Hermes gains its project scope
Hermes does read a project's skills, so the installer now writes <project>/.hermes/skills/ and the AGENTS.md pointer. It used to force every Hermes install into ~/.hermes/, which meant the one tier that outranks the global one was unreachable.
Hermes will not load skills out of a cloned repository until you say that repository is yours. That is a deliberate vendor gate, so the installer names the step instead of bypassing it. Run this once in the project after installing:
hermes skills trustProject skills are a newer Hermes feature: the official installer has it, the older hermes-agent package on PyPI does not.
The install documentation is rewritten end to end
Both routes now run from opening a terminal to a verified install. Every route documents updating and removing, and the phone and chat-window route is written down for the first time.
The guide is also restructured around what you are doing instead of which tool you picked. Install, Update, and Remove are separate sections that each cover the same routes in the same order, so finding your route in one finds it in the others.
Cursor's install command is corrected
/add-plugin <url> appears in no Cursor documentation and not in the shipped CLI bundle. The documented route is agent plugin marketplace add <url>, then Install in the Customize panel.
The installer's conflict prompt is clearer
It now says which choice updates and which one quietly leaves the old version in place, and its closing line says so when it keeps your existing folders. Previously you could pick the safer-sounding option, install nothing, and stay on the old version with nothing telling you.
Two content fixes from a live generate-test
Eyebrow Badge Above the Headline joins the pattern list: a pill parked directly above the H1, usually with a dot, holding a category label the headline already says. It had slipped between R-09, whose forbidden content list stops at "AI Powered", "Beta", and "New", and Decorative Status Dot, which opens with a glowing, pulsing dot.
The core also now states outright that three steps run in both usage modes whatever else is driving the session: direction resolution (R-37), asset clarification (R-23), and the Delivery Gate. The test that surfaced this ran another skill's planning workflow, which led the work while antislop's mandatory steps went unrun, including the gate report.
The README gains a comparison section
The core also now states outright that three steps run in both usage modes whatever else is driving the session: direction resolution (R-37), asset clarification (R-23), and the Delivery Gate. The test that surfaced this ran another skill's planning workflow, which led the work while antislop's mandatory steps went unrun, including the gate report.
The README gains a comparison section
One brief, generated four times over. Three page builds sit in a row, each captioned underneath: with nothing, with antislop alone, with DESIGN.md alone. The fourth follows at full width, both together, the only one of the four that is clean and directed at the same time.
They are separate outcomes rather than a before and after, because each is a different set of tools rather than a stage of the same work. Copy and code comments get before-and-after pairs instead, since there the filter is the whole of the job.
Updating
- The installer: run
npx antislop-aiagain and choose Overwrite them. - A plugin door: update through your agent's own plugin command.
- The single file: download
antislop.mdagain.
Every earlier release is in ROADMAP.md.