Skills

Instructions an agent reads only when the moment calls for them.

Chapter 1

An agent cannot remember everything

So the rules live on cards, and it pulls one only when the job needs it.

  1. You saycommit this
  2. The agent matchesa commit is about to run
  3. The card loadspre-commit
What a skill is

A skill is a packaged set of instructions for one kind of task. The agent sees a one-line description of each skill, and when the task matches a description it reads the full file and follows it.

A deploy checklist the agent reads before it ships. A review checklist it reads before it approves a change. A brand voice guide it reads before it writes copy. Each one is plain text, tied to a situation.

Chapter 2

One card, not the whole library

Everything else stays on the shelf and costs nothing.

  • plan-first
  • handover
  • debug-protocol
  • perf-audit
  • intent-capture
  • version-bump
  • resolve-conflicts
  • responsive-check
  • optimize-assets
  • session-budget
  • role-session
  • e2e-scaffold
  • pre-commit
Why skills exist

An agent cannot carry every rule in its head at all times. Loading a rule only when it is needed saves tokens and keeps the agent focused on the task in front of it.

Chapter 3

In this kit, a card is a file

A name, a trigger line that says when to load it, and a body.

Name
pre-commit
Trigger
Asked to commit, "ship it" or "save this", or a git commit is about to run. Load BEFORE running git commit.
Body
The commit pass: stray files, debug leftovers, secrets, message format, version bump.
How a skill gets picked

Installing the kit copies each chosen skill into your project as its own folder with a SKILL.md file.

Most descriptions end with a Load BEFORE line that names the action the skill guards. When the agent is about to take that action, it loads the skill first. Loading is judged by the model, not guaranteed, so the kit's mods are what enforce the critical gates.

A skill can ship a companion script that does the mechanical part, such as counting rule violations or bumping a version. The agent reads the script's output instead of reading every file, which keeps the cost low.

Chapter 4

Three ways a card starts

Some start themselves. Some wait for you. Some you switch on once.

  • Automatic

    The agent notices a matching situation and loads the skill itself. You do nothing except ask for the task.

  • Manual

    The skill starts only when you ask for it in words, such as installing the kit or hiding a file from git.

  • Opt-in

    Not installed by default. You turn it on once, and after that it behaves like an automatic skill.

Any skill can also be asked for by name. The mode shows how it normally starts.

Chapter 5

Nine shelves

Every skill in the kit lives on one of them.

Shelf 1 of 9

Workflow

Session and process discipline: how a task is understood, planned, checked and committed.

handoverAutomatic
What it does
Keeps handover.md current so a new session resumes with no re-explaining.
When it applies
Session start, before ending one, or on handover, wrap up, where were we, resume.
Your role
Say wrap up or resume when you want it.
AI's role
Reads the handover at the start and rewrites it at the end.
agent-usageAutomatic
What it does
A default-never policy for the Agent tool, with two scope-anchored exceptions and an approval gate.
When it applies
An agent or subagent call is being considered.
Your role
Approve when asked.
AI's role
Works inline unless an exception applies and you approve.
mode-kernelAutomatic
What it does
A central table for how the stop-and-wait gates behave in each session mode.
When it applies
A gate is about to apply in autonomous, planning or background mode.
Your role
Nothing.
AI's role
Looks up whether to pause, ask or proceed for the current mode.
session-budgetAutomatic
What it does
Token discipline: targeted reads, no restating, cheap-model delegation. Report and compact scripts ship with it.
When it applies
Tokens, cost, budget or a filling context mentioned, or a long task starts.
Your role
Ask about spend when you want a report.
AI's role
Reads only what it needs and shrinks verbose output.
debug-protocolAutomatic
What it does
Reproduce, form one hypothesis, run the cheapest disproof. No shotgun edits.
When it applies
Something fails and the cause is unknown, or after two failed fixes.
Your role
Describe the failure.
AI's role
Investigates in order before it edits anything.
intent-captureAutomatic
What it does
Pins goal, constraints and done-when before planning an ambiguous request.
When it applies
A vague request, two or more plausible readings, or no acceptance criterion.
Your role
Confirm or correct the goal it states.
AI's role
Writes the goal, constraints and done-when, then hands them to planning.
plan-firstAutomatic
What it does
A five-line plan and file list before multi-file work. In plan mode it plans for a cheaper executor and verifies the result.
When it applies
The task touches 3+ files, has ambiguous scope or several deliverables.
Your role
Redirect the plan, or let it proceed.
AI's role
States the plan, then executes without asking again. A big plan can get a /goldfish review first, and /dupe copies a block that must repeat across files.
interpretation-checkpointAutomatic
What it does
A Files, Changes and Assumptions breakdown you can correct before a wide edit.
When it applies
The task touches 2+ files and changes 2+ distinct parameters.
Your role
Correct any wrong assumption.
AI's role
Lists what it will touch and what it assumes, before the first edit.
pre-merge-gateAutomatic
What it does
Checks a diff against the loaded standards before handoff or commit. A companion script runs the mechanical checks.
When it applies
A change is about to be called done, ready for review, handed off or committed.
Your role
Nothing.
AI's role
Self-checks the diff before it says done.
pre-commitAutomatic
What it does
The commit pass: stray files, debug leftovers, secrets, message format, version bump. A script can install it as git hooks.
When it applies
You ask to commit or ship it, or a git commit is about to run.
Your role
Ask for the commit.
AI's role
Runs the check, fixes or reports findings, then commits.
version-bumpAutomatic
What it does
Bumps a package.json version and inserts the matching CHANGELOG.md entry, for the whole project, one group or one package.
When it applies
A push is about to happen in a project that follows the versioning standard, or you say bump the version.
Your role
Pick the bump kind if asked.
AI's role
Runs the bump script and checks the changelog entry.
resolve-conflictsAutomatic
What it does
Resolves merge and rebase conflicts by mode (both, ours, theirs), pace (all, auto, step) and developer role.
When it applies
A merge, rebase, cherry-pick or pull stops with conflicts.
Your role
Choose the mode and pace.
AI's role
Lists the conflict set, verifies role scope, then resolves.
responsive-checkAutomatic
What it does
A range-based viewport check and fix. A probe script plans viewports and a measure script reads them in Playwright.
When it applies
Questions about screen sizes, phones, tablets, horizontal scroll or breakpoint bugs.
Your role
Name the range or width to test.
AI's role
Measures first, then fixes what fails.
optimize-assetsAutomatic
What it does
Converts, compresses and pre-compresses assets (webp, avif, jpg, mp4, woff2). Originals are never touched.
When it applies
Converting, compressing or shrinking images, video, fonts or SVG.
Your role
Say which assets and the target format.
AI's role
Scans for heavy files, converts with the tools installed, lists references.
perf-auditAutomatic
What it does
A measured, ranked performance audit: payload, then loading, then runtime, then 3D.
When it applies
Performance, slowness, bundle size, Lighthouse or page weight, or a pre-release check.
Your role
Ask for the audit.
AI's role
Measures, ranks the findings and proposes fixes.
role-sessionAutomatic
What it does
A lane protocol for parallel sessions: role charters, atomic file claims, a git token and session ids.
When it applies
A role is assigned, or a handover board or locks folder exists.
Your role
Assign the role.
AI's role
Claims files before touching them and stays in its lane.
e2e-scaffoldAutomatic
What it does
Scaffolds a reusable Playwright config, fixtures and smoke test once per project.
When it applies
Setting up e2e or Playwright testing, or before a first spec with no fixtures layer.
Your role
Ask for end-to-end tests.
AI's role
Creates the scaffold before writing any spec.
readme-writerAutomatic
What it does
Writes or trims a README so the hook comes first and every number, flag and claim is checked against the code. A script audits the claims.
When it applies
A README.md is written, restructured, trimmed or fact-checked.
Your role
Ask for the README work.
AI's role
Runs the audit script, then edits the README to match what the code does.
untrackManual
What it does
Hides files from the Changes tree on your machine only. Nothing is committed.
When it applies
You say don't track X, hide X from git, or track X again.
Your role
Name the files to hide or show.
AI's role
Runs the hide script and reports what changed.

Shelf 2 of 9

Standards

Convention systems: the rules code is written against, and the audits that check them.

coding-standardsAutomatic
What it does
Loads and enforces the layered coding-standards chain before any edit.
When it applies
About to edit CSS, HTML, JS/TS or framework files in a project with a coding-standards folder.
Your role
Nothing.
AI's role
Loads the rules for the file's role, then edits to them.
code-auditAutomatic
What it does
A project-wide audit of the script-checkable rules. A script counts violations by rule and lists file and line.
When it applies
Whole-project cleanup, audit or retrofit, or find all violations of a rule.
Your role
Ask for the audit.
AI's role
Runs the script and reads only the files it flags.
implicit-state-auditAutomatic
What it does
Finds hidden state machines: boolean flags that allow impossible combinations, state scattered over classes or data attributes, and timer-sequenced state. A script scans first; a model sees only the flagged files.
When it applies
UI code keeps state in several flags, classes, data attributes or timers, or a bug only happens in some order.
Your role
Ask for the review, or type /state-audit.
AI's role
Names the one state variable that replaces the flags for each failing file.
security-auditAutomatic
What it does
A security audit procedure with a five-part finding bar, confirmed versus needs-validation findings and a report format.
When it applies
A security review, is this safe, or a vulnerability hunt.
Your role
Point at the code to review.
AI's role
Reports only findings that meet the bar.
ui-designAutomatic
What it does
Leads design only when starting from scratch. With a design doc or Figma it only offers a UI and UX audit.
When it applies
Designing a new page, screen or component with no design doc, or a UI/UX audit request.
Your role
Answer the ask-once questions.
AI's role
Runs the pre-flight, builds the page and sweeps it.
ux-theoryAutomatic
What it does
Designer craft values for colour, type, composition, motion and effects. Read only while making or extending a theme.
When it applies
Creating a theme or design system, extending one with a new element, or reviewing visual design.
Your role
Nothing.
AI's role
Checks the new element against the craft values.
comment-styleAutomatic
What it does
A per-project comment level: none (default), terse or descriptive. Saves tokens.
When it applies
Writing or editing code, or changing comment style or verbosity.
Your role
Pick the level at install, change it any time.
AI's role
Writes comments at that level and no more.

Shelf 3 of 9

Meta

Maintaining the library itself: writing skills, commands, mods and installers, and installing the kit.

install-kitManual
What it does
Installs, updates, uninstalls or archives skills and git hooks, all at once, hand-picked, auto-detected or by stack preset. A script does the copy and a second script verifies it. Run bare in a terminal it shows arrow-key menus with undo and redo.
When it applies
You ask to install skills or standards, or a project lacks them.
Your role
Choose what to install.
AI's role
Runs the install script and the verify script before it reports done.
mod-setupManual
What it does
An opt-in installer for the kit's mods: version check, review of what each mod touches, install, confirm. Claude Code only.
When it applies
Installing, enabling, reviewing or removing mods, or mods not loading.
Your role
Review and approve each mod.
AI's role
Explains what a mod touches, installs it and confirms it loads.
skill-scopeAutomatic
What it does
The single classification of pinned, conditional and archivable skills. Install-kit and handover both read it.
When it applies
Classifying, installing, pinning or archiving a skill.
Your role
Nothing.
AI's role
Reads the classification and edits it nowhere else.
skill-writerAutomatic
What it does
The quality bar for writing new skills: triggers, checkable rules and the improvement loop.
When it applies
Creating, reviewing or rewriting the description of a skill.
Your role
Ask for the skill.
AI's role
Writes the skill against the bar.
command-writerAutomatic
What it does
The quality bar for slash commands: script first, one scoped permission, capped output. Used only when building the kit.
When it applies
Creating or improving a slash command or anything in skills/commands.
Your role
Ask for the command.
AI's role
Builds the command and runs the registration checklist.
cli-writerAutomatic
What it does
The quality bar for CLI tools and installers: zero dependencies, prompts only on a TTY, safe re-runs. Used only when building the kit.
When it applies
Creating or improving a command-line tool, installer or setup script.
Your role
Ask for the tool.
AI's role
Writes the tool and tests a cold start.
mod-writerAutomatic
What it does
The quality bar for Claude Code mods: mod versus script versus hook, fail-open rules, a validate and test loop. Claude Code only.
When it applies
Creating or improving a Claude Code mod or anything in mods.
Your role
Ask for the mod.
AI's role
Chooses the right mechanism and validates the result.
misfire-logOpt-in
What it does
Logs a skill, mod or add-on that should have fired but did not. A report script ranks repeat offenders.
When it applies
Triggers only when configured. Never installed by default.
Your role
Turn it on and read the report.
AI's role
Appends an entry when it notices a misfire.

Shelf 4 of 9

Memory

Persistent knowledge that survives between sessions.

memory-bankAutomatic
What it does
Repo-committed project memory: decisions, context and solved mysteries.
When it applies
Starting work in a previously touched area, after a lasting decision, or on remember this and why did we do X.
Your role
Say remember this when something should last.
AI's role
Reads the bank before answering and records what is non-obvious.
memory-gardenerManual
What it does
Prunes and merges learnings files and memory banks. A script lists bullets that name deleted files.
When it applies
You ask to clean or prune skills or memory, a learnings file passes about 30 bullets, or a handover passes 120 lines.
Your role
Ask for a cleanup and approve the removals.
AI's role
Merges duplicates and deletes disproven notes.
skill-ablationManual
What it does
A periodic pass that archives skills and runs bare, restoring only the instructions real work proves are missing.
When it applies
On ablation, run bare or prune skills, after a major model upgrade, or about six months since the last run.
Your role
Start the run and judge what to restore.
AI's role
Archives, runs the task, reports what was missed.

Shelf 5 of 9

MCP

Optional tool servers for any MCP-capable agent, all set up through one manager.

mcp-managerOpt-in
What it does
The meta skill for every MCP server: a setup interview, per-tool config and an MCP-first-then-fallback rule.
When it applies
Adding, setting up or fixing an MCP, or a configured server covers the task.
Your role
Say set up MCP servers and answer the interview.
AI's role
Wires each server into your tool's config.
local-visionOpt-in
What it does
Free, offline image reading with Qwen3-VL through llama.cpp, for models without vision. The installer is hardware-aware.
When it applies
Chosen during the MCP setup.
Your role
Pick it in setup.
AI's role
Calls the server to read an image.
image-genOpt-in
What it does
Image generation with a fallback chain, plus colour and sharpen enhancement for existing photos.
When it applies
Chosen during the MCP setup.
Your role
Add keys if you want the hosted providers.
AI's role
Generates, reports each fallback step, and warns near quota limits.
stitchOpt-in
What it does
Google Stitch UI prototyping. Try it first, then fall back to the artifact or HTML flow.
When it applies
Chosen during the MCP setup.
Your role
Pick it in setup.
AI's role
Prototypes screens, and falls back if the server is unavailable.

Shelf 6 of 9

Libraries

Guides for the author's own libraries.

strata-cssAutomatic
What it does
Use, debug and fix the Strata CSS framework at its source. A coverage script shows which utilities exist.
When it applies
Using, installing, migrating to or debugging Strata classes.
Your role
Nothing.
AI's role
Checks coverage before it uses a class, never guesses names.
triforgeAutomatic
What it does
Package picking and a findings workflow for the Triforge Three.js suite.
When it applies
Any Triforge package or Blender-style node work in Three.js.
Your role
Nothing.
AI's role
Picks the package, uses its API and fixes issues at source.

Shelf 7 of 9

Stack

Technology-specific discipline.

threejs-sceneAutomatic
What it does
Three.js discipline: shaders, disposal, scroll cameras and render hygiene.
When it applies
Three.js scenes, meshes, shaders, materials, cameras or visual glitches.
Your role
Nothing.
AI's role
Loads before it edits scene code.
astro-pageAutomatic
What it does
Convention-driven Astro scaffolding: discover the project, mirror siblings, verify.
When it applies
Adding an Astro page, component, overlay or entry script, or wiring CSS or TS into Astro.
Your role
Name what to add.
AI's role
Copies the project's own pattern instead of inventing one.
shopify-toolkit-installManual
What it does
Pulls Shopify's own AI toolkit skills into a project live. They are never forked into this kit.
When it applies
You say setup for shopify, install shopify skills or shopify ai toolkit.
Your role
Ask for the install.
AI's role
Fetches the current toolkit and installs it.

Shelf 8 of 9

Add-ons

Optional capabilities you opt into. None is part of a default install.

personalityOpt-in
What it does
Six original personas, female and male at novice, mid and pro levels, that change how replies are worded. Wording only, never facts.
When it applies
You ask to set up, pick, change or remove a reply persona.
Your role
Choose a persona with /persona.
AI's role
Words replies to fit the persona.
narrateOpt-in
What it does
A stop hook that reads each reply aloud with Piper voices. It costs no tokens and never touches the model.
When it applies
You want replies read aloud.
Your role
Set voice, volume and speed with /narration.
AI's role
Nothing. The hook does the reading.
project-contextOpt-in
What it does
A private, gitignored project-context folder with an index per sub-folder. A scan script tracks changes so only new files are re-read.
When it applies
You ask to set up project docs, or a project-context folder exists.
Your role
Drop in requirements, brand and design docs.
AI's role
Reads the index first, then only what changed.
system1-prefilterOpt-in
What it does
A typed-decision prefilter that cuts fetching-stage tokens by filtering candidates before the model reads them. A local free provider is supported.
When it applies
Only when configured. Never installed by default.
Your role
Configure the provider.
AI's role
Ranks candidates and escalates when confidence is low.

Shelf 9 of 9

Models

Driving each model at full capacity: routing, effort, thinking and cost rules per model.

The 20 model protocols
  • fable-5-1
  • opus-5-5
  • sonnet-5-5
  • haiku-5-5
  • opus-5
  • fable-5
  • opus-4-8
  • sonnet-5
  • haiku-4-5
  • opus-as-fable
  • big-pickle
  • glm-5-2
  • deepseek-v4
  • kimi-k2-6
  • qwen3-coder
  • minimax-m3
  • devstral-2
  • mimo-v2-5
  • gpt-oss
  • local-small-models
claude-all-modelsAutomatic
What it does
Fleet routing for the Claude lineup: the right model per task, escalation rules and cost discipline.
When it applies
Any model-selection or cost-optimisation decision.
Your role
Nothing.
AI's role
Picks or recommends a model for the task.
opencode-all-modelsAutomatic
What it does
Fleet routing for open-weight models in OpenCode, hosted versus local. It holds the per-model protocols.
When it applies
Any OpenCode or open-weight model choice.
Your role
Nothing.
AI's role
Routes the task and applies the chosen model's protocol.
hooks-enforcementOpt-in
What it does
Optional Claude Code hooks that assist the AI standards rules mechanically, plus a git-commit gate on pre-commit. Claude Code only.
When it applies
Installing into a Claude Code project that uses the AI standards or pre-commit.
Your role
Opt in at install.
AI's role
The hooks run the checks.

Finale

Not sure which card?

Start from the situation.

  • A task touches three or more files

    plan-first

  • A request is vague

    intent-capture

  • Something fails and you do not know why

    debug-protocol

  • You are about to commit

    pre-commit

  • A merge stops with conflicts

    resolve-conflicts

  • Pages feel slow or heavy

    perf-audit

  • You want a session to resume cold

    handover

  • You want to set up MCP servers

    mcp-manager

  • You want replies read aloud

    narrate

Keep reading

On to the next page

The docs are a short run of pages. Use the buttons, or jump to any page.