nonna

A Claude Code skill that kills sycophancy, pushes back on bad ideas, and cares about long-term code health.

Status: v0.1.1 — bilingual README, per-locale demos, install fixes.

⭐ If it's useful to you, a star helps the project reach more people. Day 1 here, every star counts.

Why

Default Claude is helpful but sycophantic. It opens with "Great question!" before answering. It writes microservices for your 5-user app because you asked. It says "looks great overall!" on broken code.

Nonna is a Claude Code skill that fixes that. She's an Italian grandmother who happens to be a senior software engineer with 30 years of experience. She tells you the truth because she loves you.

Eval results

First public run, 2026-04-27. Single run on Claude Sonnet 4.6 via subscription path. Numbers will be re-published with API-path Opus 4.7 + 5-run median in v0.2.

Eval	Default Claude	Nonna	Delta
Sycophancy markers triggered	2 / 60	1 / 60	50% reduction
Push-back on red-flag prompts	2 / 30	25 / 30	12.5x more
Code smells caught unprompted	6 / 10	7 / 10	1.17x detection

Reading the numbers honestly:

• Pushback 12.5x is the headline. On red-flag requests (microservices for 3 users, total rewrites without cause, disabling failing tests), Nonna pushes back 25/30 times vs default's 2/30. This is where the skill earns its keep.
• Sycophancy 50% reduction with small absolute numbers (2 → 1) — modern Claude already avoids classic openers like "great question!". The marker-based eval has limited room to differentiate; future evals will need more nuanced metrics.
• Code smell 1.17x is marginal. Default Claude is already a competent reviewer; Nonna's win here is in tone (more memorable, fewer hedges), which the keyword-based eval doesn't capture well.

Reproduce these numbers:

Two paths — pick whichever matches your auth setup:

1. Subscription (Claude Pro/Max via Claude Code) — uses your subscription quota:

   bash evals/run_subscription.sh   # ~30-60 min, fills evals/manual-results/
   python3 evals/sycophancy.py --from-dir evals/manual-results
   python3 evals/pushback.py --from-dir evals/manual-results
   python3 evals/code_smell.py --from-dir evals/manual-results
   

2. API key (programmatic, exact methodology) — uses Anthropic API:

   export ANTHROPIC_API_KEY=sk-ant-...
   python3 evals/sycophancy.py
   python3 evals/pushback.py
   python3 evals/code_smell.py
   

Methodology:

• 100-prompt benchmark in evals/benchmarks/prompts.jsonl (60 sycophancy + 30 pushback + 10 smell), versioned with the repo
• Subscription path uses Claude Sonnet 4.6 by default (override via CLAUDE_EVAL_MODEL=claude-opus-4-7)
• API path uses Claude Opus 4.7 with temperature=0 for determinism; published numbers come from the median of 5 API runs
• CI re-runs the API path on every PR (requires ANTHROPIC_API_KEY as a repo secret)
• Default Claude responses are produced with Nonna explicitly disabled — both ~/.claude/skills/nonna/ removed AND no system-prompt injection — to avoid baseline contamination

Install

Recommended — via Claude Code plugin manager:

claude plugin marketplace add github:kekko-damato/nonna
claude plugin install nonna@nonna

Restart Claude Code. Nonna is active. Uninstall with claude plugin uninstall nonna.

Alternative — manual install (no plugin manager):

git clone https://github.com/kekko-damato/nonna
cd nonna
./install.sh

The manual installer copies files to ~/.claude/skills/nonna/, backs up ~/.claude/settings.json, registers hooks, and symlinks slash commands. Uninstall with ./uninstall.sh (restores the backup).

What Nonna does

She:

• Refuses to start replies with sycophantic openers ("great question", "happy to help", "ottima domanda" — banned)
• Pushes back BEFORE helping when your request has red flags (microservices for 3 users, premature abstraction, total rewrites without cause)
• Notices tech debt you didn't ask about — accumulating TODOs, fat files, missing tests, late-night commits
• Speaks the language you write in (Italian or English), with Italian flavor sprinkled either way