eStack: explainx.ai Open-Sources Its Agent Skill Stack for Learning, Building and Growth

Over 2026, three personal skill stacks became the way many developers set up their coding agents. Garry Tan's gstack passed 138,000 GitHub stars. Lauren Tan's pstack shipped inside Cursor's plugin repository with 34 skills and 24 principles. Theo Browne's tstack packed the whole pull request lifecycle into 9 skills. Today we are releasing our own: eStack, 14 MIT-licensed agent skills for Claude Code and Codex.

eStack is not another engineering stack. The other three make the agent a better engineer. eStack also makes you better, with skills that explain and quiz you on what the agent built. And it covers the work after the code: search, blog posts, social posts and short video. These are the workflows we use at explainx.ai to run a learning platform for 350K+ students, with agents doing most of the daily work. The source is on GitHub.

TL;DR: questions answered

What gstack, pstack and tstack actually are

Before we explain eStack, here is what the three stacks it is compared to contain. All three are free and MIT-licensed. None of them sell anything. They spread because each one captures how a respected builder works.

gstack (Garry Tan, Y Combinator). gstack turns Claude Code into a team of roles: CEO, designer, engineering manager, QA lead and release manager. Its README describes it as "a process, not a collection of tools", in the order think, plan, build, review, test, ship and reflect. The current README lists more than 50 slash commands, including /office-hours, /plan-ceo-review, /review, /qa, /ship, /land-and-deploy, /retro and safety skills such as /careful and /freeze. We covered it in depth in our gstack guide and in the story behind Garry Tan's 400x productivity claim.

pstack (Lauren Tan, known as poteto). pstack is the opposite of "move fast". Its idea is that slow, deep work lets you trust agents enough to run several at once. A router skill, /poteto-mode, matches each task to one of 23 playbooks (bug fix, perf, refactor, shipping and more). Around it sit 34 skills such as /architect, /interrogate (several models try to break a diff) and /teach, plus 24 short principle skills like "prove it works" and "subtract before you add". It started as a Cursor plugin, and the community ported it to Claude Code and Codex.

tstack (Theo Browne). tstack is the smallest and most focused. Its 9 skills automate the pull request lifecycle: full-send takes a change all the way to a merged PR, council gets three models from different providers to review until they agree, babysit-pr watches CI and review bots, and takeover finishes someone else's PR while keeping their credit. You install it by copying folders into your skills directory.

Why we built a different kind of stack

All three stacks answer one question: how do I get better code out of an agent? That is a good question, and gstack answers it very well. We saw no reason to build a fourth answer to it.

We had two other problems, and they are common among the people in our workshops.

The first problem is understanding. Agents now write code faster than people can read it. A founder ships a feature on Monday and cannot explain it on Friday. A new developer accepts a 400-line diff they could not have written. The code works, but nobody owns it. When it breaks, the person on call is the agent.

The second problem is distribution. Building is cheaper than ever, so more products compete for the same attention. The work that decides whether anyone finds your product (search, posts, video) is still mostly manual for most small teams. At explainx.ai we moved most of that work to agents over the last year, and we wrote down the rules that made it work.

So eStack is built around three stages, and two of them are the ones the other stacks skip.

The 14 skills

eStack agent skills arranged as a tree of learn, build and grow branches

Start with /estack. The router restates your goal in one sentence, picks the stage and the skill, and tells you why in one line. If no skill fits, it does the task normally instead of forcing one.

Skills chain where it helps. Run premortem before ship on anything costly to undo. Run blog, then social, then short to turn one piece of research into a post, a thread and a video script. Run explain, then quiz-me, when you need to own code you did not write.

Learn: the part no other stack has

quiz-me skill asking a multiple choice question about code an agent wrote

explainx.ai is an education company first, so this is the section we care about most.

explain writes in layers and stops at the layer that answers your question. Layer one is a single plain sentence. Layer two is why the code exists. Layer three is how it works, often with a small ASCII diagram. Layer four is the one tricky part, with the exact file:line. It uses your real names and code, not foo and bar, and it says "I checked" or "I did not check" instead of guessing.

quiz-me is the skill we expect people to talk about. After a large change, it picks the 3 to 5 ideas you must know to maintain it and asks about them one at a time: predict what happens on a 429 error, spot the bug in a changed line, explain in one sentence why the cache runs before the auth check. It grades honestly. A wrong answer gets the right answer and a pointer to the code, not praise. It ends with a score and the one area to read again. It is the same idea as the quizzes in our learning app, applied to your own codebase.

catch-up fixes the problem with AI news: almost none of it matters to your project. It first reads your dependencies, model names and agent setup and writes a one-line profile. Then it pulls the week's news and keeps only items that change a price, a limit, a deprecation date, a security issue or a better option for something you use. Zero items is a valid answer. With the explainx.ai MCP server connected, it uses our live news and trending data. Without it, it falls back to web search and says so.

pick-model finds every place your code calls a model, groups the calls by task, and compares current prices instead of prices from memory. Bulk work like tagging and extraction goes to the smallest model that passes a 10-sample check. Code that ships and security review stay on the strongest model. It shows the monthly cost before and after, and it changes nothing until you agree. If you have read our coverage of model routing cutting AI coding costs, this is that idea as a skill.

Build: small on purpose

The build section has four skills, not forty. gstack and pstack already cover engineering in depth, and you can install them next to eStack. Our four are the ones a small team needs every day.

• premortem starts from "it is 30 days later and this failed badly. Why?" It lists 8 to 12 causes across data, users, money, timing, dependencies, people and security, checks the real code for each suspected one, and turns the top risks into a prevent, detect or recover action. It ends with a go, go-with-fixes or stop call. Our earlier post on the premortem agent skill explains why starting from failure works.
• ux-review asks one question of every label and message: is this for the user or for the engineer? It removes leaked internals ("Calling API...", model names, error codes) and rewrites buttons, status messages and errors in plain words.
• verify-ui exists because a clean typecheck does not prove a page looks right. It takes screenshots at 390 px and 1280 px wide, in light and dark mode, reads every one, checks the console, fixes what it finds and repeats up to three rounds.
• ship stages files by name, never git add -A, so another agent's half-finished work in the same tree does not ride along. It confirms that each check actually ran (a test runner that found zero tests is not a pass), writes a short commit message, and touches the changelog only for changes a user would notice.

Grow: the work after the code

seo-geo and blog skills growing a site from a stack of pages

These four skills come from the content systems that run explainx.ai and our other sites every day.

seo-geo writes or audits a page for two readers: a search engine that ranks it and an AI engine that quotes it. It puts a 40 to 60 word answer right under the title, uses question-style headings, backs every number with a primary source, adds Article, FAQPage and breadcrumb schema, and checks that robots.txt does not block AI crawlers the owner wants. For the theory, see our guide to the SEO and GEO agent skill.

blog reads 3 to 5 of your existing posts first, so the new one matches your voice and format. It goes to the primary source, not a summary, separates what a vendor claims from what someone verified, reports practitioner discussion as "users report", and adds links both ways: from the new post to old ones, and from old posts to the new one. It runs your site's own validation before it hands off.

social is built on one rule: if nobody would disagree with a post, it is not a point yet. It names the subject in the first line, takes a side on ideas, products and company decisions (never on people), keeps links in the first reply, and scores each draft out of 100 before you see it. Drafts under 55 are not offered. Zero drafts is a valid answer.

short writes vertical video scripts. It rejects topics whose payoff needs the mechanism explained first, puts the most surprising fact in the first two seconds, keeps one idea per beat with a visual for each, and gives three title options with the named product or company up front.

ste100: non-slop writing everywhere

The last skill applies to every stage. ste100 uses the countable rules of ASD-STE100 Simplified Technical English, the aerospace writing standard whose Issue 9 came out in January 2025: procedure sentences of 20 words or fewer, descriptive sentences of 25 or fewer, paragraphs of 6 sentences or fewer, active voice, no "-ing" verb forms, no more than 3 nouns in a row, and warnings before the step they apply to. Its core rule is that it changes the words, never the substance. No fact, number or caveat is cut to make text shorter, and a "usually" stays a "usually".

It has two modes. Practical mode, the default, applies the countable rules with looser vocabulary, which is what Andrej Karpathy describes as asking for text "80% of the way" to STE. Strict mode is for procedures and safety warnings: it uses each word in one meaning and one part of speech, and it flags every word it is not sure the official dictionary approves. The skill never claims that text is STE compliant, because the agent does not have the official dictionary.

On top of STE it runs a slop check built from Wikipedia's editor guide to signs of AI writing: 13 patterns, each with a fix. It covers empty openers like "Great question!", inflated significance ("stands as a testament"), stock AI vocabulary ("delve", "tapestry", "intricate"), "serves as" in place of "is", "not X, but Y" frames, trailing "-ing" clauses that add fake analysis, hedge stacks, em dash chains and vague sources like "studies show". A small Python linter flags long sentences, hedges, passives, semicolons and slop words for review. The blog and social skills run the slop check before they finish. If you want the background, read our guide on using ASD-STE100 with AI.

How to install eStack

As a Claude Code plugin:

/plugin marketplace add whyashthakker/estack
/plugin install estack@estack

With the skills CLI:

npx skills add whyashthakker/estack

By copying the folders (Claude Code or Codex):

git clone https://github.com/whyashthakker/estack
cp -R estack/skills/* ~/.claude/skills/   # Claude Code
cp -R estack/skills/* ~/.codex/skills/    # Codex

To use eStack in one project only, copy the folders into that project's .claude/skills/ folder instead.

Optional live data. catch-up and pick-model work best with the explainx.ai MCP server, which serves current AI news, model price comparisons, trending topics and a directory of agents, tools and MCP servers:

claude mcp add --transport http explainx https://mcp.explainx.ai/mcp

You sign in with your explainx.ai account the first time a tool runs. Setup for ChatGPT, Codex and other clients is on the MCP page. If you are new to MCP, start with what MCP is.

What people are asking

Should I replace gstack with eStack?

No. They do different jobs. If gstack's planning and QA skills work for you, keep them. Add eStack's Learn and Grow skills, and delete whichever ship or review skill you use less, so the agent does not choose between near-duplicates. Our post on CLAUDE.md, SKILL.md and MCP explains how these layers fit together.

Do the skills cost tokens when I do not use them?

Very little. Agents load only each skill's name and description at the start. The full SKILL.md loads only when the skill is used. 14 short descriptions are a small cost next to a typical CLAUDE.md.

Are the Grow skills only for explainx.ai?

No. We removed everything specific to our sites: posting times, internal tools, bot rules and paths. The skills read your site first (your posts, your folder layout, your validation scripts) and follow it. Put your own settings, such as your voice and your posting schedule, in CLAUDE.md or AGENTS.md, so you can update eStack without losing them.

Will social post for me?

No. Every skill that publishes, posts, pushes, sends or pays asks first. social and short hand you drafts with scores. You choose and post.

Why is the build section so small?

Because the strongest engineering stacks already exist and are free. A fourth set of review and QA skills would add overlap, not value. We kept the four skills that small teams told us they reach for daily.

Limits to know before you install

• It is version 0.1. The skills come from workflows we run every day, but the public versions are new. Expect wording changes as people report what works.
• The Grow skills need your judgment. social can score a draft. It cannot know your audience as well as you do. Treat scores as a filter, not a verdict.
• Live data needs an account. catch-up and pick-model are weaker on web search alone, because general search results do not show your project's prices and deprecations.
• verify-ui needs a browser tool. It uses Playwright, a browser automation tool, or a headless script, whichever your session has. Without one, it can only tell you what to check.

Learn the workflows live

The skills are free. The judgment behind them is what we teach. Our live workshops cover agent workflows, skills and shipping with Claude Code and Codex, and members get the recordings. If you want the full picture of how skills work before you install anything, start with our agent skills guide or see how the Superpowers framework approaches the same problem from the engineering side.

Related reading

• eStack: install and skill list
• What are agent skills? Complete guide
• gstack: Garry Tan's Claude Code skills factory
• Superpowers: the skills framework that makes coding agents plan, test and review
• Matt Pocock's agent skills for real engineers
• CLAUDE.md vs SKILL.md vs MCP
• The premortem agent skill
• How to use ASD-STE100 with AI

Skill counts and star counts for gstack, pstack and tstack come from each project's README on November 1, 2026, and they change often. eStack details match version 0.1 of the eStack repository.