Kushar Raj Kashyap

Git & GitHub discipline

Every project gets its own repository from day one. Each unit of work becomes a clean commit pushed to GitHub, so there is always a saved version to revert to.

• New repo per projectIsolated history, nothing tangled.
• Clean commit messagesHistory you can read months later.
• Push after each unitStatus is never lost; revert anytime.

git add -A
git commit -m "feat: add level select screen"
git push

git log --oneline
git revert <commit-sha>

Caveman — fewer words, fewer tokens

A Claude Code plugin that makes Claude answer in terse "caveman" style. Same work, far shorter replies — so a session goes further before hitting limits.

claude plugin marketplace add JuliusBrussee/caveman
claude plugin install caveman@caveman

Source: github.com/JuliusBrussee/caveman

Graphify — a project knowledge graph

Graphify maps a project into a knowledge graph Claude can query, instead of re-reading every file. Less context to load means fewer tokens. It works on any folder — a codebase, notes, papers.

pip install graphifyy && graphify install

pipx install graphifyy && graphify install

/graphify .

Source: github.com/safishamsi/graphify

Karpathy-inspired guidelines — one CLAUDE.md

A single CLAUDE.md file, derived from Andrej Karpathy's observations on LLM coding pitfalls, that sets four principles for how Claude should work. I append it to the project's own CLAUDE.md.

echo "" >> CLAUDE.md
curl https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md >> CLAUDE.md

Source: github.com/forrestchang/andrej-karpathy-skills

The .claude folder

Every project gets the same structure, so Claude always knows where its team, commands, rules and skills live.

.claude/
├── agents/      ← AI team members
├── commands/    ← custom slash commands
├── hooks/       ← rules Claude MUST follow
├── rules/       ← context-aware instructions
├── skills/      ← situational intelligence
└── settings.json ← the control center
CLAUDE.md         ← the project brain (root)

mkdir -p .claude/{agents,commands,hooks,rules,skills/frontend-design}
touch .claude/agents/{code-reviewer,debugger,test-writer,refactorer,doc-writer,security-auditor}.md
touch .claude/commands/{fix-issue,deploy,pr-review}.md
touch .claude/hooks/{pre-commit,lint-on-save}.sh
touch .claude/rules/{frontend,database,api}.md
touch .claude/skills/frontend-design/SKILL.md
touch .claude/settings.json CLAUDE.md
chmod +x .claude/hooks/*.sh

claude-mem — persistent memory across sessions

A Claude Code plugin that captures what happens during a session, compresses it with AI and brings the relevant context back next time. No more re-explaining the project every morning.

• CapturesWhat Claude did during the session.
• CompressesInto short, reusable memory.
• RestoresRelevant context in future sessions.

Source: github.com/thedotmack/claude-mem · Docs: docs.claude-mem.ai

Hooks — rules the harness enforces

Hooks run my scripts at fixed points in Claude's lifecycle, so guardrails don't depend on the model remembering them.