eval-first-coding-plus

Built on affaan-m/ECC by @affaan-m (245,633 stars, MIT). All credit for the original idea to them. This fork improves and repackages it; upstream license preserved in UPSTREAM_LICENSE.

A Claude Code skill that helps AI write safer code with tests, small tasks, risk checks, and smart model choice.

Made for developers who use Claude Code to plan, build, test, and review software.

🤔 Why

AI can write code fast. It can also miss risks.

Tests may pass but fail to check the real problem. Large tasks can hide bad changes. Model costs can grow with each retry.

This skill gives Claude Code a clear loop. Define success, test first, make one small change, check the result, and stop when the work is unsafe or unclear.

It is a single skill file with no added tools or packages. It fits into your current Claude Code flow.

⚡ Install

Replace YOUR_GITHUB_USER with the repo owner, then run:

mkdir -p ~/.claude/skills/agentic-engineering && curl -fsSL https://raw.githubusercontent.com/YOUR_GITHUB_USER/eval-first-coding-plus/main/skill/SKILL.md -o ~/.claude/skills/agentic-engineering/SKILL.md

Restart Claude Code after the install.

🛠️ Usage

Ask Claude Code to use the skill for a coding task:

Use agentic-engineering to add a rule that blocks an empty user name.

Claude should first give you a small task plan like this:

Goal: Block an empty user name.

Main risk: A valid name may stop working.

Done rules:
- An empty name returns a clear error.
- A valid name still works.
- All old user tests pass.

Allowed files:
- The user check
- Its tests

Plan:
1. Add a test for an empty name.
2. Run it and save the failure.
3. Make one small code change.
4. Run the new and old tests.
5. Review edge cases and record the result.

The skill also tells Claude when to stop and ask you for help. This includes unclear access rules, data risk, missing rollback steps, and product or legal choices.

🔄 Core workflow

1. Write clear done rules.
2. Split the work into small tasks.
3. Pick the lowest model level that can do the work well.
4. Test before and after each change.
5. Review risky code, even when tests pass.
6. Save a short handoff between major phases.
7. Stop after two failed tries with the same plan.

✨ What we changed vs upstream

• Rewrites the skill from Japanese into clearer, simpler English while preserving its core eval-first, decomposition, model-routing, review, session, and cost principles.
• Expands the eval-first loop with repeatable checks, failure capture, test-quality safeguards, and explicit rules against changing tests merely to pass bad code.
• Makes task decomposition more actionable by requiring one goal, one risk, independent verification, done criteria, and an allowed-file list per task.
• Adds stronger human-safety controls, including stop-and-escalate conditions for data, access, rollback, conflicting requirements, and product/legal decisions.
• Introduces practical operating details absent from the original: handoff contents, retry limits, model downgrade guidance, and local-only cost tracking.

📄 License

This project is licensed under the MIT License.

The upstream MIT license and credit are preserved in UPSTREAM_LICENSE.