What Survives in AI: Skill Set

Tested shortcuts and Claude Code skills for people building with AI. Every trick has a date it was last checked, and every claim links to its source. From the What Survives in AI newsletter.

⭐ Star the repo to get new skills as they land: a new tested trick every day.

Install a skill (Claude Code)

git clone https://github.com/prandata/wsai-skill-set
cp -r wsai-skill-set/skills/claude-code-memory ~/.claude/skills/
cp -r wsai-skill-set/skills/ssh-key-audit ~/.claude/skills/
cp -r wsai-skill-set/skills/wsai-skill-set ~/.claude/skills/ && cp wsai-skill-set/tricks.yaml ~/.claude/skills/wsai-skill-set/

Skill	What it does
claude-code-memory	Sets up the four files that stop Claude Code forgetting your project: CLAUDE.md, AGENTS.md, a skill, and a hook that blocks dangerous commands.
ssh-key-audit	Read-only audit of the SSH keys on your GitHub account: which keys at github.com/<you>.keys have no .pub file on this machine, and which are DSA or RSA under 3072 bits. Changes nothing. 66 offline cases in tests/test_ssh_key_audit.py.
wsai-skill-set	Gives your agent 55+ tested shortcuts: change-one-word GitHub URLs, repo-to-AI-context tools, terminal one-liners.

Agents (Claude Code subagents)

mkdir -p .claude/agents && cp wsai-skill-set/agents/repro-reviewer/repro-reviewer.md .claude/agents/

Agent	What it does
repro-reviewer	Read-only code reviewer that must reproduce every finding before it reports it: each one comes with the command it ran and the quoted output, a claim it can't run blocks a PASS, and it ends with VERDICT: PASS|FIX @ <sha>. Three saved example runs (a real bug, a false alarm, one it can't run) and offline checks in tests/test_repro_reviewer.py.

MCP servers

Server	What it does
notes-readonly	Lets Claude Code list, search and read one folder of your Markdown notes, and nothing else. No write tool; ../, absolute paths and every symlink are refused; a per-file size cap. 99 lines on the official MCP Python SDK, 34 tests. Install: claude mcp add --scope user --transport stdio notes -- uv run --script /path/to/server.py ~/notes.

Templates

templates/claude-code-memory/: copy-paste CLAUDE.md, AGENTS.md, SKILL.md and a PreToolUse hook (block-dangerous.sh blocks recursive force-deletes of /, ~, * and ., force-pushes, and shell commands that mention .env; 143 cases in tests/test_hook.py, including quoting tricks, long flags and bash -c) plus a permissions.deny rule so the Read tool cannot open .env either.

Loops

loops/overnight-handoff/: run Claude Code on one bounded task overnight. Each claude -p session stops cleanly at a time or spend limit and leaves a HANDOFF.md (state, done, next, open questions); the next session resumes from it, and a resume check refuses to start if the handoff is malformed. No shell for the agent, no git pushes, a dry run first. The sample was run twice, the first run interrupted; both logs are in the folder.

The trick list

tricks.yaml: every trick with how to use it, what you get, status and the date it was last checked. A GitHub Action re-checks every link each morning (scripts/verify.py); two failures in a row mark a trick broken.

Contribute

Add a trick to tricks.yaml with status: partial and a check_url, and open a PR. We promote it to verified once we've seen it work in a browser. Claims need a source.

How this is made

Pranjal writes What Survives in AI. AI agents help gather and check the material; he reviews what's published.

Known limits of the hook

It is a guardrail, not a sandbox. It parses commands the way a shell does, but it does not run them, so it cannot see:

Not covered	Example
Variables and command output	d=/; rm -rf $d
Other interpreters	python3 -c "shutil.rmtree('/')"
Encoded scripts run by another interpreter	echo ... | base64 -d | python3
Aliases and functions	alias x='rm -rf /'

Simple rule, on purpose: any command whose program is a shell interpreter (bash, sh, zsh, dash, ksh, eval, source, .), and any heredoc into one, is blocked without looking inside. Intended blocks, because the program is a shell: curl ... | bash (and | sh, which also covers echo ... | base64 -d | sh), . ~/.nvm/nvm.sh and source ~/.zshrc. A command line that ends in &&, ;, || or | continues on the next line, and that line is checked as a new command. Known false positives: bash script.sh (run ./script.sh), source .venv/bin/activate (call .venv/bin/python directly), sh -c '...' in scripts.

For a hard block, use permission deny rules and the sandbox (see the Claude Code permissions and sandboxing docs). Deliberate obfuscation beyond these is out of scope by design.