Skip to content
Greg Lum

Learn

Agent Skills

Reusable instructions for tasks you want an agent to handle consistently. Each skill is a folder with a SKILL.md file for Claude Code or Codex. They cover design and layout, motion, films, reviews, writing, specs, and tests.

48 skills from greglum/agent-toolkit at dd691be, synced October 28, 2026.

View on GitHub

Start with a Task

Most jobs need one skill, sometimes two or three in order. Skip a step when the task does not need it.

  • Turn a conversation into buildable work

    1. to-spec

    A spec with scope, decisions, and conditions you can check when it is done.

  • Understand and review a change

    1. guided-review
    2. global-code-review

    A reading order tied to the actual diff, then findings checked against your coding standards.

  • Build or check a web page

    1. web-app-design
    2. fit-surface
    3. deslop-ui

    Layout, focus, and accessibility measured in the running app, not guessed from the code.

  • Add interface motion

    1. find-animation-opportunities
    2. animate
    3. review-animations

    Motion with a reason to exist, tested for interruption and reduced motion.

  • Pick up and watch a pull request

    1. continue-pr
    2. monitor-pr

    The original intent recovered, then CI and review comments followed until it is ready or blocked.

  • Write from sources

    1. extract-documents
    2. deslop

    Documents turned into text with a record of each source and its verified locations, then a plainer draft.

Install

A skill asks the agent to run commands in your shell. Read a skill before you install it, and read the dry run before the installer changes anything. Python 3.10 or later; macOS or Linux.

  1. Step 1: Clone the repository

    git clone https://github.com/greglum/agent-toolkit.git
cd agent-toolkit

  2. Step 2: Read what the installer would change

    python3 scripts/install.py --dry-run --only ui-motion

  3. Step 3: Install

    python3 scripts/install.py --only ui-motion

Leave out --only to install every skill. The installer links skills into each harness, never deletes a folder, and does not change settings, permissions, or Git hooks.

Catalog

Automatic skills are picked by the agent from their description. By-name skills run only when called. Sources name the public work a skill draws on.

  • acp

    By name, Codex only

    Review task-owned changes, stage and commit them in cohesive groups, then push to the authorized Git destination. Use for add/commit/push requests; preserve unrelated staged and working-tree changes.

  • Analyze a repository for reliable agent work, make an evidence-backed improvement plan, and implement it with bounded parallel workers. Use to prepare or optimize a codebase for agents. Assessment-only requests produce a plan without repository edits.

  • animate

    Automatic

    Build web UI motion that earns its place: gate purpose and frequency, pick the cheapest tool, then set properties, timing, interruption, exit, and reduced motion from project tokens. Edits code. Use to animate an element, add a transition, or build a find-animation-opportunities row. Grade results with review-animations.

  • app-verifier

    Automatic

    Set up a way for agents to run and check this app: create or update a project-owned skill or doc that records how to launch it, reach each feature without real credentials, and see the end state proving it works. Use to record how to verify features, or refresh that record after code changes. Not for auditing or fixing UI.

  • Assess codebase quality with source-grounded measures and a frozen local rubric. Use for a quality baseline, scored improvement workshop, or checking a requested quality target across eight dimensions.

  • Plan and review source-backed marketing campaigns for owned projects across a website, YouTube and LinkedIn. Build channel briefs, campaign links and matched-window performance reports from local exports. Use for promoting work or choosing the next content experiment; not posting or film production.

    Sources

  • continue-pr

    Automatic

    Resume a PR, branch or agent handoff, or reassess an old PR before rebasing. Recover intent, preserve contributors and local work, and check whether the goal is still useful against the current architecture.

    Sources

  • create-pr

    By name, Codex only

    Create a pull request from committed branch changes after verifying the head repository, push destination, and intended base. Use for PR creation; preserve uncommitted work and report an existing matching PR instead of creating a duplicate.

    Sources

  • deslop

    Automatic

    Remove AI tells from prose in docs, READMEs, and web UI copy (labels, messages, i18n strings). Run the tells and Vale scripts, rewrite plainer, then check. Audit mode reports and changes nothing. Use to de-slop, humanize, plain-language, or de-AI text. Not for simplifying code; use simplify-code.

  • deslop-ui

    Automatic

    Audit or fix a web UI that looks AI-generated: repeated visual, structural, and copy patterns. Use to deslop, de-AI, modernize, sharpen, or make a UI pop. Keeps the repository's design system; an audit changes nothing. For new visual direction use web-app-design; for width or overflow fit use fit-surface.

  • Create source-grounded system diagrams and standalone interactive node maps with search, upstream/downstream tracing, route inspection, before/after comparison, and PNG/SVG export. Use Mermaid for small Markdown diagrams, existing FigJam tools for editable boards, and plotting tools for statistical charts.

  • Turn content the user owns (a post, doc, pull request, release note, or page) into LinkedIn and X drafts. Each draft is traced to the source, checked against a used-source ledger, and held for the user to accept, edit, revise, or drop. Drafts only: never posts, schedules, or signs in. Invoke explicitly with /distribution-draft SOURCE or $distribution-draft SOURCE. For prose cleanup alone, use deslop.

    Sources

  • docs-sync

    By name

    Audit or update project documentation against current code and commands. Use when a change leaves README, AGENTS, or linked docs stale. Audit, review, check, and dry-run requests report findings without edits; sync, fix, and update requests apply scoped corrections.

  • Prevent repeated agent mistakes by finding their source in history or user corrections, then changing the owning architecture, types, or checks. Use for recurring review findings, correction sweeps, or requests inspired by pstack /correct. A single incident needs evidence of a recurring rule.

    Sources

  • Extract a mixed document corpus into text or Markdown with a per-source manifest of provenance, verified locators, and content loss. Use for preparing sources for comparison or search, not document authoring, browser automation, or vector ingestion.

    Sources

  • Find where a web interface should gain motion and where it should not. Read-only sweep that returns a gated, ranked ledger with file:line evidence and rejected candidates. Use for "what should animate here?" or "make this feel alive". To build motion use animate; to grade existing motion use review-animations.

  • Find other instances of an already confirmed defect within a bounded codebase. Use when asked whether the same bug occurs elsewhere or to audit sibling paths after diagnosis. Requires a known seed and cause; not a general security review or initial bug diagnosis.

    Sources

  • find-docs

    Automatic

    Retrieve current authoritative documentation when an answer depends on uncertain, version-specific, or exact API, configuration, migration, CLI, or library behavior. Do not use for general concepts or facts already established by the current repository and task.

  • fit-surface

    Automatic

    Make one existing web surface fit every supported viewport width, chrome state, theme, and data state, measured in the running app. Use for overflow, worst-case content, or controls hidden by mobile chrome or keyboards. For new design direction use web-app-design; for generic AI-looking UI, deslop-ui; for motion, ui-motion.

    Sources

  • Review a diff against a coding standards file the user names plus any standards the repo documents. Use when asked for a global code review, a review against their own standards, an over-engineering or unnecessary-complexity check, or a standards review in a repo that documents none.

  • Use repository graphs to plan and verify a change: trace dependencies, test impact, and missing relationships before choosing what to rebuild. Use for cross-file change impact and architecture questions; graph visualization belongs to diagram-design.

    Sources

  • guided-review

    Automatic

    Create a human reading guide for a pull request or significant diff. Use for guided reviews, code tours, or help understanding large agent-generated changes. Explain behavior before and after, order the review by dependencies and risk, and account for the complete diff. Complements defect review and does not issue an approval verdict.

    Sources

  • Audit, revise, and evidence-test one existing Agent Skill or the declared portfolio across Codex, Claude Code, GitHub Copilot, and compatible clients. Invoke explicitly with $harden-skill PATH, /harden-skill PATH, or /harden-skill portfolio. Do not use to create a new skill from scratch, or to research outside sources for the catalog (use skill-research).

  • Design or run a bounded engineering feedback loop with a fixed objective, observable acceptance, independent checks, and stop conditions. Use for repeated improve-review-verify work; skill-specific evaluation belongs to harden-skill.

    Sources

  • monitor-pr

    Automatic

    Monitor an existing GitHub pull request for CI, review, thread, head and base changes. Use when asked to watch or babysit a PR; verify findings and report readiness separately from event notifications.

    Sources

  • motion-film

    Automatic

    Create, edit, audit, or render films: product demos, explainers, data graphics, motion graphics, music-led typography, and character animation. Reuse the project's renderer and verify the requested output. For application interaction motion, use ui-motion instead.

    40 sources for motion-film

  • Run the whole animation loop on a web app with parallel subagents: find opportunities, build each with animate, then review and fix in bounded rounds until no Block findings remain, and verify in the running app. Explicit only. Edits code in the current workspace; never commits. Arguments: an optional scope path, max=N rows, rounds=N, or plan-only.

  • Diagnose PostgreSQL query slowness, estimate errors, lock or connection bottlenecks, and row-security surprises. Produce evidence and a bounded remedy; use the project's migration workflow for schema or policy changes.

    Sources

  • Trace a disputed dashboard number or aggregate to its definition and source records, then reconcile it independently. Use for conflicting totals, join inflation, rate denominators, weighted averages, or time-window mismatches; not general chart styling or database tuning.

    Sources

  • Assess Git worktrees against an identified base and refresh authorized clean worktrees without losing local work. Use after merges or when asked to refresh worktrees. Preserve dirty, active, or unmerged work; removal, recreation, and branch reset require authority for the exact targets.

  • Audit remote branches, local branches, worktrees and stashes; recover useful unfinished work before removing verified completed resources. Use for repository cleanup and stale-work sweeps, rather than a routine worktree refresh.

  • Cut short vertical clips from a long video the user owns or has licensed: word-level transcript, scored candidates snapped to sentences, overlap suppression, context checks, subject-tracked reframing, and a per-clip edit list. Uses existing local tools. Film and export rules: motion-film. Posting copy: distribution-draft. Campaigns: content-marketing.

    Sources

  • Review web animation and motion code, in a diff or a named surface, against a strict craft bar: purpose, frequency, timing, origin, interruption, cost, and reduced motion. Read-only; returns file:line findings with a Block or Approve verdict. Build new motion with animate; diagnose stutter with ui-motion.

  • Review client-API integration for wire-contract mismatches, transformation costs, validation, and error handling. Audit a requested feature or integration scope; keep findings read-only unless fixes are authorized.

  • review-pending-changes

    By name, Codex only

    Assess all requested uncommitted changes, including staged, unstaged, untracked, deleted, and renamed paths, for defects and over-engineering before committing. Read-only assessment with APPROVE/REVISE/REJECT verdict.

  • review-pr

    By name, Codex only

    Review a pull request for code quality, bugs, over-engineering, and guideline compliance. Analyzes the full PR diff with confidence-scored findings and APPROVE/REVISE/REJECT verdict. Read-only assessment.

  • review-recent-commits

    By name, Codex only

    Review recent or named Git commits against repository rules and current file state. Use for an evidence-based, read-only APPROVE/REVISE/REVERT verdict on committed work. Use review-pending-changes for uncommitted work and review-pr for a PR or branch diff.

  • Review named source files, a feature, or a codebase for reachable defects, contract failures, and maintenance problems with concrete impact, including unchanged code. Read-only source audit; not a PR verdict or commit-history assessment. Discover the stack and apply only relevant runtime checks.

  • Audit a repository's security boundaries and verify suspected vulnerabilities; remediate confirmed findings when requested. Use for whole-system security reviews and audit follow-up, rather than routine diff review or a search for variants of one known bug.

    Sources

  • simplify-code

    Automatic

    Simplify code in named files or directories, including unchanged code, while preserving behavior and project conventions. Use for recursive, filtered, shallow, or dry-run simplification. For an explicit comment-only request, audit or clean comments without changing code behavior. For prose use deslop.

  • Vet linked articles, threads, videos, or repositories with independent web research, identify functional gaps, and build supported improvements into custom skills and tooling, or run a catalog-wide research round. Invoke /skill-research or $skill-research with URLs, a focus, skills, or a date window. To audit and evidence-test one skill, use harden-skill instead.

  • Research public posts and comment threads from other accounts, peers or a niche to find what works and what audiences ask. Use for competitor outliers, topic or comment listening, and audience questions; writes a dated evidence record. For owned campaign measurement use content-marketing; for drafts, distribution-draft.

    Sources

  • Create a source-grounded interactive anatomy, cutaway, or isometric explainer of a system or mechanism. Use when parts, layers, and user-controlled state changes teach how it works. For ordinary architecture graphs use diagram-design; for timeline videos use motion-film.

    Sources

  • Create or revise YouTube thumbnails, LinkedIn video covers, and social post images from a film brief, script, or existing video. Research concepts, use cleared photos or generated artwork, and deliver checked platform exports. Use motion-film for video production.

  • to-spec

    Automatic

    Synthesize the current conversation into an implementation spec without another interview. Publish to the identified issue tracker when the user authorizes it, or deliver a complete local draft.

    Sources

  • tune-tests

    By name

    Audit, prune, or speed up an existing test suite without losing failure detection. Audit mode scores every in-scope test and returns a ledger without edits. Speed mode makes behavior-preserving speedups with measured before and after timings. Prune mode proposes removals and deletes only ledger items the user approves. Use when a suite is slow, flaky, bloated, or due for a value review.

  • ui-motion

    Automatic

    Fix or diagnose existing application motion: loading continuity, state handoffs, overlay lifecycles, interruptible gestures, or animations that stutter under load, with rendered checks. Preserve tokens and real state. New motion: find-animation-opportunities, then animate; grading a diff: review-animations; videos: motion-film.

    6 sources for ui-motion

  • Build or redesign a web app from researched references, or document its existing design language in DESIGN.md. To audit generic or AI-looking UI use deslop-ui; for interaction motion use ui-motion; to make an existing surface fit widths and states use fit-surface.