Never run Claude Code inside a freshly cloned repo before checking .claude/settings․json first.
A malicious SessionStart hook executes shell commands before you even type a prompt.
Do you audit repo hooks with cat first or run inside a container?
A 400-line CLAUDE․md makes Claude Code worse, not smarter.
After testing dozens of repo configs, stripping mine down to these 12 lines cut wasted token burn by ~45% and stopped random file edits:
# WORKFLOW & BLAST RADIUS
- Search with rg/glob before reading any file over 200 lines.
- Touch ONLY files directly required by the prompt. Never refactor adjacent code.
- If a fix requires editing >3 files, list the files and wait for approval.
# TOOL OUTPUT DIET
- Never dump full JSON logs or build outputs into context; pipe through head -n 40 or jq.
- Run single targeted test files, never the full test suite unless asked.
# VERIFICATION GATE
- Every code edit must be followed by a typecheck or test command before reporting done.
- When a test fails twice for the same reason, stop and state the root cause.
Keep project-specific commands in package․json scripts, and leave CLAUDE․md purely for behavioral guardrails.
An AI agent that grades its own homework will eventually break prod.
6 surging GitHub repos (+4,810 stars in 24h, 134k total) that wire Claude Code and Codex into a gated delivery pipeline:
01 backnotprop/plannotator (12․2k stars) -> Annotate and approve the implementation plan before any file is edited
$ plannotator review plan․md
02 GoogleChrome/modern-web-guidance-src (7․7k stars) -> Official Chrome team skill that injects current web specs so agents stop writing outdated CSS and DOM code
$ npx modern-web-guidance install
03 Tencent/LoopForge (2․1k stars) -> Self-hosted workflow runner chaining Claude Code and Codex with human approval gates
$ npx -y loopforge-cli install claude
04 ChromeDevTools/chrome-devtools-mcp (57․4k stars) -> Let your agent inspect live DOM state, console errors, and network waterfalls directly
$ npm i -g chrome-devtools-mcp
05 alibaba/open-code-review (52․6k stars) -> Hybrid reviewer running deterministic NPE, XSS, and race-condition rules before LLM diff review
$ ocr delegate rule src/main․go
06 NVIDIA/SkillEvaluator (2․2k stars) -> Multi-tier QA gate that catches conflicting SKILL․md files before they degrade agent output
$ pip install skillevaluator
Plan -> Specs -> Execute -> Inspect -> Review -> Eval.
Which stage in your workflow is still manual right now: plan sign-off, browser check, or PR review?
Letting the same Claude Code session write a feature and grade its own PR is how subtle bugs ship to main.
One worktree writes the diff. A second clean session tries to break it.
Do you review agent PRs in a fresh session or the same chat?
Half the Coding Agent repos trending on GitHub today do one thing: keep junk out of the context window.
Headroom strips 60-95% of raw JSON logs. Caveman cuts 65% of output tokens.
Most teams do not have a model problem. They treat context like /tmp.
wild contrast on my feed today: claude code is shipping animated desktop themes while cursor is publishing a postmortem on leaked client data.
pretty clear which tab is staying open on my machine this weekend.
most people blame opus 5.5 for getting sloppy after 30 minutes, when their context window is just full of dead debugging turns.
use `/btw` for side questions and `/fork` before risky refactors.
6 built-in claude code commands that stop context rot:
if you use Claude Code, go check your `.gitignore` right now.
clicking "yes, don't ask again" on a bash command with an inline token dumps the raw secret straight into `.claude/settings.local.json`.
1,122 public GitHub repos already leaked 1,379 live keys. Bad default tbh.