AI Dev Kit
  • Docs
  • Skills
  • Commands
  • Mods
  • Standards
  • MCP
GitHub
AI Dev Kit
  • Docs
  • Skills
  • Commands
  • Mods
  • Standards
  • MCP
  • Add-ons
  • Libraries
GitHub

AI Dev Kit

A free, MIT-licensed toolkit of layered coding standards and on-demand skills for AI coding agents.

Built for Claude Code, OpenCode, Codex and any agent that reads markdown.

Explore

DocsSkillsCommandsMods

The kit

StandardsMCPAdd-onsLibraries

AI Dev Kit

ReadmeLicenseContributeRepository
GitHub
StartChapter 1Chapter 2GroupFinderNext
Docs

Mods

Hook-based plugins that run in code and can block an action.

Chapter 1

Skills ask. Mods enforce.

A mod runs in code, so unlike a skill it fires every time.

What a mod is

A skill is judged by the model and may not load. A mod is a hook on a tool call or a session event, so it runs whether or not the model remembers the rule.

Mods are optional and fail open: if one breaks, your work continues. No kit rule depends on one.

Mods need Claude Code 2.1.287 or later. They are Claude Code only.

Chapter 2

Review before you install

A mod runs with your permissions and sees every prompt and tool call.

How to install one

The mod-setup skill checks your Claude Code version, lets you pick mods, validates each one and shows what it can touch. You say yes per mod.

Add the kit marketplace once with claude plugin marketplace add <path-to-ai-dev-kit>/mods, then install by name with claude plugin install <name>@ai-dev-kit-mods.

Four mods run a script that ships with a skill. A mod whose skill is missing loads fine but does nothing, by design.

For a new user the order is budget-ledger first, because it measures the rest. Turn one off in /plugin, or start with --safe-mode to turn them all off for a session.

Group 1 of 4

Blockers

These deny an action before it happens.

precommit-gateBlockers

Gates git commit and push, in Bash and PowerShell. Blocks leaked secrets, secret files, bad messages and --no-verify. Needs the pre-commit skill's check.js.

guardBlockers

Denies edits to secret files, lockfiles and .git, and destructive commands: root-like recursive deletes, force push, reset --hard, clean -f, publish, download piped into a shell. Config in .claude/guard.json.

standards-chainBlockers

Denies the first Edit or Write of a file until its coding-standards chain has been read, and again after a compaction. Does nothing without a coding-standards folder.

aiignore-guardBlockers

Keeps the AI out of files it should not read: your .aiignore, paths that belong to another role, the current task's scope. Does nothing without these files.

Group 2 of 4

Result shapers

These change what the model sees, not what it may do.

edit-checkResult shapers

After each Edit or Write, runs the pre-merge-gate skill's edit-check.js and shows the model only the problems: syntax, ESLint, Stylelint and coding-standards rules.

output-compactResult shapers

Runs session-budget's compact.js on every Bash or PowerShell result before the model reads it. Drops noise, keeps every failure line.

prefilter-rankResult shapers

Ranks Grep and Glob results of 5 to 40 files by relevance to the task, using the local Laya prefilter. Calls 127.0.0.1 only. Does nothing if Laya is not set up.

Group 3 of 4

Reminders

One line, once, and never a block.

debug-nudgeReminders

At the second failed fix of a test or build command, reminds the model to stop editing and apply the debug-protocol skill.

eta-nudgeReminders

Nudges once when a command is backgrounded, then every 10 minutes after the first 5, to restate a time estimate. Never computes one.

design-audit-offerReminders

After the second distinct UI file edit in a project with a design source, tells the model to offer a UI/UX audit once. Never audits or changes the design.

state-auditReminders

After each Edit or Write of UI code, runs the implicit-state-audit scan with no model and shows only the hidden-state findings. Adds nothing for a clean file. An optional second reviewer runs only if the project sets one in .claude/state-audit.json. Never blocks.

skill-hintReminders

On a typed prompt, matches keywords in code with no model call and attaches one line naming skills that may apply, once per skill per session. An archived skill is named with the /unarchive command. It never loads, archives or restores anything.

marker-guardReminders

When a reply does not start with the [CX] marker your protocol file defines, shows one warning line beneath it. After a compaction it shows one notice. Never blocks.

Group 4 of 4

Context and measuring

Add context, or just watch.

budget-ledgerContext and measuring

Logs real token usage, cost, cache and tool-result sizes per turn, one file per session, and adds /budget. Observes only.

feature-usageContext and measuring

Records which skills load, commands are typed and kit scripts run, one name-only row each in ~/.ai-dev-kit/usage/log, and adds /kit-usage (on, off, notify, status, report). Observes only. The report needs session-budget's usage.js.

session-journalContext and measuring

Keeps a script-written journal of one session outside the project in ~/.ai-dev-kit/sessions/: the ask, files edited, commands run with secrets scrubbed, failures. After a compaction the summarizer is handed those facts. Journals idle 5 hours or more are deleted at the next session start. Never touches handover.md.

gap-logContext and measuring

Appends each STANDARDS GAP line an answer contains to .claude/standards-gaps.jsonl (time, session, gap), so the holes in the standards pile up for the maintainer. Adds no text to the answer or context, and stops at 1 MB.

learnings-injectContext and measuring

Adds a skill's learnings.md to the Skill call result once per session, so the skill's start-of-skill read needs no tool call. Context only.

Finale

Not sure which one?

Start from the situation.

  • You want leaked secrets kept out of commits

    precommit-gate

  • You want destructive commands stopped

    guard

  • You want edits to wait until the standards were read

    standards-chain

  • You want to see what a session costs

    budget-ledger

  • You want a warning when the [CX] marker is dropped

    marker-guard

Keep reading

On to the next page

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

Page 3 of 7

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
Previous pageCommandsNext pageCoding standardsLayered rules for CSS, HTML, scripts, git, SEO, performance, accessibility, QA, design and AI behaviour.