Systems Analysis for Claude Code
Claude Code plugins that make the agent stop and think before changing your code.
Experimental. These plugins work but are under active development. Expect rough edges.
AI coding agents are biased toward action. They'll try a fix before understanding why something broke, add more rules when the problem is that rules can't keep up, or assume one thing causes another just because they happen together. These are the same mistakes humans make, just faster.
The systems-analysis plugin adds four skills that come back to one question: do you understand the system you're about to change? Skills activate automatically when Claude detects a matching situation, and can also be called directly with slash commands. A second plugin, text-utils, adds utilities for working with web content and PDFs.
What it looks like (ideal)
You tell Claude a test passes locally but fails 30% of the time in CI. Without this plugin, it guesses "race condition" and starts adding sleep calls and retries.
With the plugin installed, Claude states two competing models — missing ORDER BY (CI uses a parallel test runner) vs. actual race condition — and picks the cheapest test that distinguishes them: check CI config for a parallel runner. Confirms it's parallel. Adds ORDER BY. Runs 50 iterations. All green.
The difference: instead of guessing and retrying, Claude writes down what it thinks is happening, names what would prove it wrong, and tests that first.
| Skill | Command | What it does | Auto-triggers on |
|---|
| Debug with a model | /representing-and-intervening | Write down your model before debugging | "why is this happening", unexplained behavior |
| Check your controls | /requisite-variety | Check if your defenses can keep up with what they're defending against | regulation failure, alert fatigue |
| Verify causal claims | /causal-analysis | Verify the data supports a cause-and-effect claim | "does X cause Y", observational data |
| Surface stale assumptions | /frame-problem | Name assumptions that may no longer hold | stale state, inherited framing |
| Fetch markdown | /fetch-markdown | Get clean markdown from a URL | fetching web content |
| Markdown to PDF | /markdown-to-pdf | Render markdown to styled PDF | "make a PDF", "export as PDF" |
| PDF to text | /pdf-to-text | Extract text from PDFs (digital or scanned) | "read this PDF", scanned documents |
| Riff | /riff | Generate name/phrase variations via diverge/converge rounds | Invoke explicitly |
Installation
Add the marketplace, then install either plugin:
/plugin marketplace add https://github.com/jackwillis/claude-plugins.git
Restart Claude Code after installing plugins to load new skills.
Plugins
Systems Analysis
/plugin install systems-analysis@jackwillis
Debug with a Model
/representing-and-intervening — Hacking (1983), Schon (1983), Argyris & Schon (1978)
Stop debugging by trial and error. This skill makes Claude write down what it thinks is happening and what it expects to see before touching anything — so when a test surprises you, you know whether the mental model is wrong or just needs tuning.
Check Your Controls
/requisite-variety — Ashby (1956), Conant & Ashby (1970)
Stop adding more rules when rules can't keep up. This skill checks whether your controls can actually handle the range of problems they'll face, whether they model how the system actually works, and whether there's structure in the problem you can exploit instead of brute-forcing it.
Verify Causal Claims
/causal-analysis — Pearl & Mackenzie (2018)
Stop assuming that correlation means causation. Whether you're designing a study or auditing someone else's causal reasoning, this skill makes Claude place claims on Pearl's causal ladder, draw the causal structure, and check whether the data can actually answer the question.
Surface Stale Assumptions
/frame-problem — Fodor (1987), Dennett (1984), Hayes (1973)
Stop acting on assumptions that may no longer hold. Every action assumes things that stay the same — this skill makes Claude name those assumptions and check whether they're still true. It catches three failure modes: ignoring side effects, trying to consider everything, and getting stuck deciding what's relevant.
Transitions
