# Skill Cauldron: Brew a Claude Skill That Works Like You Do Canonical: https://www.kartikaydhar.com/blog/skill-cauldron-brew-personal-claude-skills Published: 2026-10-25T10:34:06.077635+00:00 Updated: 2026-10-25T10:34:06.077635+00:00 Category: Technical I built Skill Cauldron to turn your work habits, installed skills, and skills.sh discoveries into one focused Claude skill—with a blind taste test and a reusable recipe card. We have more ways than ever to customize an AI assistant: prompts, MCP servers, plugins, and now agent skills. Yet there is a strange irony. We often install the same popular skills as everyone else, then spend our time correcting the output to match how we actually work. That's the idea behind Skill Cauldron, an open-source Claude skill I published on October 24, 2026. Rather than installing another generic workflow, it helps you brew a focused skill from your own habits, existing skills, and relevant skills discovered on skills.sh. It gives you a way to inspect the ingredients, test the result, and package something reusable. ## Why I built Skill Cauldron A prompt tells an assistant what you want right now. A reusable skill should encode how to do a particular job consistently: which steps to follow, what to leave out, what good output looks like, and where your preferences override generic advice. But a skill downloaded from the internet cannot know that you prefer one-page product briefs, always want sources next to claims, or consistently ask for a runnable example instead of an abstract explanation. Those working habits are often the missing ingredient. Skill Cauldron starts there. Its guiding principle is simple: the job you want done comes first, your working habits come second, and borrowed techniques come third. It is not a tool for blindly concatenating prompt files. Image: Skill Cauldron interface inside Claude showing the ingredient picker and brewed skill recipe card ## Install Skill Cauldron The quickest route is the agent skills installer. Run this in your terminal: npx skills add kaidhar/skill-cauldron Or manually copy the repository's skills/skill-cauldron folder into your skills directory. For Claude Code, that is typically ~/.claude/skills/. The project is MIT licensed. After installation, open Claude in an environment where the skill is available and ask: Brew me a skill for my weekly product and AI updates. You do not have to memorize a CLI for the brewing process. The skill guides the interaction in Claude. Its visual ingredient picker appears inline when a supported HTML widget is available; in other environments, it can provide an HTML artifact to open separately. ## No Terminal? Install Skill Cauldron from Chat You don't have to be a developer to get started. If you use Claude Desktop or ChatGPT Desktop, copy the prompt below into a new chat. It asks the assistant to inspect the GitHub repository, check what installation options your app actually supports, and help you prepare the skill if a manual upload is needed. I want to install Skill Cauldron as a reusable AI skill. GitHub repository: https://github.com/kaidhar/skill-cauldron Skill folder: skills/skill-cauldron Please: 1. Read the README, SKILL.md, and referenced files to understand the skill. 2. Check whether this app and my account support installing custom skills. 3. If installation is supported and you have the required tools, help me install the complete skill, preserving its folder structure. If a ZIP upload is required, prepare it and show me where to upload it. 4. If installation is not supported, explain the limitation and offer to use the repository instructions in this conversation instead. Do not claim the skill is installed unless it really is. 5. Do not execute unreviewed repository scripts or share my personal conversation history. Once the skill is available, help me begin with: "What kind of skill would you like to brew?" For Claude Desktop: If Claude prepares a skill ZIP but cannot install it itself, open Customize → Skills → + → Create skill → Upload a skill, upload the ZIP containing the complete skill-cauldron folder, and enable it. Code execution and file creation may need to be enabled. Claude's official instructions explain the current steps. For ChatGPT Desktop: On eligible accounts, open Plugins → Skills → Create → Upload from your computer, or use the skill-creation flow in chat. At the time of writing, ChatGPT Skills are generally available for Business, Enterprise, Healthcare, and Edu accounts, subject to workspace settings, and availability may differ by app. Check ChatGPT's official Skills guide for current eligibility and installation. Important: Pasting the prompt is not the same as installing a persistent skill. If the assistant cannot install skills, it should say so. It may still help you explore Skill Cauldron in the current chat when it has access to the complete repository, but the skill will not automatically carry into future conversations. ## Step 1: Gather three kinds of ingredients Skill Cauldron builds its menu from three different sources: - Your working habits: recurring tasks, explicit preferences, and repeated corrections in conversation history the assistant can actually access. It should not invent personality traits or include sensitive details by default. - Your installed skills: local or already installed skills that overlap the job, with special attention to your own customized versions. - Skills from skills.sh: relevant skills discovered from the live directory, along with available install counts and security-audit signals. Skills with failed audits are excluded by the workflow. If Claude cannot access earlier conversations, the skill asks a few short questions instead of pretending it remembers you. That's a feature, not a limitation to hide. ## Step 2: Mix, but do not make a kitchen-sink skill In the cauldron interface, you select the ingredients worth combining. The experience nudges you toward about three and warns when you pile on more. The point is to select useful techniques, not dump entire instruction sets into one giant prompt. Image: Animated Skill Cauldron brew showing selected ingredients entering the cauldron and a recipe card appearing For example, imagine I want a weekly roundup of new AI tools that I can immediately evaluate as a product manager. My ingredients might be a personal preference for practical takeaways, an installed summarization skill, and a vetted research-oriented skill discovered through skills.sh. Those are example choices, not a claim that the tool picks the same skills for everyone. ## Step 3: Brew one skill with one job Before writing anything, Skill Cauldron asks you to confirm a single job sentence. For this example: Turn new AI and GitHub developments into a five-minute weekly product briefing with primary sources, practical implications, and one idea worth testing. The skill then reads the selected ingredient materials and writes two files: - SKILL.md — the concise, runnable instructions and activation description. - RECIPE.md — the provenance: which ingredients were used, where they came from, what was retained, and what was deliberately discarded. The distinction matters. RECIPE.md is an audit trail for the skill's design. If a technique doesn't serve the approved job, the brewing rules say to leave it out—even if it came from a popular skill. ## Step 4: Taste-test the skill before trusting it A good-looking SKILL.md is not evidence that it improves work. Skill Cauldron therefore asks for real previous tasks and prepares two versions of each: one following the brewed skill and another without it. You compare the outputs as anonymous A/B options and decide which is better. That is a human preference check, not an automatic scientific benchmark. If the outputs tie, lose, or cannot be tested, the recipe card should say so. It must not manufacture a win rate or treat the assistant's own opinion as your vote. For a stronger test, reuse the same set of tasks over several iterations and track things that matter to you: number of factual corrections, number of follow-up instructions, time-to-usable-output, and whether sources are correct. ## Step 5: Bottle and reuse Once the files pass the built-in checks, the repository's packaging script creates a .skill archive. For example, within a working directory containing your brew: python skills/skill-cauldron/scripts/bottle.py brews/weekly-ai-brief Replace the script path with the actual location of your installed Skill Cauldron folder. The command above illustrates the packaging step from the repository, not a command you must run to start brewing. You also get a shareable recipe card that records the ingredients and the taste-test outcome. With the right browser tooling available, the project can render the card as a PNG; otherwise, an HTML card works too. ## The research behind the three-ingredient idea The motivation is informed by SkillsBench, research evaluating how agent skills affect task outcomes. Its early experiments found that curated, task-relevant skills helped more than having an assistant generate generic procedural guidance from scratch, and that adding too many skills led to diminishing returns. The research has evolved across versions, so the exact figures depend on the evaluation release. Importantly, those are benchmark results about agent skills generally—not measured performance improvements for Skill Cauldron. Selecting three ingredients is a practical design heuristic. Your own blind comparison is how you decide whether a specific brew helps. ## What Skill Cauldron does well—and where to be careful What interests me most is that it treats customization as a repeatable workflow: gather evidence of a user's preferences, inspect source material, compose a focused procedure, record provenance, and compare it with a baseline. There are also real constraints: - Conversation-based habits depend on what history the current assistant can access. No access means you must provide preferences yourself. - The visual interface depends on the host environment. Not every Claude surface can display an interactive inline cauldron. - Live skills.sh discovery depends on network access and third-party metadata. Audit signals are useful filters, not guarantees of safety; review skills and dependencies before use. - The blind taste test is qualitative and can be biased. Treat it as a useful first filter rather than a substitute for repeatable evaluations. - Shared recipe cards can disclose habits or source names. Review them for personal or client information before sharing. ## Try this first My suggested starting point is not a grand, general-purpose “personal AI” skill. Pick a task you repeat every week and already know how to judge. I want to brew one skill for my weekly AI product update. Use my stated preferences only, show me the ingredients, and keep the final skill narrowly focused. After brewing, compare it against my normal workflow on two real examples. That prompt gives you a clear job, a way to avoid fabricated personalization, and an evaluation criterion. If the new workflow doesn't improve the output, change the ingredients and brew again. Repository: github.com/kaidhar/skill-cauldronInstall: npx skills add kaidhar/skill-cauldronLicense: MIT Everyone can install the same skills. The interesting part starts when you build one that actually reflects the way you work.