IMPLICATOR
Five projects climbed GitHub this week and none of them is a model. Each sits in the layer a team can own outright: the rules an agent reads before it types, the diagrams it hands back, the workspace it runs in, the harness keeping a half-dozen from colliding, and the weights underneath.
go-modern-guidelines
JetBrains ships its GoLand team's idiom rules as an agent skill instead of IDE documentation. The plugin reads the Go version out of go.mod, then tells the agent which language features that version permits, so it writes slices.Contains rather than a hand-rolled loop and errors.AsType[T] where Go 1.26 allows. Installs into Junie, Claude Code, Codex and Cursor.
Archify
The agent writes typed JSON, not a picture. Archify validates that intermediate form against a schema, then compiles it deterministically into one self-contained HTML file covering five diagram types, dark and light themes, and PNG, SVG and WebM export. It also diffs two validated snapshots as before, delta and after, so an architecture change can be read at review time.
FREE WEEKDAY MORNING BRIEFING
Don’t miss the next AI story that matters.
The Implicator Morning Briefing filters the AI news cycle to the stories worth your attention and explains their consequences. From San Francisco, every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern.
About five minutes. No hype. No spam.
Worktrunk
A Rust CLI that addresses git worktrees by branch name, so wt switch -c -x claude feat replaces the usual sequence of git worktree add, then cd, then starting the agent. Adds hooks on create and merge, a unique dev-server port per tree, build-cache copying between trees, and a list view carrying CI status per branch.
wt binary collides with Windows Terminal, so the Winget package installs it as git-wt and every documented command has to be retyped.Munder Difflin
An Electron app that wraps the terminal agent CLIs a developer already pays for, Claude Code, Codex, Grok, Gemini CLI and nine others, as real processes in a pseudo-terminal, then gives each one a mailbox, a markdown memory file and a desk on a 2D office floor. A supervisor agent routes work between them and escalates only when it has to.
Heretic
One command strips safety alignment out of an open-weights model. Heretic runs directional ablation steered by an Optuna parameter search that co-minimizes refusals and KL divergence from the original, so the modified model keeps more of its ability than hand-tuned attempts do. On Gemma 3 12B it took refusals from 97 of 100 prompts down to 3.
Heretic
Heretic earns the slot for what it measures, not for what it removes. Refusals on Gemma 3 12B fall from 97 of 100 prompts to 3, and the modified model sits 0.16 KL divergence from the original on harmless prompts, roughly a third of the next-best hand-tuned abliteration. Those two numbers describe how thin the safety layer is: a direction in the residual stream, located by a parameter search, gone in twenty to thirty minutes on an RTX 3090. More than 5,000 models on Hugging Face now carry the tag.
Treat it as an instrument rather than a product. Take an open-weights model your company already ships or plans to ship, run Heretic against it on an isolated machine, and write down the two figures: how far the refusal rate drops, and how much capability survives the drop. That pair belongs next to the model card in the risk register, because it describes what an outside party can do to those weights the day they go public. The tool's author, Philipp Emanuel Weidmann, also ships the interpretability half: pass --plot-residuals and it charts how harmful and harmless prompts separate layer by layer, which is the same picture a safety team would want anyway.
Frequently Asked Questions
How were these projects selected?
Current GitHub metadata, recent activity, README clarity, practical setup path, and relevance to builders working with AI systems.
Are stars enough?
No. Stars measure attention. Push dates, license, issues, docs, and whether the project solves a specific workflow decide usefulness.
What does the difficulty score mean?
It estimates how hard the project is to test or adapt, not how impressive the underlying engineering is.
Which repo should readers try first?
go-modern-guidelines is the cheapest test at 1/5, and Archify is the fastest way to see output. Heretic is the more strategic experiment.
What should teams check before production use?
License, data retention, credential access, update speed, maintainer responsiveness, and whether the repo has a realistic rollback path.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
