Personal, reusable agent skills for planning, browser and terminal automation.
Captures below for the quick concept.
| Skill | Capture |
|---|---|
| code-orientation Turn "I don't know this code yet" into numbers, a map, and a picture of the hot path. Main fiew is: SLOC breakdown, annotated file tree, and diagrams served locally. Point it at a repo, a subfolder, a dependency, or a PR range. Shown: redis 7.4.1 at 74b289a, 175,383 SLOC across 547 files, tracing one SET foo bar. |
 |
| presenting-html-plans Present a finished plan as a source-controlled local HTML page with Mermaid diagrams, content nav, and print styles. HTML viewer is quick to read and understand, but the plan stays MDX so it is also quick for the agent to read and understand. Shown: the plan for this gallery, rendered by the bundled scaffold. |
 |
| making-isometric-system-maps Draft: interactive isometric system maps. Select a component to light up its relationships, or trace a flow step by step with automatic camera moves Shown: a Kubernetes Deployment and ClusterIP Service, mid flow-trace. More captures. |
 |
| remote-browser-cdp-kvm Drive one persistent, logged-in Chrome over CDP through small repeatable verbs, and hand human-only steps (login, CAPTCHA, SSO, 2FA, sign-off) to a private phone-friendly KVM over your tailnet. Shown: agent shares "I'm not a robot" screen to the human, and tailnet plus Android split screen makes it easy to interact with the agent and browser both. Real recording, obviously sped up. |
 |
| tui-puppeteering-with-tmux Isolated tmux sessions for automating and testing TUI and CLI applications. tui-start brings a program up on a dedicated socket, tui-send delivers keystrokes, tui-capture dumps the screen as text, and tui-assert refuses to continue until the expected state is actually on screen (the wait for action needed to drive the interactions). Nothing it does can reach your own tmux.Shown: Nori CLI at launch, and then same CLI with its session config picker open. Started and driven by this another agent. |
  |
| tui-capture-with-ghostty-web Capture a tmux-driven TUI as paired plain-text and Ghostty Web PNG artifacts, for visual QA, documentation, and screenshot regression evidence. Every capture writes screen.txt, screen.ansi, screen.png, and a metadata.json carrying the grid, the pinned renderer version, and a PNG checksum.Shown: bat rendering this repo's own sloc.py, captured at 120×40 by ghostty-web@0.4.0. The paired text is committed beside it. |
 |
Each skill's code and instructions live under its own directory.
Original work is available under the MIT License. The one credited preceding concept, and the third-party dependencies that are not vendored here, are documented in NOTICE.
The NetHack harness, plus two advisory skills that ride alongside it. The two advisory skills have no visual output, so they are listed but not illustrated.
| Skill | Capture |
|---|---|
| nethack Compact, observed NetHack control through the bundled Rust nh harness. Movement, combat, prompts, inventory, and run-state reporting against a live game in an isolated tmux session. The harness captures before and after every key, classifies the event, and refuses input it cannot interpret rather than guessing.Shown: a real game at turn 31, Claude the Stripling, a lawful human Valkyrie on Dlvl 1, two rooms and the connecting corridor explored, a lichen ( F) east and an unidentified potion (!) southwest. |
 |
| nethack-strategy Live strategy guidance: useful exploration, early retreat, timely item use, limited pet rescue, and clear reasons to leave a bad fight. |
— |
| nethack-wiki-research Bounded local-wiki research delegated to a sub-agent, so raw source material never enters the playing agent's context. |
— |
Large local NetHackWiki and official-source archives are runtime data and intentionally are not tracked.