woz-benchmark

WOZCODE Savings Benchmark

Run a side-by-side comparison of WOZCODE vs vanilla Claude Code on the user's own codebase. Each prompt runs twice against a fresh copy of the repo with git reset --hard between runs, so the target MUST be a clean git repo.

TRIGGER: "compare woz", "how much does woz save", "benchmark woz", "woz vs claude", "show me the savings", "is woz worth it", or /woz-benchmark.

Prerequisites

• User logged in to WOZCODE (if not, stop and ask them to /woz-login).
• Target directory is a git repo with a clean working tree.

Steps

1. Gather inputs — BE BRIEF

Ask for all three in ONE short message (< 10 lines). Do not re-explain what the benchmark does — the user already invoked it.

1. Target directory — absolute path to a clean git repo to run the test on.
2. Prompts — 2–10 real coding tasks. Tell them briefly: "meaty feature/refactor/bugfix work, not one-liners — trivial prompts hide WOZCODE's advantage". If they don't have prompts in mind, offer to suggest some after looking at their repo.
3. Environment setup (optional) — one line: "Anything Claude needs already in place (DB seeded, services running, credentials in .env)? Skip if the repo is self-contained."

Do NOT ask about the model. Default to opus in the YAML config. Only switch to sonnet or haiku if the user volunteers a different choice in their answer (e.g. "use sonnet" or "try it on haiku").

Keep examples OUT of the user message unless they ask for help picking prompts. The user doesn't need 4 bullet points of good-vs-bad prompt examples.

2. Validate the target

Before writing any config, verify the target is usable:

test -d <target>
git -C <target> rev-parse --git-dir
git -C <target> status --porcelain

If the directory doesn't exist, isn't a git repo, or has uncommitted changes, STOP and tell the user how to fix it (e.g. "please commit or stash your changes — the benchmark resets the repo between runs and would lose your work").

3. Write a temporary benchmark config

Use the Write tool to create a YAML file at /tmp/woz-benchmark-<timestamp>.yaml (get the timestamp from date +%s). Format:

model: opus
maxTurns: 15
prompts:
  - "first prompt from the user"
  - "second prompt from the user"
setup:
  commands:
    - "curl -L https://example.com/dataset.csv -o data/sample.csv"
    - "psql $DATABASE_URL -f seed.sql"

Notes:

• Default to model: opus. Only use a different model if the user volunteered one in their answer.
• Quote every prompt string. If a prompt contains a double quote, escape it with \".
• Omit the entire setup: block if the user didn't give any environment setup commands.
• Keep maxTurns: 15 as a safety cap so a single prompt can't run away.

4. Run the benchmark

One-line warning: "This'll take several minutes — each prompt runs twice." Then run:

node --no-warnings=ExperimentalWarning ${CLAUDE_PLUGIN_ROOT}/scripts/benchmark.js --target <target> --config <yaml-path> --user-env

--user-env loads the user's project CLAUDE.md hierarchy on BOTH sides. Do NOT pass --screenshots, --codex, --judge, or --trace.

5. Present the results as a savings report

The benchmark prints a detailed text report at the end. Relay the full report to the user, then add a clear, sales-oriented savings summary at the top. Compute the deltas from the report's totals:

💰 WOZCODE Savings Summary
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  Cost saved:       $X.XX  (Y% cheaper)
  Tokens saved:     X,XXX  (Y% fewer)
  Turns saved:      N      (Y% fewer)
  Time saved:       X min  (Y% faster)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Frame the numbers positively. If WOZCODE was slower or more expensive on a specific prompt, call it out honestly but note the aggregate.

Finally, tell the user where the detailed JSON report was saved (the benchmark prints this path).

Tips

• If the user has no prompts in mind, read a few files in their repo and suggest 2-3 realistic tasks tailored to what you see.
• The temp YAML file is safe to leave in /tmp — the OS cleans it up.
• If the user wants to re-run with different prompts, just generate a new YAML and call the script again.
• If the benchmark fails midway, its cleanup handler resets the repo to its original HEAD — the user's work is safe.