From the lecture "Using agents to deal with high dimensions" by Kareem Hegazy and Maksim Melnichenko (UC Berkeley, Stat 157, Fall 2026).
Download everything as a zip: stat157-agent-skills-main.zip
Slides of the lecture: slides/2026-10-15-using-agents-high-dimensions.pdf
A profile tells Claude who you are and how you want it to work, once, instead of in every chat. In a fresh chat, paste:
Help me build a persistent context profile. Ask ONE question at a time about: my role and courses; my code stack; my accuracy standards; preferred output format; what you should NEVER do; standing rules. Validate each answer. End with a complete Markdown profile in my exact words, and save it as ~/.claude/CLAUDE.md; if that file already exists, show me the changes and ask before overwriting it. If you cannot write files, tell me where to paste the profile instead.
Where the profile goes:
- Claude Code:
~/.claude/CLAUDE.mdfor all your projects, orCLAUDE.mdin one project folder. - claude.ai: the personal instructions in Settings; they apply to every chat.
- Codex:
~/.codex/AGENTS.md, orAGENTS.mdin one project folder.
Keep it under a page, and once a month ask Claude to review it against your recent chats. More in prompts/profile-interview.md. The prompt comes from Kareem Hegazy's workshop "Working with Claude: a practical guide for researchers" (ICSI, June 2026).
slides/: the lecture slides (PDF).prompts/: the profile-interview prompt.skills/: the three skills, one folder each, for Claude Code (also Codex and Gemini CLI).claude-ai-uploads/: the same three skills zipped for claude.ai (Claude in the browser), which accepts a skill only as a zip upload. Its copies of answer-key and office-hours say in their descriptions when to run, because claude.ai does not accept the "run only when typed" setting.
| Skill | What it does | Claude Code | claude.ai |
|---|---|---|---|
| sci-plot | House style for scientific figures: the exact formula in red on the simulation in blue, readable at the size it will be shown, no text on the data. Claude uses it whenever it plots. | SKILL.md | sci-plot.zip |
| answer-key | Complete worked answers at once, like the solutions at the back of the book. | SKILL.md | answer-key.zip |
| office-hours | A tutor who asks questions, gives one hint at a time, and never hands over the answer. | SKILL.md | office-hours.zip |
Read a skill before you install it: it is someone else's instructions running in your agent with
your permissions. Each one is a single short SKILL.md file.
Free add-ons for Claude Code, by category. Plugins install from inside Claude Code with
/plugin install <name>@claude-plugins-official (the official marketplace is preloaded). Read what
you install: a plugin or skill runs with your permissions.
| Name | Kind | What it does | Install |
|---|---|---|---|
| Correctness signal | |||
| clangd-lsp | plugin, first-party | C/C++ diagnostics on every edit; find references | /plugin install clangd-lsp@claude-plugins-official |
| pyright-lsp | plugin, first-party | Same for Python (type and import errors) | /plugin install pyright-lsp@claude-plugins-official |
| context7 | MCP server, community | Current, version-specific library docs | /plugin install context7@claude-plugins-official |
| Discipline, review, cleanup | |||
| superpowers | plugin, community | Spec, failing test, implement, debug | /plugin install superpowers@claude-plugins-official |
| code-review | plugin, first-party | Multi-agent PR review, confidence-scored | /plugin install code-review@claude-plugins-official |
| code-simplifier | plugin, first-party | Behavior-preserving cleanup of new code | /plugin install code-simplifier@claude-plugins-official |
| security-guidance | plugin, first-party | Flags injection, unsafe deserialization, leaked secrets | /plugin install security-guidance@claude-plugins-official |
| Tokens and cost | |||
| codeburn | CLI tool, open source | Token and cost meter for 31 tools, local only | npx codeburn in a terminal |
| caveman | skill, community | Terse output, about two-thirds fewer tokens | /plugin marketplace add JuliusBrussee/caveman, then /plugin install caveman@caveman |
| Workflow | |||
| commit-commands | plugin, first-party | /commit, push, full PR in one bundle |
/plugin install commit-commands@claude-plugins-official |
| skill-creator | plugin, first-party | Scaffold, evaluate and audit your own skills | /plugin install skill-creator@claude-plugins-official |
The catch: of 3,984 public skills that Snyk audited in 2026, 36.8% had a flaw and 76 carried malware (ToxicSkills).
Copy the three folders in skills/ into .claude/skills/ inside a project folder (that
project only) or into ~/.claude/skills/ (all your projects). If Claude Code is running, restart it,
then type /skills to check that they are listed.
- sci-plot loads by itself when Claude draws a figure.
- answer-key and office-hours run only when you type them, each in a new session:
/answer-key <your question>or/office-hours <your question>.
- Settings > Capabilities: turn on "Code execution and file creation". Skills need it.
- Customize > Skills > + > Create skill > Upload a skill, once for each zip in
claude-ai-uploads/. - Start a new chat with "Answer key: " or "Office hours: ". sci-plot is used whenever Claude makes a figure.
claude.ai does not accept the setting that makes a skill run only when you type it, so the claude.ai copies of answer-key and office-hours say in their descriptions when to use them.
The same SKILL.md files work there. Put the folders from skills/ in .agents/skills/
inside your project folder.