Side projects · Open source tools and experiments

Things I built and shipped

Two kinds of thing live here. Open source tools, where every claim is checkable at the source and the code is in the repo. And experiments, which are running products with real users, real money, or both.

Open source projects

Eight tools, one rule: every claim they make is checkable at the source. All MIT, all on GitHub, and the receipts are in the repos.

New · Skills pack · Grok Bot and Cursor

Wainwright

Hire Grok Bots like roles. Every new bot gets one job, a 14-check lint on its draft and a verified profile before it counts as hired.

  • Six steps, in order: ask only what can't be decided later, draft four fields (one job, anti-jobs, voice, wake), lint, create, read the new bot's profile.json to verify it, then, for a bot with a routine, hand off its standing shift and wait for the bot to confirm.
  • The designer bot lints its own draft against fourteen checks and is told to hold the hire on any fail. The target is the multi-job "helpful assistant", because the designer bot that creates it has no tool to delete it, and a human has to remove it from the sidebar.
  • Memory packed by tier: profile for who you are, log for dated rules, note for short-lived context. The skill tells bots never to store secrets, pasted contact lists or chat trivia.
  • I ran the trail on my own Grok Bot account on September 15, 2026, and the repo publishes that log next to the simulated tests that failed first. It is Markdown instructions that a model follows, it installs in Cursor as a local plugin, and the twelve Candor personas ship with it under new names.
About Wainwright Import the Manager GitHub
New · Agent skill · Any agent

Cutlist

Turns any long video into a clip plan, every title and description, and one thumbnail. Any coding agent, any recorder, any editor. Nothing published.

  • Give it the video and the captions from one export, from a podcast, a webinar, a lecture, a tutorial, a stream, or a keynote. It reads the whole transcript, scores every candidate on being self-contained, hooking in two seconds, and paying off, then confirms with audio energy and still frames. Five to eight clips with exact cut times, in publish order, each with one line on why.
  • Every field ready to paste: title, description, hashtags, tags, category, for the episode and each clip, validated against YouTube's documented limits. A number a guest says stays a quote. Every clip points to real words with a timestamp.
  • One thumbnail mockup that matches the first ten seconds, the title, and the first description line, at 1280x720 and at the 320-pixel size viewers see in the feed.
  • Open Agent Skills format: the same folder works in Claude Code, Codex, Cursor, Gemini CLI, and Copilot. Standard-library Python plus Pillow and ffmpeg. No API key, no server, no account. Thirteen self-test checks in the open repo.
About Cutlist GitHub
Live on the Chrome Web Store

Bangermeter

Score any tweet with X's published production ranking weights.

  • A free Chrome extension that badges every post in your X feed with what X's ranking formula makes of it. On August 13, 2026 X published the real production weights for the For You timeline, and Bangermeter shipped them the same day. Every factor is cited to the line of source code.
  • The engine is a port of X's own ranking_scorer.rs: the weighted sum across all 26 ranking heads, then the negative-rescale, author-diversity and out-of-network steps. The weights are X's. The probabilities they multiply are Bangermeter's estimates, and every screen labels which is which. It does not predict reach.
  • Zero network requests, zero data collected, no account. It all runs in your browser. 255 self-tests in the open repo, and every published weight is checked against X's param.rs individually, as of September 18, 2026.
The full story Add to Chrome GitHub
New · Runs in your browser

Inkwash

Strips invisible characters and machine typography out of finished prose. Nothing you paste leaves the page.

  • Finds zero-width characters, smuggled Unicode tag payloads, homoglyph letters borrowed from other alphabets, exotic whitespace and machine typography, then itemizes every change with its exact codepoint and the text around it.
  • A homoglyph is the interesting one. A Cyrillic с inside the word "ceramics" looks completely normal on screen, but you are then betting that every system reading that page folds it back to a Latin c before it matches anything. Some do. You have no way of knowing which.
  • The whole program is one file with no build step, no dependencies and no network request, so you can save it to your desktop and use it with the wifi off. That is the only way a privacy claim is checkable. It reads .docx too, and reports the author name, company and revision count riding along inside.
  • Measured rather than assumed: a scan of 2.4 million characters across every corpus built for the studies behind it found zero homoglyphs and 27 zero-width spaces, all inside one scraped court opinion. Cheap insurance, not a crisis.
Use it now GitHub
New · Python, standard library only

HUMANIST

Makes an AI-assisted draft measurably better to read. It does not make it undetectable, and the studies proving that are published with it.

  • A rewrite pass, a mechanical checker that reports and never edits, and a composition read for the five tells no regular expression catches. In a blind study where editors judged quality alone, with authorship never mentioned, the processed draft beat the raw draft 22 times out of 22.
  • It started as an attempt to beat AI detection and failed three separate ways: the pipeline moved detection from 100% to 97.8% against a design that could not resolve four points, an explicit "write more human" ruleset fooled 0 of 17 judges, and length changed nothing. All three negatives are published beside the one positive.
  • The checker's false-positive rate is published rather than hidden. Swept over 24 known-human documents from 1854 to 2018 it fails four of them. Its rules encode one person's taste in prose; they do not identify machines.
  • The rewrite pass is Siqi Chen's MIT-licensed humanizer, vendored. It carries the entire measured benefit at 16 of 16, and none of this would exist without it.
About HUMANIST GitHub
New · Python, local-first

LenkRaster

A toolkit for inspecting and preparing pixel art. It reports on your sprites and never draws for you.

  • It scores a PNG sprite against pixel-craft checks and hands back structured findings with the measurement attached, so you get a specific thing to go and check rather than a grade. Palette and contrast reporting, quantization, ordered dithering and OKLCH color ramps sit alongside it.
  • Animation-cycle QA measures frame-to-frame motion, timing and loop consistency using rendered-pixel math. On the four-frame hammer cycle in the examples, a fixed-palette pass took the artwork from 77,766 colors down to 48, with the original left untouched beside it for comparison.
  • Python API, command line, an optional Aseprite CLI bridge, and a local stdio MCP server for agent workflows. No network listener, no API key, create-only outputs, and every report is advisory rather than approval.
  • It does not produce finished artwork and it does not stand in for an artist. Art direction, provenance review and your own release gates all still apply, and none of those belong to a script.
The full story GitHub
Desktop app · Open source

Assay

Grades business websites for AI-readiness and builds evidence-based outreach. Local-first.

  • Six deterministic checks run against raw page source. No AI summaries of page content, ever: if it isn't in view-source, it isn't a finding.
  • Every claim in a deliverable can be reproduced by the recipient with Ctrl+U. That reproducibility is the product.
  • It never sends anything. There is no sender in the codebase; every outreach item leaves by your hand, one approved item at a time.
About Assay GitHub
Claude Code plugin · v0.5.0 experimental

Candor

Anti-sycophancy behavioral modes for Claude Code.

  • Twelve skills: one shared core plus eleven work-mode personas (coding, debugging, security, architecture, writing, data, and more) that cut reflexive agreement and hollow praise without cutting helpfulness.
  • Prompt-level instructions only, Markdown and JSON. No code execution; the whole plugin is text you can read.
  • The repo's own blind A/B benchmark has Candor responses preferred about 3 to 1 overall, with 39/39 trigger-routing accuracy. Its numbers, published to be re-run.
  • The oldest tool on this page, and in my opinion, the most important one: Candor's agents and skills ran the early sessions that put Krafty Bandit, my mom's shop, on track, and several of the skills in daily use here today are descendants of candor-core.
GitHub

Experiments

Running products rather than repos. One of them takes real money and one of them is a joke that ships weekly. Neither is client work, and neither is evidence about anybody's search visibility.

Live · Takes real money

Bid Club

A public paid leaderboard where rank is the bid and nothing else, and ten percent of it is pledged to the American Heart Association.

Ten percent to the American Heart Association. Every bid, boost and suppress is recorded each week and pledged to the AHA, in remembrance and honor of family. The running total sits on the board and the raw ledger is public at giveback.json. It is an operator pledge, not your tax receipt, and the AHA does not endorse it. The other ninety percent is what pays for the open source work above and keeps me working for myself.
  • Anyone with a website, a social handle, or an app store listing can pay for a seat on a public board. Rank is the bid, in whole dollars. The two at the top sit in a ring where anyone can pay to boost or bury either one. The fight is public and who paid is not.
  • A bid buys a seat on that board and nothing else. Not a backlink, not a citation, not search ranking, not app store ranking, not AI answer placement, and not traffic anywhere else. Rank there exists only there, and the site says exactly that on the page where your money would change hands.
  • You do not have to own what you list, and the paid boosts and suppressions are anonymous. Somebody can put your site on that board without asking you, and other people can pay to push it up or down without ever showing a name. That is the design. If you are on it and you did not put yourself there, mail hello@ryanlenk.com and it comes down.
  • I sell AI search visibility for a living and I also run this, so you should hear it from me rather than find it out. If you hire me, none of your work touches that board and none of your budget goes near it. They are two different things I run, and I will say so here if that ever changes.
  • It is advertising, not gambling and not a contest. A bid is the price of a sponsored seat. Sponsored listings are labeled as sponsored on the board itself.
The full story Bid Club The rules The 10% ledger
New · Episode one is up

Vibe Wheel

A weekly parody series. One 2005 pixel ad, rebuilt as a 2026 web app you can actually play with.

  • In 2005 a single web page sold a million pixels of advertising, almost all of it at a dollar a pixel. Vibe Wheel spins one of those ads every week and asks what it would be now, with a 2026 twist. Then it gets built. The joke has to run in a browser or it does not count.
  • Episode one is tBite, a photo status game where you put up a picture and everybody else votes PIMP or NAH. Hot rod red, chrome, too much shine. You can play as a guest, and signing in with Google is only there so your points stick.
  • Parody, and unaffiliated. No connection to the Million Dollar Homepage, to anyone who bought pixels on it, or to anyone who owns those brands today, and no endorsement from any of them. Each episode names the ad it came from as history and credit, and that is all.
  • Free to play, nothing to buy, no ads and no sponsorships anywhere in it. It is a side project I build on weekends, not a client deliverable, and it tells you nothing about anybody's search visibility.
The full story Open Vibe Wheel Play episode one

Everything on this page is an independent project by Ryan Lenk. The eight open source tools are MIT licensed and public on GitHub. Bid Club and Vibe Wheel are running experiments rather than repos. Not affiliated with, endorsed by, or sponsored by X Corp, xAI, Anthropic, Google, Cursor, Aseprite, the American Heart Association, the Million Dollar Homepage, or any advertiser who bought pixels on it.