Stuff you don't need on day one but will want by week one — VS Code layout, recommended plugins, multi-model second opinions, custom skills, project workflows, cost guidance. Read top to bottom or jump to whatever you need.
Agnes is terminal-driven. You spend most of your time talking to Claude in a terminal — not in a web app. VS Code's split-terminal layout gives you one panel for the conversation, one for diffs and tool output, and an optional third for parallel work. Files open as tabs when you need them.
~/Agnes: File → Open Folder… → select ~/Agnes.claude (or yolo if you set that alias).~/Agnes workspace anatomyYour workspace was created by agnes init and the bundled Agnes plugin. Here's what each folder is for.
~/Agnes/ ├── .claude/ │ ├── skills/ # Reusable workflow templates Claude follows │ ├── rules/ # Behavioral rules injected into every message │ ├── hooks/ # Scripts that fire on events (session start, etc.) │ └── settings.local.json # Permissions allowlist (your safety net) ├── docs/ │ └── plans/ # Timestamped design docs and plans ├── projects/ │ ├── work/ # Company projects │ ├── personal/ # Personal projects │ └── learning/ # Experiments and tutorials ├── tasks/ # Task lists, lessons learned ├── temp/ # Scratch files, downloads ├── logs/ # Action logs (auto-generated) └── CLAUDE.md # Master instructions Claude reads every session
The CLAUDE.md at the root is yours to customize. Add preferences, conventions, rules — Claude reads it at the start of every session. Tell Claude: "Add a rule to CLAUDE.md that we always use metric units."
Every piece of work lives in a project folder. Always start Claude from the ~/Agnes root — the welcome menu and prompt-history logging both depend on it.
~/Agnes.claude (or yolo).Just tell Claude in plain English:
cds and loads project context.~/Agnes.The agnes-welcome hook fires only when Claude starts in ~/Agnes. Prompt-history logging uses $PWD to route prompts to the right project's history. Launch Claude from inside a project folder and you skip both.
Plugins extend Claude's capabilities. Install only what you need. Browse the operator-curated set at /marketplace for the marketplace tied to this Agnes instance.
Inside any Claude Code session, run the slash command shown in the row's Install column — e.g. /plugin install superpowers@every-marketplace. Click Copy, switch to Claude Code, paste, hit Enter. If the marketplace isn't registered yet, Claude will tell you and you can add it with /plugin marketplace add <url> first. Or paste the catch-all prompt at the end of this section and let Claude do the registration + install in one go.
| Plugin | What it does | Install |
|---|---|---|
| superpowers | The workflow engine — brainstorming, planning, code review, parallel agents, skill creation. Most important plugin. | /plugin install superpowers@every-marketplace |
| context7 | Fetches current library/framework docs. Bypasses Claude's knowledge cutoff. | /plugin install context7@anthropic |
| github | Read/write GitHub repos, PRs, issues. If you use GitHub for anything. | /plugin install github@anthropic |
| Plugin | What it does | Install |
|---|---|---|
| playwright | Browser automation — screenshots, form fill, web scraping, end-to-end tests. | /plugin install playwright@anthropic |
| atlassian | Read/write Jira issues + Confluence pages. | /plugin install atlassian@anthropic |
| code-simplifier | Reviews + cleans up code for clarity and consistency. | /plugin install code-simplifier@every-marketplace |
| Plugin | What it does | Install |
|---|---|---|
| figma | Read Figma designs, generate code from them. | /plugin install figma@anthropic |
| playground | Create interactive HTML tools and explorers. | /plugin install playground@every-marketplace |
| ralph-loop | Iterative refinement loop — runs after implementation to catch what the first pass missed. | /plugin install ralph-loop@every-marketplace |
Paste this prompt instead — Claude detects which marketplaces are registered, registers any missing ones, and installs in the right order:
Install the superpowers, context7, and github plugins for me. If they're not in any of my registered marketplaces, find the canonical source on GitHub for each one (look up the marketplace.json), register the marketplace via `/plugin marketplace add <url>`, then run `/plugin install` for each. Verify with `/plugin list` and tell me which ones are now active.
Send your work to multiple AI models for independent review. They catch blind spots, challenge assumptions, improve quality. One of the most useful workflows you can run inside Agnes.
Paste this into Claude Code:
Install OpenAI's Codex CLI for me. Walk me through it: 1. `npm install -g @openai/codex` (run via Bash tool — npm prefix is under nvm/fnm so no sudo). 2. Symlink the binary into ~/.local/bin so non-interactive subshells find it: `ln -sf "$(command -v codex)" ~/.local/bin/codex`. 3. Authentication — pick whichever I prefer: - ChatGPT login (easiest, Plus/Pro included): run `codex` and select "Sign in with ChatGPT". - API key path: have me create a key at platform.openai.com, then store it in ~/.config/agnes/openai.env via a `read -srp` helper script (no chat exposure). 4. Verify: `codex --version`.
Paste this into Claude Code:
Install Google's Gemini CLI for me. Walk me through it: 1. `npm install -g @google/gemini-cli`. Symlink to ~/.local/bin same as Codex. 2. Get a free API key at https://aistudio.google.com/app/apikey — open the page in my browser via `open` / `xdg-open` / `Start-Process`. 3. Tell me to copy the key. Prepare a `read -srp` helper script at ~/.claude/agnes/bin/store-gemini.sh that pipes my pasted key into the OS keychain (macOS: `security add-generic-password -U -s 'agnes-gemini-key' -a "$USER" -w "$t"`; Linux: secret-tool; Windows: cmdkey). Tell me to run `bash ~/.claude/agnes/bin/store-gemini.sh` in my terminal. 4. Verify: `gemini --version` and a low-cost test prompt.
The bundled Agnes plugin ships two skills:
Trigger them by saying "Get a second opinion on this plan". Claude runs review-panel first, then second-opinion across all available external models.
The mechanics that turn Claude from a blank slate into a workspace that follows your conventions. Most of Agnes's day-to-day power lives here.
Reusable workflow templates at .claude/skills/<name>/SKILL.md. Invoked by name or automatically.
| Skill | Fires when |
|---|---|
| brainstorming | Before any creative or design work — explores options before committing. |
| writing-plans | Before any multi-step task — produces a timestamped plan in docs/plans/. |
| executing-plans | Step-by-step plan execution with checkpoints. |
| systematic-debugging | Before any bug fix — narrows down the failure before changing anything. |
| verification-before-completion | Before claiming "done" — runs the actual checks. |
| second-opinion | After any plan or analysis — external model review (see §5). |
| ralph-loop | After implementation — iterative refinement to catch the first pass's misses. |
Create your own: "Create a skill called weekly-digest that searches Gmail for important emails from the past 7 days, categorizes them by urgency, creates an HTML report, and lists action items with deadlines." Claude writes the skill file. Then "Run the weekly digest" follows the recipe.
Files in .claude/rules/ injected into every Claude message. Stronger than CLAUDE.md prose — Claude can't miss them. The most important is skill-enforcement: a decision matrix that maps task types to required skills (entry gates: brainstorming for creative work, systematic-debugging for bugs; exit gates: second-opinion after plans, verification-before-completion before "done"). Without this rule, Claude optimizes for speed and skips the skills.
Scripts in .claude/hooks/ that fire automatically on events:
agnes pull --quiet so data is fresh, fires the agnes-welcome project picker.agnes push --quiet so session jsonl + CLAUDE.local.md flow back to the server.Pick whichever you can run today.
Paste this into Claude Code at the ~/Agnes root:
Create a new project called "Q4 strategy review" under work. Then enter plan mode and draft a 6-week project plan covering: stakeholder interviews, current-state analysis, gap identification, recommendations, and stakeholder review. Save the plan to docs/plans/.
Once GWS is set up (see /home § "connect your tools"), paste:
Search my Gmail for unread messages from the past 7 days. Extract anything that looks like a task or decision request, categorize by urgency (urgent / important / fyi), and produce an HTML dashboard at temp/email-digest.html with a one-line summary and link back to each message. Then tell me how many minutes it should take to clear urgent.
After Claude produces any plan or analysis:
Get a second opinion on this. Run the review-panel first (4 internal personas), then send the result to all available external models (Codex, Gemini). Show me where they agree (high confidence) and where they disagree (needs my judgment).
Auto-accept mode (recommended on /home) auto-approves edits but still gates Bash. YOLO mode goes further — auto-approves everything. Faster, broader blast radius.
~/.claude/settings.local.json first. The permissions file is the actual safety net — YOLO mode + reviewed allowlist = fast AND safe. YOLO mode + missing allowlist = unbounded shell execution. Read the file before flipping the switch.
Run with the flag once: claude --dangerously-skip-permissions. Useful for trusted batch operations.
macOS / Linux (zsh):
echo 'alias yolo="claude --dangerously-skip-permissions"' >> ~/.zshrc && source ~/.zshrc
Windows PowerShell:
if (!(Test-Path $PROFILE)) { New-Item $PROFILE -Force }
Add-Content $PROFILE 'function yolo { claude --dangerously-skip-permissions @args }'
Now yolo launches Claude in YOLO mode. The reference ~/.claude/settings.local.json ships with safe defaults (read freely, run common dev tools, gate destructive ops). Customize as you learn what you trust.
| Plan | Price | Best for |
|---|---|---|
| Pro | $20/mo | Trying Agnes — base limits, resets every 5–8h. Will hit limits within first week of real use. |
| Max 5× | $100/mo | The natural home for daily Agnes use. 5× Pro limits, weekly reset. |
| Max 20× | $200/mo | Heavy parallel work — multiple subagents, long sessions, power users. |
Subagents and long conversations burn through quota faster than quick questions. Treat it like a data plan: Pro is the trial, Max 5× is the real subscription.
gws not found" after npm install — nvm/fnm puts node binaries in a path that's only on PATH inside interactive shells. Symlink to ~/.local/bin/: ln -sf "$(command -v gws)" ~/.local/bin/gws.CLAUDE.md or .claude/rules/; they survive across sessions.~/.claude/settings.local.json allowlist. The reference one ships with sensible gates; customize as you learn what you trust.The fastest debugger is Claude itself. Paste your error message into the chat. Most setup issues resolve in one or two turns.