern | Problem it addresses |
|---|---|---|
multi-llm-debate |
Proponent / Opponent / Moderator → verdict | One-sided or overconfident answers on judgment calls |
multi-llm-reflection |
Generator → Critic → Refiner | A draft that needs a sharper, outside critique to improve |
multi-llm-recursive-meta-cognition |
Decompose → Solve → Verify → Integrate → Reflect | Hard, multi-step problems where one pass reasons too shallowly |
reflect |
Trigger → 5 Whys → Placement → Cure + Prevent → Ledger | Bugs and near-misses that recur because the fix lands at the wrong level |
claude-code-steering |
Route / Audit → 4 axes → right mechanism | Instructions that live in the wrong mechanism, so CLAUDE.md bloats and rules get ignored |
loop-engineering |
Scan → Analyze → Build → Harden → Measure | Repetitive agent work looped without a gate, so it drifts, burns tokens, and ships unreviewed code |
spec-preview |
Mode + target → (cached) UI scan → self-contained HTML → open & decide | Specs and UI ideas skimmed as prose, so review stalls and decisions slip |
The
multi-llm-*skills orchestrate vendor CLIs that you install yourself (agy/ Antigravity,claude/ Claude Code,codex/ Codex); single-agent skills likereflectneed no external CLI. See each skill's README for setup, model overrides, and an offlinemockmode.
- multi-llm-debate — Architecture and tech-stack choices, build-vs-buy, risk assessment, "should we ship this?" calls — weighing trade-offs where you do not want a single model's bias to decide.
- multi-llm-reflection — Improving high-stakes writing and design: proposals, RFCs, docs, marketing copy, or an analysis you want critiqued by a model other than the one that wrote it.
- multi-llm-recursive-meta-cognition — Complex, multi-step reasoning: migration plans, debugging strategies, research-style questions — anything that benefits from decomposition, independent verification, and a final meta-review.
- reflect — Postmortems and root-cause analysis after a bug, near-miss, or repeated friction: turn the incident into a fix at the right control tier plus a written ledger, instead of a quick patch that lets it come back.
- claude-code-steering — Deciding where a Claude Code instruction belongs (CLAUDE.md vs rule vs skill vs subagent vs hook vs output style), and auditing an existing .claude config that has drifted or bloated.
- loop-engineering — Deciding whether a recurring agent task should become a self-prompting loop — and proving when it shouldn't — then building the minimum viable loop (one automation + skill + state file + gate) and hardening it against goal drift, runaway token spend, and unreviewed merges.
- spec-preview — Turning a spec, a set of requirements, or a UI proposal into a self-contained HTML bundle — compare options, a requirements dashboard, or a before/after diff — that opens in the browser so a human can eyeball it and decide, instead of skimming prose. Target-surface aware (mobile app / web / document / diagram) with a cached, framework-agnostic UI scan that reproduces the product's look.
Reach for these when the stakes justify the extra time and tokens — a hard decision, a deliverable that has to be right, a thorny multi-step problem. The multi-LLM skills run several CLIs in sequence, so they are slower and cost more than a single prompt; for quick lookups or simple edits, a normal single-model call is the better tool. Multi-model orchestration reduces blind spots — it does not guarantee a correct answer, so always review the output.
Using the skills CLI (works with Claude Code, Codex, Cursor, Gemini CLI, and many more agents):
# Browse/add skills from this repo
npx skills add buddypia/agent-skillsOr install manually by copying a skill folder into your agent's skills directory — for example ~/.claude/skills/<name>/ (Claude Code) or ~/.agents/skills/<name>/ (Codex). Each skill is a self-contained directory with a SKILL.md and its own README.md (script-based skills also bundle scripts/).
The multi-llm-* skills drive the official vendor CLIs you install yourself, under your own login: agy (Antigravity), claude (Claude Code), codex (Codex). Verify with command -v agy claude codex. Python dependencies (pydantic / python-dotenv / pyyaml) are auto-prepared by each skill's run.sh (uv if available, else venv + pip). Single-agent skills like reflect need none of this; spec-preview needs only node for its UI scan.
These skills orchestrate official CLIs that you install yourself; they do not circumvent authentication or billing, and API keys are supported as an alternative. You are responsible for complying with each provider's and CLI's terms of service when automating them — automating subscription-authenticated CLIs may be subject to usage restrictions, and any account or usage consequences are your own.
"Claude" / "Claude Code" (Anthropic), "GPT" / "ChatGPT" / "Codex" (OpenAI), and "Gemini" / "Antigravity" (Google) are trademarks of their respective owners. This is an independent project and is not affiliated with, endorsed by, or sponsored by Anthropic, OpenAI, or Google.
Gemini and Claude defaults (for example gemini-3.7-flash and claude-opus-5) are defined once per skill in scripts/workflow/models.py. Codex uses codex-default by default, deliberately delegating to the authenticated CLI's supported model rather than pinning an account-incompatible ID. Concrete retired or superseded model IDs are refused before a run starts. Multi-model orchestration is a design choice and does not guarantee better results. Treat model outputs as untrusted and review them, and be mindful of prompt-injection when feeding in third-party content.
MIT © 2026 buddypia. See LICENSE. Each skill also bundles its own LICENSE and a per-skill References & Attribution section crediting the research that inspired it.