Claude Code skills from Satisfying Systems

A small pack of Agent Skills for Claude Code. Each skill is a folder with a SKILL.md (the instructions Claude loads when the skill is relevant), plus the templates, prompts and scripts it points to.

They're generalised from workflows we use at Satisfying Systems, and most of the rules come from something that went wrong.

Install

Copy a skill folder into one of the places Claude Code looks for skills:

git clone <this repo> claude-skills

# For you, in every project:
mkdir -p ~/.claude/skills
cp -r claude-skills/skills/writer-reviewer-split ~/.claude/skills/

# Or for one project only (commit it so your team gets it too):
mkdir -p <your-repo>/.claude/skills
cp -r claude-skills/skills/humanize-without-changing-facts <your-repo>/.claude/skills/

Claude Code picks skills up by their description, so there's nothing else to configure. The scripts need bash and Python 3.8+ (standard library only).

Skills

writer-reviewer-split

One model writes the code in an isolated git worktree. A different model reviews the exact commit, read-only, in a fresh session. Nothing merges unless the reviewer's final message ends with the line VERDICT: PASS. The skill covers what goes in the coding brief, how to keep the reviewer independent, why execution and verdict are separate evidence, what to do on FAIL (quarantine, failing regression tests, a re-review that proves each test fails on the old commit and passes on the new one), how to recover a run that stopped part way, and when to stop patching and rethink the design. It includes fill-in brief and review templates, a pitfalls list, and extract_verdict.sh, which reads the reviewer's final message from the file the CLI writes it to (for example codex exec -o) or from Claude Code's JSON result, never from a raw transcript, ignores code blocks, accepts a verdict only when it's the last line, and refuses to report one when the run hit a usage limit or a capacity error.

humanize-without-changing-facts

An editor pass that makes AI-written prose read like a person wrote it, guarded by checks on the facts it mustn't change. The editor rewrites prose only. Then check_invariants.py compares the original and edited text and rejects the edit if the numbers inside any citation span changed or changed order, a citation marker moved to a different paragraph or list item, a double-quoted span was added, removed or changed, a link, reference definition, email, handle, hashtag or piece of code changed, the heading and paragraph structure shifted, a protected literal changed, or the length drifted out of band. Those checks don't know which subject a number or claim belongs to inside a sentence; that's the judge's job. A separate judge model has to reply with exactly {"changes": []}. Any failure keeps the original text. The skill includes the editor prompt, the judge prompt with a strict reply parser, and a tested checker.

Run the tests for both skills from the repo root:

python3 -m unittest discover -s skills/humanize-without-changing-facts/tests
python3 -m unittest discover -s skills/writer-reviewer-split/tests

Licence

MIT. See LICENSE.

About

Generalised from workflows we use at Satisfying Systems. More at https://satisfyingsystems.com/toolkit.