lord-eagle/skills

Public skill collection. Three themes:

1. Solo MCP multi-agent orchestration — solo-orchestration and solo-multi-agent-brainstorm. Built on Solo MCP.
2. GitHub PR review tooling — review-triage. Independent skill, no Solo dependency.
3. Editorial cleanup of marketing pages — reduce-to-max. Audits a running website for information overload and proposes editorial cuts.

What's inside

Solo multi-agent skills

• solo-orchestration — Foundation library for any Solo multi-agent workflow. The core value is push-based reporting: spawned agents call back on terminal events instead of the orchestrator polling on a timer. Zero token cost while nothing happens. This works with a single Claude agent — no multi-vendor setup required. Cross-project scoping and the multi-vendor preflight are bolt-on accelerators, not prerequisites.
• solo-multi-agent-brainstorm — Run a moderated panel of 3 AI agents on a hard decision. Mix vendors (Claude, Codex, Gemini, …) for cross-vendor disagreement. Situational and expensive — a 3-panelist × 3-round panel burns real tokens. Reserve for high-stakes, hard-to-reverse decisions (architecture calls, cross-repo migrations), not routine questions a single agent answers fine.

If you only use Claude: you need none of the multi-vendor machinery. Solo auto-wires Solo MCP into Claude. Install solo-orchestration, use the push-based reporting patterns, and skip every "multi-vendor" / "preflight" section below.

Standalone skills

• review-triage — Triage GitHub PR review feedback into Must fix / Should fix / Nice to have / Nit; bundles non-blocking findings into one issue per severity. Not Solo-specific — works on any GitHub repo with gh CLI access. Project board placement is opt-in per repo via .github/review-triage.yml; without that file, issues are created with priority labels and no project placement.
• reduce-to-max — Audit a running website for information overload, redundant sections, and comprehension gaps. Recommends or applies editorial cuts so each page does its one job with the minimum sections and words required. Markdown-first skill with dev-browser helpers for section extraction, page-level metrics, content scanning, route discovery, and live browser overlay during apply mode.

What you get from a multi-agent brainstorm

A panel of 3 AI agents argue a decision for 3 rounds. Orchestrator synthesizes between rounds. You get a verdict that includes:

• TL;DR of the locked decisions
• Per-panelist POSITION / TOP_RISK / DELTA quoted verbatim (no smoothing-away of dissent)
• Diversity assessment — single-vendor panels are flagged as echo-chamber risk
• Token-cost estimate, panel composition, convergence trigger

Use it for: cross-repo decisions, architecture calls, anything where one LLM's opinion is not enough.

Install

Via Scribe (recommended)

scribe install solo-orchestration solo-multi-agent-brainstorm

Scribe handles linking into Claude Code, Cursor, Codex, etc.

Manual

Clone this repo, then symlink the skill directories into your agent's skills folder. For Claude Code:

git clone https://github.com/lord-eagle/skills.git ~/skills-src
ln -s ~/skills-src/solo-orchestration         ~/.claude/skills/solo-orchestration
ln -s ~/skills-src/solo-multi-agent-brainstorm ~/.claude/skills/solo-multi-agent-brainstorm

Setup for multi-vendor brainstorms

Skip this whole section if you only use Claude. Solo's default Claude wiring is enough for push-based orchestration and single-vendor panels. Multi-vendor setup only matters when you deliberately want cross-vendor disagreement on a panel.

Solo MCP is auto-wired only into Claude. For Codex, Gemini, Amp, or OpenCode panelists, the runtime's config needs an mcp_servers.solo entry.

See docs/multi-vendor-setup.md for the per-vendor config snippets and known footguns.

The solo-orchestration skill includes a Multi-Vendor Preflight that patches the configs idempotently and probes each runtime — you don't need to do this by hand once the skill is loaded.

Quickstart

After install, in Claude Code or any Solo-aware agent:

Brainstorm with a 3-vendor panel: should I migrate from Postgres to ClickHouse for the analytics workload?

The skill picks up: solo-multi-agent-brainstorm runs preflight, spawns Claude + Codex + Gemini panelists (or whichever are OK), runs 3 rounds, synthesizes, archives the verdict.

Status

• solo-orchestration — v1.2 (multi-vendor preflight added 2026-05-05)
• solo-multi-agent-brainstorm — v1.2 (multi-vendor enabled 2026-05-05)
• review-triage — generic GitHub-PR tooling, opt-in project placement via .github/review-triage.yml
• reduce-to-max — v0.7 (route discovery added 2026-05-10)

Active development. Open issues track planned work:

• Mid-flight panel growth (NEED_PANELIST) — issue #2 Feature B
• Verdict template with DoD + reasoning trail — issue #5
• Per-round synthesis pad persistence — issue #7

Repo layout

.
├── solo-orchestration/         # foundation library skill
│   ├── SKILL.md
│   └── versions/               # version snapshots
├── solo-multi-agent-brainstorm/ # panel discussion skill
│   ├── SKILL.md
│   └── versions/
├── review-triage/
├── reduce-to-max/                # editorial cleanup skill
│   ├── SKILL.md
│   ├── principles/               # rule definitions (C, X, TS, CL, MO, DC, HI, F)
│   ├── workflows/                # per-phase workflow modules
│   ├── scripts/                  # dev-browser helpers
│   ├── evals/                    # expected-findings fixtures
│   └── references/               # citations
└── docs/
    └── multi-vendor-setup.md   # per-vendor MCP wiring

Each skill is self-contained — install one without the others if you want.

Contributing

Issues + PRs welcome. See open issues for what's planned. The Multi-Vendor Preflight footguns section (in solo-orchestration/SKILL.md) is a living document — if you hit a new vendor edge case, file an issue with the probe output.

License

See LICENSE.