Week 1 with Claude Code, I typed: "improve this project."

Claude read 4 files, edited 2, broke 1. 34,000 tokens. Nothing was fixed.
I typed /cost. $0.09 for zero work.

After a month of figuring out what actually works, the same changes now ship with 60–70% fewer tokens. This repo is the condensed version.

The 3 Things That Mattered Most

1. Specific prompts cost 8.5× less

- "improve this project"                             → 34,000 tokens
+ "add RFC 5322 validation to validateEmail()
+  at src/auth/login.ts:87"                          →  4,000 tokens

Same outcome. One-eighth the cost.

2. /compact is useless without an instruction

# Trims almost nothing
/compact

# Cuts 70% of context, quality unchanged
/compact Keep only changed functions and relevant test output

In one long session this took me from 82,000 → 11,000 tokens.

3. .claudeignore is doing nothing for you right now

node_modules/
dist/
*.log
coverage/
*.lock

Six lines. File reads drop 3× to 10× depending on the project.

What's In This Repo

Skill Library

The Workflow

flowchart LR
    A["Vague prompt"] -.->|Wastes tokens| X(["Fail"])
    B["CLAUDE.md<br/>+ .claudeignore"] --> C["Specific prompt"]
    C --> D["Run a skill"]
    D -->|Long session| E["compact with hint"]
    E --> F(["Ship"])
    D --> F

    style X fill:#ff6b6b,stroke:#000,color:#fff
    style F fill:#51cf66,stroke:#000,color:#000
    style B fill:#7B4FFF,stroke:#000,color:#fff

Quick Start

One command (installs core skills, .claudeignore, and CLAUDE.md scaffold):

curl -fsSL https://raw.githubusercontent.com/berkcangumusisik/claude-code-practices/main/install.sh | bash

Flags: --all for every skill, --skills a,b,c to pick.

Or manually:

git clone https://github.com/berkcangumusisik/claude-code-practices.git
cp -r claude-code-practices/skills ./.claude/skills
claude
> /cost-audit

Where To Start

FAQ

Yes. Every skill and pattern here is model-agnostic — they're about how you prompt Claude Code, not which model it's pointed at.

No. Start with 3–5 you'd actually use this week. Add more as the pain points show up.

No. It only controls what Claude reads — your build and editor are unaffected.

Run /cost before and after a session. The numbers in this README are from my own /cost outputs, not estimates.

Yes. Commit .claude/skills/ and a CLAUDE.md to your repo and the whole team picks them up automatically.

Support the project

If this repo saved you tokens or hours, there are three ways to give back:

• Star the repo so the next person finds it
• Sponsor on GitHub to keep new skills and templates coming
• Open a PR with your own patterns (see below)

Contributing

Found a better pattern, wrote a skill, spotted a bug? PRs welcome. See CONTRIBUTING.md.

Star History