Anthropic Claude Code

CLAUDE.md loads into every session, so every line costs context forever. Keep only what Claude can't work out by reading the code:

• Keep the non-derivable: pitfalls, rationale, and conventions that differ from tool defaults. Claude can read your directory layout; it can't guess that make test is broken and you use pytest -x.
• Cut the derivable: directory listings, dependency lists, and architecture overviews belong in the codebase, not in permanent context. /doctor will offer to trim these for you.
• Push detail down: move on-demand guidance into skills and nested per-package CLAUDE.md files so it loads only when relevant.
• Never mid-session: CLAUDE.md edits don't apply until the next session — the prompt cache is already built. Use /memory to edit and restart to apply.

Match the mode to the blast radius, not to your patience:

• Exploring an unfamiliar repo: plan mode. Claude reads and proposes, nothing is written.
• Normal feature work: acceptEdits or auto, with a project allowlist so read-only commands stop prompting. Build it with /fewer-permission-prompts.
• Unattended CI: --strict-mcp-config, an explicit --tools list, --max-turns, and --max-budget-usd. Constrain the agent rather than trusting the prompt.
• bypassPermissions: containers and disposable runners only. It skips the checks that stop a bad command, and the mode persists across supervisor restarts for background sessions.

Context quality degrades long before the window is full. Defaults worth adopting:

• New task, new session. /clear beats fighting drift from an unrelated conversation. Project memory survives.
• Name sessions you'll come back to with -n or /rename, so /resume is a list of tasks rather than a list of timestamps.
• Branch instead of arguing. Trying a second approach? /branch preserves the first so you can compare rather than reconstruct.
• Review before you ship: /diff to see it, /code-review for correctness, /security-review for vulnerabilities.

• SuperClaude is third-party, not from Anthropic. It is an independent open-source project (MIT, ~23.8k stars, actively maintained) that writes command and configuration files into your Claude Code setup.
• It adds 30 commands, plus optional cognitive personas and MCP server integrations.
• It needs Python 3.10+ — it installs from PyPI, not npm, which surprises people.
• Because it modifies your Claude configuration, install it the way you would any tool with write access to your dotfiles: read the repo first, and prefer pipx so it lives in its own isolated environment.
• Nothing here is required for Claude Code to work. The 88 built-in commands above are complete on their own.

• Reading a large CSV into context is the mistake. A 50 MB file will exhaust the context window and produce a worse answer than a five-line pandas script would.
• Say "analyse it with pandas, don't read it into context". Claude then works on the whole dataset regardless of size, because the file never enters the conversation.
• Read the file directly only to inspect its shape — headers, a few sample rows, a delimiter check.
• There is no meaningful upper bound on data processed this way — the limit is your machine's memory and patience, not Claude's.

• XLSX: 1,048,576 rows × 16,384 columns per sheet. That is the Excel file format itself, not a Claude limit — exceed it and you need multiple sheets, or Parquet.
• Excel cells cap at 32,767 characters, and a formula at 8,192.
• CSV has no format limit at all; the limit is whatever opens it afterwards.
• Very large single-file HTML dashboards get sluggish in the browser well before they get large on disk — paginate or aggregate.

• Binary formats need the library present. No openpyxl, no XLSX. Claude will offer to install it, which is a permission prompt, not magic.
• Sandbox and permission settings govern where files may be written. The default workspace-write preset confines writes to the workspace and temp directories.
• The docx, pdf, pptx and xlsx skills are plugin-provided, not bundled. Without them the formats are still generatable — you just lose the encoded gotchas, so expect more iterations.
• Long-running scripts can hit a tool timeout. For genuinely heavy jobs, have Claude write the script and run it yourself in the background.

• Claude cannot see the rendered output. It writes plotting code; it does not look at the chart. Overlapping labels, clipped legends and terrible colour choices go unnoticed unless you look or ask it to screenshot the result.
• Statistical conclusions deserve the same scepticism as any generated code — check the aggregation logic, especially group-bys, joins and date handling.
• Timezone and date-parsing bugs are the single most common defect in generated data reports. Ask explicitly what timezone the dates are in.

• # to create a memory — this was a genuine, well-known shortcut in earlier Claude Code versions. It does not appear in the current docs' list of input-start triggers (/, !, @, :, ?). It may still work as undocumented legacy behaviour — try it — but don't build a habit or a team doc around something that isn't guaranteed anymore. The current, documented path is /memory, or simply telling Claude "remember that…" or "add this to CLAUDE.md."
• /vibe — not found anywhere in the current official documentation. It reads as a demo easter egg from the talk rather than a supported command. Type /help in a session for the live, authoritative list instead of trusting this — or any — cheatsheet.

• Type # at any time (if it still works for you — see above) or ask directly to add something quickly; run /init to generate a starting CLAUDE.md from your codebase.
• Beyond hand-written CLAUDE.md files, Claude Code also keeps auto memory — notes it writes itself about your corrections and preferences, stored per-project at ~/.claude/projects/<project>/memory/ and loaded automatically. Browse or edit it with /memory.
• Run /context in any session to see exactly which memory files actually loaded — the fastest way to debug "why isn't Claude following my instructions."
• One thing from the talk worth naming honestly: a demo showed a Claude-team-internal ~/.claude/ULTRACLAUDE.md as an "[ANT-ONLY]" tier. That's exactly what the label says — an internal Anthropic mechanism, not a public product feature. It won't do anything special on your machine; it's here for accuracy about what was on screen, not as something to set up.

• Slash commands have technically merged into skills — a file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy and work identically. Existing commands/ files keep working, so the table above is still accurate — just know skills are where new capability is landing.
• MCP has three real scopes, and "user" is easy to misread as "global default." local (the default, no flag needed) is private to you and to the current project; project lives in .mcp.json and is genuinely shared via git; user is private to you but follows you to every project. When the same server is defined in more than one scope, local wins, then project, then user.