Skills are inspectable source packages. Most contain instructions only. Some include scripts for deterministic checks. Installation does not execute those scripts. An agent may execute them during use, subject to the host's permissions. Review all instructions and executable files before installation.

Everything this repository ships inside a skill folder is source you can read — never a binary, never minified or obfuscated. (Your own machine may add to that: running a skill that ships a script leaves compiled __pycache__/*.pyc behind. Your interpreter wrote those, nobody shipped them, and commitment 3 says exactly when they are fine.) Installing one:

• runs nothing at install time — no install hooks, no build step, no code executed to put the files in place;
• copies the whole folder onto your machine, scripts included, so what you read in this repository is what you have locally;
• adds instructions your AI assistant will read and may act on. Your assistant can run commands, so a malicious skill could instruct it to do harmful things. That is the real threat model for every skill collection, including this one.

1. Human-readable by design. Every skill's instructions are short plain English — if you cannot read a skill's markdown end to end in a few minutes, we consider that a defect, so report it. Two skills also ship Python, and that is longer: im-down and im-up carry a validator and its test suite. They are held to being readable and commented rather than to the few-minutes bar, and commitment 3 covers them.

2. No fetch-and-execute. No skill in this collection instructs the agent to download and run remote code, pipe URLs to a shell, or fetch instructions from an external source at run time.

3. Any code a skill ships is readable source, and nothing here fetches or generates it. Some skills bundle a script, committed as source in this repository. A skill's own SKILL.md names the scripts it asks the agent to run, so you can read them before you run them. Not every shipped .py is one of those — im-down and im-up also ship their test suites, which our CI runs and no skill invokes. Read those too; they are part of what you installed.

   CI detects any file inside a skill folder that is not .md, .txt, .py or .json — plus compiled Python, which is allowed only as __pycache__/<name>.pyc sitting directly beside the <name>.py it came from. Adding a format is a reviewed change to this policy, not a silent commit.

   ⚠ That check detects violations; it does not prevent them. main has no required status checks, so a nonzero exit is a signal, not a gate. Which is the point of the next paragraph.

   Do not take our word for it — check the copy you actually have. Point this at the folder you installed into: npx skills add writes to .claude/skills/<name> under the directory you ran it in unless you pass --global, so confirm where yours landed rather than assuming.

   # 1. Nothing but the declared formats. Compiled Python is step 2.
   find -L <the folder you installed> -type f \
     ! -name '*.md' ! -name '*.txt' ! -name '*.py' ! -name '*.json' \
     ! \( -path '*/__pycache__/*.pyc' ! -path '*/__pycache__/*/*' \)
   
   # 2. Every compiled file sits directly in __pycache__ with its source beside it.
   find -L <the folder you installed> -path '*/__pycache__/*.pyc' ! -path '*/__pycache__/*/*' -exec sh -c \
     'for f; do d=${f%/__pycache__/*}; b=${f##*/}; [ -f "$d/${b%%.*}.py" ] || echo "$f"; done' _ {} +

   Both print nothing when this commitment holds on your machine. (-L matters: installs are symlinked, and find without it silently skips them. Step 2 exists because a skill that ships a script leaves __pycache__ behind the first time it runs. The ! -path '*/__pycache__/*/*' on both steps is load-bearing: * spans slashes, so without it a file buried a level deeper inside __pycache__ is skipped by step 1 and waved through by step 2.)

   These two commands are generated from the same suffix list the CI walker enforces, and scripts/check-installed-skills.sh in this repository is the same pair in a runnable file.

4. No secrets handling. No skill asks the agent to read, move, or transmit credentials, tokens, or keys.

5. Explicit updates only. Installation copies files locally. Nothing self-updates; you diff and adopt changes deliberately.

6. Provenance. Skills with a real-incident origin carry a dated EVIDENCE.md. History is append-only in git.

Review the diff of any update as you would a pull request — that is the intended trust mechanism, not a substitute for it.

If any skill in this repo violates the commitments above, or you find a way a skill's instructions could be abused (prompt-injection amplification, instruction smuggling, destructive-command patterns), please open a GitHub issue — or, for anything sensitive, use GitHub's private vulnerability reporting on this repository.

Reports are acknowledged as fast as a one-maintainer project honestly can — typically within a few days. Confirmed violations get the skill pulled first and discussed second.