| FazBrowse GitHub Viewer | Trending | | Home |
| Tools: [Download Repo ZIP] [Original HTTPS Page] |
| Name | Name | Last commit date | ||
|---|---|---|---|---|
A collection of AI agent skills we're collecting as we find gaps and needs in our day-to-day work in applied AI. These address real friction points we've encountered and will grow over time.
Install one skill into the current project:
npx skills add gpu-cli/skills --skill decision-trail
npx skills add gpu-cli/skills --skill lyrebirdOr install the whole collection:
npx skills add gpu-cli/skillsCreates brand-aware content strategy and platform-native posts for Blog, LinkedIn, Reddit, and X. Lyrebird establishes a durable VOICE.md file, researches current platform guidance and factual claims, and writes social output under social/<proposal-slug>/.
Invoke: /lyrebird
How it works:
Link tracking: When VOICE.md defines a UTM convention, write and modify tag backlinks to your own domains with utm_* parameters (utm_source set per platform, so a LinkedIn→X conversion updates the source). Third-party and citation links are never tagged.
Path: skills/lyrebird
Outputs: VOICE.md for reusable voice context and social/<proposal-slug>/ for generated post files. X output is always named x.md.
Analyzes staged git changes and evaluates agentic context files to suggest additions or removals. Keeps your AI context (CLAUDE.md, .cursorrules, AGENTS.md, etc.) synchronized with actual code.
Invoke: /context-curation after staging changes
How it works:
Ensure context evolves with your codebase by adding a precommit hook.
Add to .claude/settings.json:
{
"hooks": {
"PreToolUse": [
{
"matcher": "Bash",
"hooks": [
{
"type": "command",
"command": ".claude/hooks/precommit-context.sh"
}
]
}
]
}
}Create .claude/hooks/precommit-context.sh:
#!/bin/bash
input=$(cat)
cmd=$(echo "$input" | jq -r '.tool_input.command // ""')
if [[ "$cmd" =~ ^git[[:space:]]+commit ]]; then
echo "Run /context-curation first to check if context files need updates." >&2
fi
exit 0Make it executable: chmod +x .claude/hooks/precommit-context.sh
Conducts time-boxed technical investigations with structured, citation-backed output. Produces markdown reports suitable for sharing with stakeholders or preserving decision context.
Invoke: /research-spike
How it works:
Asks upfront questions to scope the work:
Follows a structured methodology:
Outputs a structured report to research/[topic-slug]/spike-report.md including:
Applies anti-hallucination protocols: only states what sources confirm, uses hedging for inferences, marks opinions vs facts.
Test CLI applications interactively using tmux sessions. Provides a helper library for launching commands, waiting on conditions (never sleeping), sending keypresses, capturing output, and asserting on content. Works with any CLI or TUI — GPU CLI patterns included. Also supports testing inside Docker containers.
Invoke: /tmux-cli-test
How it works:
Sources tmux_helpers.sh for a full set of primitives:
Test pattern:
Start session → wait for ready signal → interact → assert → kill
Docker support via tmux_docker_helpers.sh — routes all tmux commands through docker exec for container-based testing.
Requires: tmux
Explore and analyze TUI applications to produce clone-ready documentation. Launches the target TUI in a tmux session, systematically explores all views and keybindings, captures ASCII diagrams of each screen, and writes findings incrementally to a timestamped markdown file (survives context compaction).
Invoke: /tui-clone
How it works:
Launches target TUI in a sized tmux session using tmux_helpers.sh
Immediately writes each discovered screen to the output file as it's found — context compaction safe
Captures across 12 dimensions:
Outputs tui-analysis/[app-name]-[timestamp].md — a complete clone spec
Supports: Claude Code, OpenCode, Codex, lazygit, lazydocker, htop, btop, k9s, and any ratatui/ncurses app.
Critically review terminal user interfaces across 10 UX dimensions. Launches the TUI in tmux, takes color screenshots at key states using freeze, visually inspects each one, and produces a graded report. Benchmarks against Claude Code, OpenCode, and Codex — the three best-in-class AI terminal UIs.
Invoke: /tui-review
How it works:
Requires: tmux, freeze (brew install charmbracelet/tap/freeze)
Security audit and active remediation for agent skills. Analyzes SKILL.md instructions and bundled scripts for prompt injection, data exfiltration, excessive permissions, supply chain risks, and other threats aligned with OWASP Top 10 for LLMs (2025), NIST AI RMF, SLSA, and OpenSSF Scorecard best practices. Presents findings inline first, then optionally generates structured reports (Markdown + JSON) and can rewrite skills to remove security concerns.
Invoke: /skill-shield <path>, /skill-shield <path> --remediate, or /skill-shield --all
How it works:
Requires: sha256sum (or shasum), git
Keeps a reviewable decision trail for a branch worked by several agents and the user at once: what was decided, why, with what perceived confidence, on what evidence, and by whom. Adapted from Cursor's show-me-your-work skill into one beads-native, multi-agent, PR-oriented skill with three subcommands.
Invoke:
How it works:
Requires: git; jq and bd (beads) for the default backend; gh for --pr. Degrades to a TSV trail without bd.
Test: bash skills/decision-trail/tests/selftest.sh
Deletes agent commentary from code and rewrites the comments worth keeping as one plain line. Agents restate the code, narrate their own edits, carry issue IDs no reader can resolve, and spend three lines where none were needed. This removes that layer without touching a single line of executable code.
Invoke: /clean-comments [path], /clean-comments all, /clean-comments check, /clean-comments install
How it works:
Prevention: /clean-comments install writes a fifteen-line comment policy into CLAUDE.md, AGENTS.md, .cursorrules, and friends, fenced by markers so reruns update in place. Preventing the comments beats scrubbing them, so this is the part worth running first.
Checking: /clean-comments check reports without editing, for CI or PR review. --ci fails only on the four mechanically decidable rules. There is deliberately no pre-commit rewrite hook: an LLM pass at commit time is slow, non-deterministic, and rewrites code the author has already reviewed.
Requires: git, node
Test: bash skills/clean-comments/tests/selftest.sh
Re-renders the previous response at a length you choose — one sentence, two, or a paragraph — without redoing the work that produced it. For the moments when the answer was right and three times longer than you needed.
Invoke: /less 1, /less 2, /less paragraph, /less 2 set, /less unset
How it works:
The guarantee: compression reads the previous response and nothing else. No tool calls, no re-reading files, no fresh reasoning, and no claim that was not already made — which is what separates it from asking "shorter please", where the agent frequently re-derives the whole answer and the preference evaporates next turn.
What survives: ranked, in this order — errors and failures; warnings about destructive actions and risks you are walking toward; what you must do next, including a decision only you can make; and the concrete values those depend on — numbers, paths, commands, URLs. Everything is folded into the sentences available rather than given its own, and when the first three genuinely cannot fit, they may bend the cap by a single sentence — concrete values never do. The full response is still on screen, so whatever cannot fit is pointed at ("the four steps above") rather than dropped silently.
Under a sticky cap, the cap governs the response, never the work. A capped turn still gets the full investigation; only the report is short.
Limitations: a sticky cap lives in conversation context, so it can fade when the context is compacted and does not survive into a new session — re-issue /less N set if it stops applying. That is the deliberate trade: the durable alternative is a hook and a state file, which would tie the skill to one agent.
Path: skills/less
Test: bash skills/less/tests/selftest.sh
Skills are billed in two places, and the difference drives how they are written. A frontmatter description is injected into every session whether the skill runs or not; a SKILL.md body loads once per invocation; everything under references/ loads only when a phase asks for it. So descriptions are routing triggers, bodies are orchestration outlines, and detail lives in references.
Budgets are 45 tokens for a description and 2,000 for a body, enforced by:
bash scripts/skill-lint.sh # per-skill description and body token estimates
bash scripts/selftest.sh # the lint plus every skill's own self-testThe lint reproduces Claude Code's own arithmetic, including the two parts that are easy to get wrong: the skill name is billed alongside the description, and current models estimate at 3 characters per token rather than 4. Its SHOWN column is the figure /skills displays, so the two can be compared directly. See AGENTS.md for the formula.
The lint also fails if generated output (reports, analyses, audit artifacts) is tracked inside a skill directory — those ship to every install and an agent reading the package can mistake them for instructions.
See AGENTS.md for the full convention.
| Back | FazBrowse Home | New Git URL |