Agent Skills Bundle

Custom portable skills plus dependency management for skills maintained in other repositories, behind one minimal installer powered by npx skills.

Each local skill directory contains a SKILL.md with operating instructions and optional helper scripts. Skills are agent-agnostic where possible; tool-specific skills still assume the named CLI exists on PATH.

Skills

• dev-loop-build — autonomously implement an agreed vertical slice with TDD, risk-adaptive adversarial review, and an atomic local commit.
• dev-loop-design — shape unclear/larger work with pragmatic grilling, bounded research, competing designs, and adversarial review.
• dev-loop-pair — implement an agreed slice through a tight staged-review loop with the user.
• deep-research-pro — run ChatGPT Deep Research through Oracle using Adi's logged-in ChatGPT Pro profile.
• diagram — render Mermaid diagrams to high-resolution PNGs.
• github-pr-review-comments — read and manage GitHub PR review threads with gh + GraphQL.
• grill-me — stress-test a plan or design through one-question-at-a-time grilling.
• grill-socratic — interview without hints, options, examples, or leading answers.
• grill-with-docs — grill against project language and propose CONTEXT.md/ADR updates with approval gates.
• hunk-review — interact with live Hunk diff review sessions via CLI.
• ios-device-runner — build, install, and launch an iOS app on a physical device; configure with CLI flags, env vars, or a private --config file.
• phoenix-grilling — extract success criteria and unresolved decisions through proportionate, non-leading, Phoenix-aware questioning.
• project-catchup — quickly bootstrap context for active repo work.
• repo-audit — copy/clone a repo into /tmp and run a critical audit.
• security-privacy-review — audit a PR, diff, or codebase for security and privacy risks.
• ste-writing — write in ASD-STE100 Simplified Technical English, with a bundled linter and a dictionary lookup for the standard.
• tmux-agents — orchestrate parallel sub-agents in tmux/Supacode panes.

The installer also installs these directly from their upstream repositories:

• Agent Browser
• Herdr
• tuicr
• Phoenix Architecture
• Thermos
• Torvalds Doctrine
• Matt Pocock's skills

They are not vendored, and their full catalogs are not snapshotted here. The installer stores only source URLs and curated default names, reads current skill names from npx skills, and installs content directly from upstream; plugin hooks and tools stay out of scope.

Layout

<skill>/
  SKILL.md
  scripts/            # optional helper scripts
  state/              # optional local state, ignored by git
bin/install.mjs       # live unified installer
skill-overrides.json  # preferred source for same-name skills

Install

Run the interactive installer directly from GitHub:

npx github:adstastic/agent-skills

The installer loads every current upstream catalog, lets you search and select skills from each source, then asks project/global scope, coding agents, and symlink/copy method once. It validates each parsed catalog count before showing choices and aborts rather than silently presenting a partial list. --yes installs the curated defaults without prompts.

Overrides

skill-overrides.json records intentional same-name forks as skill name → preferred source. When both versions are selected, the preferred source wins; selecting only the non-preferred source still installs it. Unmapped collisions remain an interactive choice.

The local grill-me and grill-with-docs are standalone workflows, while Matt Pocock's current versions delegate to his grilling and domain-modeling skills, so both are explicitly overridden to adstastic/agent-skills.

For the non-interactive personal preset, clone the repo and run:

./sync-skills

This installs every local skill plus the curated upstream subset globally for Claude Code and Codex. Pi discovers the same canonical ~/.agents/skills installation. Run it again to fetch upstream changes or restore the bundle on another machine.

For a one-off Pi session without installing globally:

pi --skill /path/to/agent-skills

Helper scripts use relative paths in docs. If your agent runs from another directory, set a SKILLS_DIR env var or replace ./<skill>/scripts/... with an absolute path in your local copy.