Skip to the setup

My Favorite Skills

This is my Claude Code setup. First the kitchen, my picture of how the pieces around Claude fit together. Then the twenty skills that do real work, with what each one is bad at and credit for the ones I borrowed.

Numbers come from my own Claude Code session logs, June 2 to September 24, 2026, about 16 weeks. Older logs are gone, so anything I used before June is undercounted.

The kitchen

Skills are half of it. This is the setup around them that keeps a long day of work from falling apart.

This is my own way of picturing how a Claude Code setup works, written for people learning to build one. It is not an official diagram.

The counter Context

Context is counter space. It fills up fast, and everything on it gets wiped at the end of the night.

Claude can only work with what's on the counter: its context window, meaning everything it has read or written this session. When the counter gets crowded, the work gets sloppier. I've told Claude it's “getting kind of bloated” more often than I'd like.

Two things keep it clear. context-mode, a plugin by mksglu, sends big command output to a side index so only a summary lands on the counter. And session-handoff gives every session a fresh counter, with a prep list from the last shift. I do override context-mode sometimes: “Read it in full, not context mode.”

Claude's own replies were crowding the counter too. I wrote that it's “always, always verbose unless I specify,” so I had two hooks built. One reminds Claude of my length rule on every message I send, and the other checks each reply when it's done. caveman is the extreme version.

The walk-in Long-term memory

What survives the night: a vault of notes Claude can search, and a lock on who writes to it.

The walk-in is my Second Brain, an Obsidian vault (a folder of linked notes): raw sources in one place, cleaned-up pages in another. I wanted “a lasting source of truth for everything I'm working on,” so Claude stops meeting my projects for the first time every session.

Python hooks stock it. When a session ends, one saves a raw summary. When my message names a known person or project, another puts the matching page's location on the counter. A local search tool finds notes by meaning, so a loose description still turns up the right page.

Writing is locked. Only the capture step in session-handoff and end-session adds curated pages, and identity, money or legal changes wait for my OK. I retired an automatic sync in July because nothing should write to memory unattended. Reading is the weak spot. I've asked Claude, “Are you using your second brain at all?”

Sous chefs Subagents

The head chef hands prep work to sous chefs, so the main session keeps a clean counter.

The main Claude session is the head chef. Subagents are the sous chefs: separate Claude workers that take one job, do it on their own counter, and hand back just the result. A lot of my requests end with “to keep your context window clean.”

Sonnet (the faster, cheaper model) by default, Opus (the stronger one) when the job needs more judgment. I also make sure each one gets enough background, because a sous chef who doesn't know the menu gives confident, wrong feedback.

Every sous chef that touches files gets a security preamble at the top of its instructions. Subagents don't inherit my main rules file, so the rules have to travel with them.

The expediter Hooks

Hooks are the expediter at the pass: small scripts that check every plate before it leaves the kitchen.

A hook is a small script that runs automatically at a set moment, like before a tool runs or when a session ends. I have 53 hook entries across 17 of those moments, backed by 62 scripts and 28 helper modules. GSD brought its own; plenty are mine.

One exists because of a mess. My workspace folder used to be one big git repository (git keeps version history), and it quietly absorbed commits from the projects inside it. Now a guard hook blocks git changes there, and the folder isn't a repository anymore.

Others ask before I spend money or open a browser: one before the first browser automation in a session, another before a paid search. The browser gate once made me approve every single request, so now it asks once per session and gets out of the way. The reply-length hooks live at The counter.

Access badges Guardrails

Access badges decide which doors Claude can open. A few stay locked no matter who asks.

Guardrails are the rules for what Claude may touch. It never reads credential files, and it checks whether a secret exists without ever looking at the value. I asked for that early on, in four words: “Hard block ingesting credentials.”

Nothing sends, deletes or publishes without my OK. That one has been tested: Claude once published a page I hadn't approved, and my reply was in all caps.

Client work has its own badge. Some engagements are restricted, and nothing from them appears on any public surface, this page included. The cost is friction. I've also typed “Why do you keep asking for my permission?”

Knives and trucks CLIs and MCPs

Command-line tools are my knives, fast and always at hand. MCPs are delivery trucks for what knives can't reach.

CLI first. A CLI is a command-line tool Claude runs directly, like gws for Gmail, Calendar and Drive, or gh for GitHub. They're fast, and Claude still forgets it has them. I have typed, word for word, “YOU HAVE GWS CLI ARE YOU KIDDING ME”.

MCPs are the delivery trucks: connectors that bring an outside service into Claude. I keep them for what a knife can't do, like Notion, searching the vault by meaning, current library docs, or driving a real browser.

The rule is simplest tool first. I went through which connectors were loading into every session and cut the ones a CLI already covered, like the Gmail connector once gws could do the job.

Off the menu What I turned off

Things I tried, then turned off or skipped, and why.

An automatic sync that wrote to my vault on its own. I retired it in July, because nothing should write to memory while I'm not watching. Now only the capture step at the end of a session writes, and I approve what it proposes.

Three connectors are switched off: claude-flow for agent swarms, stitch for UI design, and one for n8n. So is a prompt-classifier hook that's waiting on a test. I evaluated agentmemory, an add-on memory system, and skipped it.

I also avoid one Codex model, gpt-5.6-sol, because it hangs. That's a little funny, since I used to insist on it by name. Before I cut anything, I want to be “very clear that we are not going to have a use for it later.”

The recipe book

Twenty skills I actually use. A skill is a written recipe Claude follows for one kind of job.

  • I ask for it
  • Claude reaches for it on its own
  • One mark per week I used it, June to September 2026
  • Built by me or credit to whoever wrote it

Closing the kitchen End of every session

Recipe card coming soon

Both of us

session-handoff

Built by me

My most-used skill by a mile. When a session gets bloated, I leave a prep list and start fresh.

What it does
Saves where a session stands to a handoff file, with exactly one next action, so a fresh session can pick it up.
When I reach for it
When the context is getting full, or before a big build that deserves a fresh window. Often I ask Claude first whether we should hand off or keep going.
Not for
Rescuing a session that lost the plot. A handoff only carries what's clear, so I make Claude state the goal in plain English first.
Goes well with
om-idea-capture rode along in 45 sessions. It also closes most of my codex-loop and codex sessions.

249 sessions on 59 days. Typed 104, Claude 210.

Recipe card coming soon

Both of us

end-session

Built by me

For the days I type “simple closeout please” and mean it.

What it does
Closes a session without a handoff file. It saves a raw summary and proposes anything worth keeping in long-term memory.
When I reach for it
When the work is done and there's no next session to set up. Often right after I've sent whatever we were working on.
Not for
Work I'm coming back to tomorrow. A fresh session picking up cold needs the detailed file that session-handoff writes.
Goes well with
codex shows up with it most, in 8 sessions. It shares the memory capture step with session-handoff.

30 sessions on 19 days. Typed 18, Claude 13.

Recipe card coming soon

Claude reaches for it

om-idea-capture

Built by me

I have never once typed this. It ran 52 times anyway, because the handoff calls it.

What it does
Saves content ideas that come up mid-session to my Notion ideas list, merging repeats so the list stays clean.
When I reach for it
I never ask for it by name. session-handoff runs it at close when a session turned up something worth writing about later.
Not for
Writing the post. It only files the idea. Drafting happens elsewhere, under my voice rules.
Goes well with
session-handoff (45 sessions together) and, now and then, grilling (5).

51 sessions on 27 days. Typed 0, Claude 52.

Reading the ticket Before I build anything

Recipe card coming soon

I ask for it

grilling

Borrowed from Matt Pocock

I usually ask to be grilled lightly. Then I type “keep grilling till we're totally aligned.”

What it does
Claude questions me about a plan, one decision at a time, and pushes back until my answers hold up.
When I reach for it
Before building anything with more than a couple of real decisions in it. Also at the start of a session, straight after a detailed handoff.
Not for
Re-deciding things we already settled. Those get a five-minute check at most.
Goes well with
session-handoff (15 sessions) and om-idea-capture. Then codex-loop, once there's an agreed plan to review.

27 sessions on 21 days. Typed 2, Claude 27.

Recipe card coming soon

Both of us

brainstorming

Borrowed from Jesse Vincent (Superpowers)

For the big open questions, like what on a sprawling roadmap actually deserves to go first.

What it does
Claude asks about intent and requirements first, then lays out a few directions before anything gets designed or built.
When I reach for it
When a project is at a fork, like deciding what to prioritize or how to move something from my laptop to a real hosted product.
Not for
Small, clear tasks. If I already know what I want, the open-ended questions mostly slow things down.
Goes well with
session-handoff (7 sessions together), and grilling when the ideas need stress-testing.

10 sessions on 10 days. Typed 0, Claude 10.

Recipe card coming soon

I ask for it

explain-like-im-12

Built by me

I've typed “plain English” in about 400 messages. Once I asked Claude to explain why it keeps failing to explain things.

What it does
Explains a technical idea with an everyday analogy first, then the real mechanism, with no jargon left undefined.
When I reach for it
When I'm lost and a short answer won't do. I ask for it by name now, since it kept firing whenever I said “plain English.”
Not for
Quick answers. When I ask for two plain English sentences, I want exactly two sentences and zero analogies.
Goes well with
explain-deck, when I want the analogy turned into slides. Claude also ran it next to codex in 4 sessions.

10 sessions on 10 days. Typed 1, Claude 9.

A clean station Build lean

Recipe card coming soon

Claude reaches for it

ponytail

Borrowed from Dietrich Gebert

I complain about over-engineering constantly. Now a rule makes Claude read this before it writes any code.

What it does
Pushes Claude toward the smallest change that works, using what's already there before adding anything new.
When I reach for it
Every coding task, automatically. It went from zero to 56 sessions in the 11 days after I asked for that rule.
Not for
Risky changes where the extra checks are the point. Minimal is my default, and I still ask for more care when the stakes are high.
Goes well with
codex (6 sessions) and test-first development (6). Also ui-ux-pro-max on front-end work (4).

56 sessions on 11 days. Typed 0, Claude 56.

Plating Design

Recipe card coming soon

I ask for it

impeccable

Borrowed from Paul Bakaus

The design skill I name most often, often with a hedge like “if you think genuinely helpful” attached.

What it does
Critiques and polishes a web interface, from layout and type down to empty states and error messages.
When I reach for it
When something I'm building looks like generic AI output, or a deck or page needs to look clean, modern and sleek.
Not for
Deciding what the page should say. It makes a layout better, and the words and structure still have to come from me.
Goes well with
ui-ux-pro-max and taste. I often name all three in one breath. Also session-handoff (5 sessions).

18 sessions on 12 days. Typed 0, Claude 18.

Recipe card coming soon

I ask for it

ui-ux-pro-max

Borrowed from Next Level Builder

Two weeks old in my kit, and I'm already writing “make sure that it uses the UI/UX Pro Max skill.”

What it does
A big reference library of interface styles, color palettes, font pairings and chart types that Claude checks before designing.
When I reach for it
When I want Claude choosing from real design options before it falls back on its usual defaults. This portfolio is one of those.
Not for
Judgment. It offers a huge menu of options, and picking the right one is still my job.
Goes well with
impeccable and taste, nearly always together. ponytail too, in 4 sessions.

6 sessions on 2 days. Typed 0, Claude 6.

Recipe card coming soon

I ask for it

taste

Borrowed from Leonxlnx

The anti-slop one. I've handed it to subagents with a one-line order: “use taste and impeccable with each agent.”

What it does
Pushes landing pages and portfolios away from templated AI looks, with a strict check before anything ships.
When I reach for it
When a landing page or portfolio needs to stop looking templated and start looking, in my words, “really cool and modern and sleek.”
Not for
Pages where the content is still unsettled. Polishing a layout before the words are right just means doing it twice.
Goes well with
impeccable, almost always. copywriting once, for case studies that needed design and words together.

6 sessions on 3 days. Typed 0, Claude 6.

A second taste Second opinions

Recipe card coming soon

I ask for it

codex

Built by me

My most-used second opinion. Claude plans, then a model from a different company tries to poke holes in it.

What it does
Hands a plan, bug or research question to OpenAI's Codex command-line tool and brings its answer back to Claude.
When I reach for it
Before I commit to a plan, and for deep research. I ask in words, like “can you get codex to weigh in on this as well?”
Not for
Accepting every note it gives. It gets nitpicky, so Claude decides what's a real gap and what's noise.
Goes well with
session-handoff (30 sessions) and end-session (8). codex-loop is the heavier version.

54 sessions on 31 days. Typed 0, Claude 54.

Recipe card coming soon

I ask for it

codex-loop

Built by me

My heavyweight build loop. I had it changed so Claude can argue back, because Codex kept nitpicking.

What it does
Claude plans, Codex reviews until they agree, then Codex builds while Claude checks. I approve at two gates.
When I reach for it
Long, multi-step builds where an early wrong turn gets expensive. “For more complex plans, use the codex loop” is basically my rule.
Not for
Small fixes. It's heavy on purpose, and even on big jobs I cap the review rounds so it doesn't over-engineer.
Goes well with
session-handoff (32 sessions together), and grilling first so there's an agreed plan to review.

44 sessions on 24 days. Typed 0, Claude 44.

Recipe card coming soon

Claude reaches for it

systematic-debugging

Borrowed from Jesse Vincent (Superpowers)

It fires on its own when something breaks, and it makes Claude do what I keep asking for anyway: root cause first.

What it does
Makes Claude find and prove the cause of a bug before it proposes any fix.
When I reach for it
Claude decides. It kicks in when a bug or a failing test shows up.
Not for
Work where “80% there and debug” is good enough. Sometimes I'd rather get close fast and fix what breaks.
Goes well with
codex, when a bug has several possible causes and I want an outside investigator on it.

8 sessions on 8 days. Typed 0, Claude 8.

Recipe card coming soon

I ask for it

agy

Built by me

A Google model on call for research. I use it rarely, mostly to run a parallel research pass next to Claude's.

What it does
Hands a research or build task to Google's Antigravity agent, which runs on Gemini, and brings back what it finds.
When I reach for it
Deep research where a different model might find different sources, like reading a vendor's terms or scouting alternatives.
Not for
Anything that needs my files or project history. It starts cold, so it's best on outside research.
Goes well with
codex, as another outside view. Sometimes a Sonnet subagent runs the same question alongside it.

4 sessions on 3 days. Typed 0, Claude 4.

Morning prep Daily work

Recipe card coming soon

I ask for it

morning2

Built by me

My morning briefing. It got heavy enough that I asked for “a fast version” that only gathers intel.

What it does
Sends agents out in parallel to read my email, calendar, tasks and news, then writes one briefing with suggested next steps.
When I reach for it
On mornings I want the full picture, inbox to news. Honestly not most mornings: 5 days in these logs.
Not for
A quick glance. It takes its time, which is why a lighter morning version exists.
Goes well with
email-sweep, when the inbox needs more than a skim.

5 sessions on 5 days. Typed 5, Claude 0.

Recipe card coming soon

I ask for it

email-sweep

Built by me

For when my inbox has clearly won. It sorts the pile and shows me the plan before touching anything.

What it does
Labels my unread Gmail, flags what's truly important, marks the noise read and starts drafts for the few that need replies.
When I reach for it
After a stretch of not keeping up. I type it myself, sometimes telling it to go back further than usual.
Not for
Sending anything. It drafts replies and moves labels, and it waits for my OK before doing either.
Goes well with
morning2, which covers the rest of the morning.

3 sessions on 3 days. Typed 2, Claude 1.

Recipe card coming soon

I ask for it

explain-deck

Built by me

Built in September. I wanted to learn big topics in layers, simple first, then deeper once I'm comfortable.

What it does
Turns one topic from my own files into a short offline deck, with a source on every slide and a closing quiz.
When I reach for it
When a project gets complicated enough that a two-sentence summary stops helping and I need the whole picture in layers.
Not for
Anything from the web. It only uses my local material, on purpose, so every slide can point to a source I have.
Goes well with
explain-like-im-12, when the whole deck should run on one big metaphor.

2 sessions on 2 days. Typed 0, Claude 3.

Recipe card coming soon

I ask for it

om-contract-generator

Built by me

The paperwork end of a project. It turns an accepted proposal into a statement of work.

What it does
Fills my standard statement of work template from a proposal, and adds a data processing appendix when personal data is involved.
When I reach for it
After a client says yes to a proposal and the work needs a statement of work before it starts.
Not for
Legal advice. It fills in my standard terms and points back to the master agreement. It does not decide new terms.
Goes well with
om-proposal-generator, which writes the proposal it reads from.

No runs in these logs, June to September.

Front of house Client-facing writing

Recipe card coming soon

I ask for it

copywriting

Borrowed from Corey Haines

Borrowed for web pages. Once I had to remind Claude: “You're supposed to be the copywriting expert here.”

What it does
Makes Claude gather context and write a short copy brief before drafting any landing page, headline or call to action.
When I reach for it
Page copy that has a job to do, like a case study or a portfolio section. I ask for it by name.
Not for
My own posts and opinions. Those go through my voice docs and have to start from things I actually said.
Goes well with
taste, for case studies that need design and words at once. My voice rules always run after.

4 sessions on 4 days. Typed 0, Claude 4.

Recipe card coming soon

I ask for it

om-proposal-generator

Built by me

Where a discovery call turns into a proposal with real numbers in it.

What it does
Takes discovery call notes and prospect research, scopes the project, and drafts a formal proposal with pricing and expected return.
When I reach for it
After a discovery call, once I know what the client needs and roughly what it's worth to them.
Not for
Sending it. The proposal is a draft that I read, edit and send myself.
Goes well with
om-contract-generator, once the client says yes.

No runs in these logs, June to September.

Also in the kit Used less, still earning shelf space

  • cavemanMakes Claude talk like a caveman to save words. I've wondered if I should use it more often.Julius Brussee
  • find-skillsSearches the public skills directory for a skill that already does what I'm about to build.Vercel
  • watchLets Claude actually watch a video link, frames and transcript both. My rebuild of someone else's skill.Rebuilt from bradautomates (claude-video)
  • video-clipperCuts a long talking-head video into a website cut and short clips, with my approval before rendering.Built by me
  • notebooklmDrives Google NotebookLM from Claude, so research I've gathered can become explainers.teng-lin
  • gemini-image-generatorMakes images with Google's Gemini, mostly reference images and graphics for slides.Credit to confirm
  • frontend-slidesBuilds HTML slide decks. I like mine light on text, since I'd rather talk over them.Zara Zhang
  • double-reviewTwo separate reviewers, Codex and a Claude subagent, check the same change, then Claude merges their verdicts.Credit to confirm

Used together Sessions where both ran

  1. session-handoffom-idea-capture45
  2. session-handoffcodex-loop32
  3. session-handoffcodex30
  4. session-handoffgrilling15
  5. codexend-session8
  6. session-handoffbrainstorming7
  7. codexponytail6
  8. grillingom-idea-capture5
  9. session-handoffimpeccable5
  10. ponytailui-ux-pro-max4
  11. codexexplain-like-im-124