lippstack

Mods and skills for Claude Code, tuned to how we work. The skills also run in Codex, Gemini CLI, Kimi, OpenCode, Cursor and other agents: see Other agents.

By Filipe Moreira.

Plugins

Plugin	Type	What it does
savvy-progress	Mod	/agents-info opens a live panel of subagents with model, context, tokens and time, plus the account's plan quota (5-hour and weekly). Each tier gets its own pixel Clawd costume.
stage-bar	Mod	An animated bar above the prompt that tracks the stages of your task, one color per stage. Claude names the stages and moves the bar as it works. Try /stages-demo.
done-ping	Mod	A notification when a task of a minute or more finishes, or when Claude is waiting for a permission or your answer (macOS Notification Center / Linux & Windows terminal notification channel).
guardrails	Mod	Hooks, not requests: /careful (on by default) asks you before rm -r, git reset --hard, force-push, DROP TABLE and the like run; /freeze <path> refuses edits outside a folder. Works in bypass mode too.
where-line	Mod	A status line under the prompt: project, git branch, model and how full the context is (🟢 🟡 🔴).

Skills

Plugin	Source	What it does
design-skills	Copied here	From Emil Kowalski: apple-design (Apple-style motion, gestures and materials on the web), review-animations, break-ui (worst-case data), prototype (several versions behind a picker) and animation-performance (animations that drop frames). hallmark (Nutlope, 2.0): pages that don't look AI-generated, with audit and redesign.
po-skills	Ours, plus two adapted	jira-stories and release-notes (ours): stories into Jira, finished work into release notes on Confluence. brainstorming (enriched with YC forcing questions) and writing-plans (with Decision Briefs and Ponytail debt tracking) from obra/superpowers.
eng-skills	Ours	Engineering, QA & delivery: investigate (systematic root cause debugging with the 3-strike rule), qa (exploratory verification across UI/API/CLI with mandatory regression tests), and ship (pre-flight checks, doc sync, PR). freeze and careful live in the guardrails mod.
ponytail	DietrichGebert/ponytail	Lazy senior dev mode: the simplest code that works. YAGNI, stdlib first, no unrequested abstractions. Commands /ponytail-audit, /ponytail-review and more.
humanizer	blader/humanizer	Rewrites AI-sounding text so it reads like the writer, without changing what it says. Based on Wikipedia's "Signs of AI writing".
video-editor-skills (opt-in)	appariciojunior/video-editor-skills	Claude as a first-pass video editor in DaVinci Resolve Studio: cuts retakes and dead air from the transcript, keeps B-roll, scores, captions, motion graphics, YouTube export. Needs Resolve Studio, ffmpeg, Python and a speech-to-text API. CC BY-NC 4.0: no commercial use.

design-skills is copied into this repo; refresh it with scripts/sync-vendor.sh. po-skills and eng-skills are maintained in this repo. ponytail, humanizer and video-editor-skills are listed in our marketplace but install straight from their own repos, so they always get upstream updates. video-editor-skills is not copied here on purpose: its CC BY-NC 4.0 license does not match this repo's MIT, and it is only useful if you edit video in Resolve. So setup.sh and setup.ps1 do not install it; add it when you want it.

PM skills by Dean Peters

From deanpeters/Product-Manager-Skills, installed from its own marketplace (pm-skills). It has 77 skills; setup.sh installs only these 12, so Claude does not have 77 descriptions to choose from on every request:

Area	Skills
Backlog	user-story, user-story-splitting, epic-breakdown-advisor, epic-hypothesis
Spec and planning	prd-development, prioritization-advisor, roadmap-planning
Discovery	discovery-process, jobs-to-be-done, problem-statement
Stakeholders	incoming-request-advisor, stakeholder-mapping

Add another with claude plugin install <skill>@pm-skills. They are CC BY-NC-SA 4.0: fine to use at work, not to resell, so we don't copy them into this repo.

Official Anthropic plugins

From anthropics/claude-plugins-official, installed by setup.sh:

Plugin	What it does
hookify	Rules that warn or block, written in plain words: /hookify Warn me before any rm -rf. /hookify alone looks at the conversation for things you corrected. /hookify:list shows the rules
claude-md-management	Audits a project's CLAUDE.md and, with /revise-claude-md, adds what this session learned
session-report	An HTML report of where the tokens went: cache use, subagents, skills and the most expensive prompts
playwright	Lets Claude open a browser, click through a page, fill forms and take screenshots, to check what it built

Built-in mods we turn on

Mod	What it does
cc-plugin-you-should-know	A side agent that reads what Claude writes and shows a note above the prompt when it spots something you might miss, such as a retry that could charge a customer twice. Off by default, it uses a bit more of the plan, and it stays quiet on a strong model unless it finds something.

It ships inside Claude Code, so there is nothing to copy here: setup.sh / setup.ps1 just switches it on.

Install

From a clone of this repo:

macOS / Linux

./setup.sh

Windows (PowerShell)

powershell -ExecutionPolicy Bypass -File .\setup.ps1

The -ExecutionPolicy Bypass applies only to this run; Windows blocks unsigned scripts by default.

Or by hand

claude plugin marketplace add lippdev/lippstack
claude plugin install savvy-progress@lippstack
claude plugin install stage-bar@lippstack
claude plugin install done-ping@lippstack
claude plugin install where-line@lippstack
claude plugin install guardrails@lippstack
claude plugin install design-skills@lippstack
claude plugin install po-skills@lippstack
claude plugin install eng-skills@lippstack
claude plugin install ponytail@lippstack
claude plugin install humanizer@lippstack
claude plugin install video-editor-skills@lippstack   # optional: DaVinci Resolve video editing, non-commercial license

Then restart Claude Code, or run /reload-plugins in an open session.

Other agents (Codex, Gemini CLI, Kimi, OpenCode, Cursor…)

The skills follow the open Agent Skills format, so any agent that reads SKILL.md can use them. The mods can't: they run on Claude Code's own hooks.

	Claude Code	Other agents
Skills: design-skills, po-skills, eng-skills, ponytail, humanizer, video-editor-skills (optional), PM skills	✅	✅
Mods: savvy-progress, stage-bar, done-ping, where-line, guardrails	✅	❌
Official Anthropic plugins, cc-plugin-you-should-know	✅	❌

From a clone, add --agents with the agent ids. Claude Code is installed too when its CLI is present, and skipped when it isn't:

./setup.sh --agents codex,gemini-cli,kimi-code-cli

powershell -ExecutionPolicy Bypass -File .\setup.ps1 -Agents codex,gemini-cli,kimi-code-cli

Or without a clone, with the skills CLI (needs Node.js):

npx skills add lippdev/lippstack -g                       # pick agents and skills interactively
npx skills add lippdev/lippstack -g -a codex -a opencode --skill '*' -y
npx skills add DietrichGebert/ponytail -g --skill '*'
npx skills add blader/humanizer -g --skill '*'
npx skills add appariciojunior/video-editor-skills -g --skill '*'   # optional
npx skills add deanpeters/Product-Manager-Skills -g -s user-story -s prd-development   # pick the PM skills you want

Common ids: codex, gemini-cli, kimi-code-cli, opencode, cursor, github-copilot, windsurf, antigravity; '*' targets every agent the CLI knows. A few things differ outside Claude Code: /skill-name may be spelled differently (Kimi uses /skill:name), skills that offer subagents fall back to doing the work in a fresh pass, and jira-stories / release-notes need an Atlassian MCP server in that agent.

How to use

Skills load on their own when a request matches their description. To make sure one runs, name it in the request ("use user-story to…") or type / and pick it from the list; plugin skills show up as plugin:skill, for example /po-skills:brainstorming or /eng-skills:investigate.

From idea to shipped feature

The skills are meant to be used in this order. Skip the steps a small task doesn't need: a one-line fix goes straight to step 7.

Step	Use	Ask something like	You get
1. Understand the ask	incoming-request-advisor	"A director sent this: … What are they really asking for?"	The literal ask separated from the job behind it, and the questions to send back
2. Frame the problem	problem-statement, jobs-to-be-done, discovery-process	"Write a problem statement for investors who can't find their portfolio report"	Who has the problem, what it costs them, how you'd know it's solved
3. Decide what to build	prd-development, prioritization-advisor, roadmap-planning	"Turn this into a PRD" / "Help me prioritize these 8 items for Q1"	A PRD with success criteria; a framework picked for your context and a ranked list
4. Shape the work	epic-hypothesis, epic-breakdown-advisor, jira-stories	"Break epic INV-120 into stories and create them in Jira"	Stories with Gherkin criteria in the team's style, created under the epic after you approve
5. Design the build	brainstorming	"Let's build an export-to-PDF button for the report"	Questions one at a time with YC forcing perspectives (10-star, narrowest wedge, elimination), a size for the work, and a design doc in docs/specs/
6. Plan the build	writing-plans	"Write the plan for the approved design"	Small tasks with files, names and tests in docs/plans/, plus Decision Briefs and Ponytail shortcut: debt tags
7. Build it	ponytail (always on), /freeze, investigate	"Run the plan" / "investigate why the worker crashes"	Smallest code that does the job. When debugging: Iron Law (root cause first, 3-strike rule) and scope freezing
8. Verify it	qa, /ponytail-review, hallmark audit, break-ui, animation-performance	"QA this new endpoint" / "break the portfolio card" / "/ponytail-review"	Adversarial exploration beyond happy path; mandatory automated regression tests for any bug found; visual & code reviews
9. Ship it	ship	"Ship this feature"	Pre-flight git cleanliness, full test pass, Diataxis doc synchronization, atomic commit, push and PR
10. Tell people	release-notes, humanizer	"Release notes for sprint 42 of INV" / "humanize the PR description"	A team version and a business version, published to Confluence after you approve; then text that doesn't read as AI-written

Product (PM skills)

• Give context in the first message, such as product, user, constraint and deadline; the skills skip the questions you already answered.
• Most of them run as a short interview when you arrive with nothing ("use prd-development") and as a straight draft when you paste material ("use prd-development on these meeting notes: …").
• stakeholder-mapping before a big alignment meeting: list the people and it ranks who to talk to first and how.
• The other 65 skills in the library (pricing, TAM/SAM/SOM, positioning, SWOT, end-of-life…) are one install away: claude plugin install <skill>@pm-skills. The full list is in their README.

Jira and Confluence

Both need the Atlassian connector, and both show you the result before touching Jira or Confluence.

• jira-stories reads a few recent stories in the project first and copies their language, title style and labels. Give it an epic key, a PRD or a paragraph. It checks for duplicates under the epic and never fills in estimates or assignees on its own.
• release-notes takes a sprint, a fix version or dates. It lists what it left out (sub-tasks, spikes, tech debt) so you can pull an item back in, and can save the Confluence page as a draft.

Brainstorming and plans

• brainstorming asks one question at a time. It uses YC-style forcing questions to uncover the 10-star vision, the narrowest wedge (minimum slice of value), and what can be eliminated completely.
• It reads the project before its first question. Before it writes anything, it lays out two or three ways to build it, with the trade-offs and a recommendation, and you pick one.
• It then sizes the work: a quick task is just done, a small change gets a short design doc, and a project gets the full one in docs/specs/. Both record the approaches considered and the one you chose.
• For anything visual it can open a local page (localhost) with mockups side by side.
• writing-plans produces bite-sized, testable tasks in docs/plans/. If architectural trade-offs or intentional shortcuts are taken, it frames a Decision Brief and marks the code with shortcut: <reason>, connecting straight into ponytail-debt.
• At the end it asks how to run the plan: subagent-driven (one fresh agent per task plus review) or native (this session does it, review at the end).

Engineering, QA and Delivery (eng-skills)

Professional development practices built directly into the agent workflow:

Command / Skill	Methodology
investigate	The Iron Law of Debugging: No fixes without proven root cause. Traces data flow backward. Enforces the 3-strike rule: if 3 fix attempts fail, the agent hard-stops, admits the mental model was incomplete, and asks for guidance instead of thrashing.
qa	Exploratory Verification: Systematically tests beyond the happy path (empty values, hostile inputs, duplicate clicks, interruptions). Enforces the Golden Regression Rule: every bug found must have an automated test that fails before the fix and passes after.
ship	Pre-flight & Delivery (only when you call /ship): Synchronizes base branch, runs tests and typechecks, performs a Diataxis documentation sanity check (ensuring README/specs/docs match code changes), creates an atomic commit and opens a PR.

/freeze <path> and /careful come from the guardrails mod: hooks enforce them, so they hold even in bypass mode.

Ponytail

On by default in every session, at the full level.

Say or type	What happens
/ponytail lite	Builds what you asked, and names the smaller option in one line
/ponytail full	The default: smallest complete change
/ponytail ultra	Also pushes back on parts of the request the need doesn't justify, before building
"stop ponytail" / "normal mode"	Turns it off for the session
/ponytail-review	Reviews the current change: logic, safety, load, tests, lines that aren't needed
/ponytail-audit	The same review over the whole repo, ranked
/ponytail-debt	Lists every shortcut: comment it left, as a debt ledger
/ponytail-gain, /ponytail-help	Benchmark numbers; quick reference

Humanizer

Takes text an AI wrote and rewrites it so it reads like a person wrote it, keeping the meaning. It is built on Wikipedia's "Signs of AI writing" and checks for 26 patterns. The five strongest: "not X but Y", one-line closers ("Let that sink in."), sayings that sound deep, a staged run-up ("Here's the thing."), and arguing with no one. It edits for human readers. It is not a way to beat AI detectors, and the author says detectors still flag most of its output.

In Claude Code it answers to /humanizer:humanizer; other agents use /humanizer. You can also just ask in plain language.

Say or type	What happens
/humanizer:humanizer and paste text	Rewrites the text and keeps what it says. Doesn't invent facts
"Humanize the prose in docs/launch-post.md"	Rewrites a file's prose
Add a sample of your own writing	The rewrite follows your voice instead of a generic one

Where it fits in the workflow:

• Step 10, tell people. Run it on the team and business versions of the release notes, and on the PR description ship writes, before you publish.
• Steps 3 to 6, documents. PRDs, problem statements and design docs written by an agent pick up the same tics. One pass before you share them with people outside the session.
• Messages and posts. Emails to stakeholders, announcements, LinkedIn, README intros: anything with your name on it.
• Don't run it on code, specs or plans that an agent will read next. They need precision, not a voice.

It only edits text you point it at. Nothing runs by itself.

Video (optional)

video-editor-skills is not installed by setup.sh. Install it with claude plugin install video-editor-skills@lippstack when you want it.

You film and decide the taste; Claude does the first pass in DaVinci Resolve Studio. Raw footage is never modified, and Resolve only gets new timelines and clips. It needs Resolve Studio, ffmpeg, Python 3.9+ and a speech-to-text API with word timestamps.

Say	What happens
"Set up my video editor"	video-editor-setup: workspace folders, keys kept out of the chat, your Resolve template, the small scripts to build first
"Edit the new video"	edit-video: transcribes raw video/, cuts retakes and dead air with vlog pacing, keeps B-roll, builds the timeline in Resolve
"Add captions" / "graphics" / "music"	captions, motion-overlays (Remotion, optional Figma) and music-and-sound (Epidemic Sound MCP, optional)
"Export for YouTube"	youtube-export: Deliver settings for good quality without huge files

It fits the workflow beside step 10: use it for the video that goes with a release (demo, class, vlog), then run humanizer on the title, description and captions text. The license is CC BY-NC 4.0, so keep it out of anything you sell.

Design

Ask	What happens
"Build a landing page for …"	hallmark asks once about material, audience, use and tone, deals a theme and a page structure that isn't the usual hero → 3 cards → CTA, and runs its lint (sloplint, needs Node) until clean before handing it over. Given a URL or a running app, it screenshots the real product instead of mocking it
hallmark audit <file or folder>	A ranked list of what looks generated. Edits nothing
hallmark redesign <target>	New visual layer, same routes, copy and brand. Asks before deleting anything
"Lock the system" (after a hallmark build)	Saves the design as a design.md; later pages reuse it
"Make this sheet feel like iOS"	apple-design: springs, gestures, interruptible transitions, materials, reduced motion
"Review the animations here"	review-animations: a strict review of the motion code
"This animation is janky"	animation-performance: finds what costs the frame (layout properties, React re-renders, blur, will-change) and swaps it for a cheaper path
"Break the portfolio card"	break-ui: renders the component with worst-case data (long names, empty lists, huge numbers) behind a toggle and lists what broke
"Use prototype for a hold-to-delete button"	prototype: several genuinely different versions behind a picker. Runs only when named

Mods

• /agents-info opens or closes the panel of subagents and plan quota.
• The stage bar appears by itself on tasks with 3 or more steps. /stages-demo shows it.
• done-ping needs nothing: walk away during a long task and you get a notification when it's done or stuck on a question. Change the threshold and sound in /config. On macOS the first one may ask for permission; on Linux and Windows it depends on the terminal supporting notifications.
• guardrails asks before a destructive command runs; /careful off silences it for the session. /freeze src/auth keeps edits inside that folder until /freeze off.
• where-line sits under the prompt. When the dot turns 🔴, finish up or start a new session before the context compacts.
• /hookify when Claude keeps doing something you don't want; session-report when you want to know where the plan went.

Credits

savvy-progress is a fork of johnnyvizz/claude-kit, MIT licensed. Its original license is kept in plugins/savvy-progress/LICENSE.