Telling a model "sound human, don't be cringe" doesn't reliably work — the tells leak right back in on the next draft. So instead of asking, this lints: a deterministic checker for the patterns that scream "an AI wrote this."
Two composable layers:
check_text.py — the base. Catches the universal AI tells: puffery
("boasts a", "rich tapestry"), importance-emphasis ("plays a vital role"),A small shell guard that blocks any command from printing the raw contents of a secret-store
file (fnox.toml or any *.env) to stdout. Agent transcripts and shell histories persist, so
even an age-encrypted fnox.toml should never be echoed.
It catches the file-read vector: cat/grep/sed/awk/head/tail/less/xxd/..., language
one-liners like open(".env"), input redirection < .env, and git diff/show/blame/log of
the file.
A tiny script that turns a markdown or text file into an mp3 using Kokoro-82M, a small open-weight TTS model. Runs on CPU, no GPU required.
I built this so I could listen to long articles I'd saved without paying for a TTS subscription. Sharing the core because the rest of my pipeline (Nextcloud upload, fuzzy filename matching, an LLM skill to invoke it) is wired into my own setup — but the part that does the actual work is small and reusable.
pip install kokoro-onnx soundfile numpyA small Telegram bot that bridges forum topics to Claude Code sessions. Each forum topic in a Telegram group maps to one Claude Code session (a UUID), stored in sqlite. Photo uploads work — Telegram downloads the image, the bot drops the path into the prompt, and Claude's Read tool loads it.
Why this exists: I had a more complicated version of this that glued each Telegram topic to a live tmux window running an interactive Claude session. It broke constantly — three layers of state, all with independent lifetimes. Replaced all of it with one sqlite row per topic. If a session is lost, the next message mints a new one. No processes to babysit.
If you use both Claude Code (terminal, desktop) and claude.ai (web, mobile), you have two Claudes with two separate memory systems. Whatever you teach one, the other doesn't know. Over months the claude.ai side drifts.
This skill fixes that by using a Slack channel as a human-gated bridge. Claude Code drafts a structured memory digest. You approve. It posts. Your claude.ai Claude reads the channel at session start and proposes memory edits on that side. You approve there too.
Two files in this gist:
brain-sync.md — the Claude Code skill. Drop it in .claude/skills/ (project-scoped) or ~/.claude/skills/ (global).
You are running a weekly skill health review.
Read skill-usage.jsonl in the repo root. Each line is JSON: {"skill":"name","ts":"ISO8601","cwd":"/path"}. If the file doesn't exist or is empty, note that and skip to Step 3.