system-comparison

A Claude Code skill that turns "compare these two systems" into an editorial-quality PDF with forced verdicts, a ranked steal list, and a fitness scorecard.

The Problem

When you evaluate an external tool ecosystem, you get one of two outputs:

• A vague "both have strengths" document that doesn't tell you what to do
• A biased teardown that misses what the other system actually does well

Neither is useful. The real question is: what do we steal, and what do we protect?

What This Does

Spawns 4 parallel research agents. Each reads actual source files, not directory names. Each is forced to commit to a WHO WINS verdict per dimension — no weasel words.

3 domain agents cover capability tracks (infrastructure, copy/QA, measurement/automation — or whatever 3 tracks fit your domain).

1 fitness agent scores both systems on 10 structural vectors:

Vector	What It Measures
Prerequisites	Dependencies, setup complexity, what must exist before it runs
Cost Impact	Token burn, API calls, agent spawns per typical run
Quality Control	Does it validate output, or does quality depend on the model?
User Experience	Friction level, interaction model, output format fit
Skill Standard Compliance	Hits the canonical good-skill bar (scoped, testable, clear triggers)
Security	API key handling, prompt injection surface, data leaving workspace
System Architecture Fit	Skill vs pipeline vs full system — is the scope right?
Workspace Compatibility	Plugin namespace fit, routing conventions
Trigger Precision	Unambiguous triggers, no overlap with adjacent skills
Maintenance Decay Rate	How fast does this rot? (scraped data vs stable internal conventions)

Phase 1 is a grill-me style intake — one question at a time, recommended answer provided, reads workspace before asking. Phase 2 spawns all 4 agents simultaneously. Phase 3 synthesizes domain matrix + fitness scorecard + ranked steal list. Phase 4 generates the PDF.

The deliverable is a PDF. Magazine layout. Cover page. SVG strength diagram. Per-track comparison blocks (orange winner, gray loser). Fitness scorecard. Full matrix. Numbered steal list ranked by operational impact. Moats callout.

Three Modes

• external-vs-internal — external repo or competitor vs your internal stack. Gap-finding bias.
• skill-vs-skill — two skills in your workspace. Decide keep/merge/deprecate.
• skill-vs-system — is this skill actually a system? Produces an architecture verdict.

Install

Copy SKILL.md into your Claude Code workspace:

mkdir -p .claude/skills/system-comparison
cp SKILL.md .claude/skills/system-comparison/SKILL.md

Requires puppeteer for PDF generation:

bun add puppeteer
# or: npm install puppeteer

Trigger Phrases

Say any of these to invoke the skill:

• "compare our skills to [X]"
• "gap assessment between [A] and [B]"
• "what should we steal from [X]"
• "evaluate this repo"
• "is this a skill or a system"
• "benchmark against [X]"

Example Output

The first validated run compared LeadGrow's internal outbound stack (GTM orchestrator + 20 lg-outbound skills) against coldoutboundskills (28 open-source skills, 1,000+ campaigns).

Finding: LeadGrow was back-loaded (strong post-launch, weak pre-launch). coldoutboundskills was front-loaded (strong pre-launch, weak post-launch). Symmetrical gap. 12 specific steal items ranked by operational impact.

That's the output this skill produces on every run.

Requirements

• Claude Code with agent spawning support
• bun or node + puppeteer for PDF generation
• Google Fonts access for PDF rendering (Poppins + Lora via CDN)

Design System

PDF uses cream (#faf9f5) / black (#141413) / orange (#d97757). Poppins for headings and labels, Lora for body copy. Letter size. The HTML component library is documented in SKILL.md under Phase 4.

---

Built by LeadGrow — GTM agency running Claude Code as operating infrastructure.