Call wiki

Things I claimed or wondered about on calls, checked against primary sources.

One short entry per topic, every claim linked to its source. Claims sit on a confidence ladder — tried reported documented inferred — each page states its default under the title, and only deviations are underlined in the text (hover for the note).

Start here

Primers: broader entry points that group the single-question entries, built from what I keep finding myself explaining on coaching calls. Each also sits pinned at the top of its sidebar category.

• Where to start — the nine-step setup order the coaching calls actually follow.
• Cowork intro — everything you need to start working with Claude Cowork.
• Connectors intro — how Claude actually sees your Drive, Slack, and email.
• Browser intro — what an AI driving your browser can do, and when it’s the right tool.
• Meetings intro — turning your calls into something your AI can read.
• Recurring intro — putting Claude on a schedule, and what can trigger it.
• Documents intro — moving words between agents and Docs, Word, and Sheets.

Recurring tasks

Getting AI to act on a schedule or on events, without babysitting it.

• Event triggers — no-code Claude can’t react to events; hourly scheduled tasks are the floor, routines’ API endpoints the code-adjacent workaround.
• Local scheduled runs — a non-bare claude -p run inherits your MCP servers, connector logins, and saved permissions; cloud routines are only needed when the machine may be off.
• Newsletter digest — one plus-address (or forwarded second account) under a Gmail label, read by the Gmail connector, summarized by a weekly scheduled task.
• Task state — every scheduler starts each run as a fresh session; keep a ledger file the run reads first and rewrites last (working-folder markdown, committed file, or a Google Doc via connector).

Connectors

What Claude actually sees when it reads Slack, Drive, or email through a connector.

• Doc freshness — connector reads are live tool calls and Drive-added docs sync; plain uploads are snapshots.
• Connector scope — connectors see the whole connected account (Drive entire, Slack incl. your DMs); scope with a limited second account and permission settings, not instructions.
• Drive search — AI Drive search is a live ~10-result keyword query that misses non-Google file types; a navigation map or localized copies beat it.
• Files by editor — no Drive filter searches by who edited; approximate with modified-since plus last-modifying-user metadata, or the Activity API.
• Feeding spreadsheets — export as CSV: pasted Sheets links fail, the connector reads Sheets only as tool calls, and raw .xlsx needs an agent that can run code.
• Two Google accounts — each Google connector holds one sign-in and a second replaces it; forward/share on Google’s side or swap connections.
• Asana access — Asana’s stock connector handles most task/project reads and writes; tags, sections, and bulk automation need a community MCP server or the API.
• Site corpus — mirror the site locally (CMS export > sitemap pull > crawl) so consistency checks become exhaustive greps.
• WhatsApp access — only via a community MCP bridge joining as a linked device: no official connector, local hosting required, outside WhatsApp’s ToS.
• Signal access — only via unofficial signal-cli bridges linked as a secondary device; new messages only, unless importing Signal Desktop’s local database.
• Ashby access — the MCP server reads the pipeline but writes only notes, stage changes, candidates and consider-for-job; feedback credited to a named user and tied to an interview needs the API’s applicationFeedback.submit.
• Local vs remote MCP — a .mcpb bundle installs a local MCP server in one click with credentials pasted at setup (Desktop and Claude Code only); per-user login needs a hosted OAuth 2.1 server, which then works on every surface.

Claude

What each Claude surface — browser, desktop, mobile — can and can’t do.

• Desktop vs browser — chat is identical; desktop adds local files (Cowork), quick entry with dictation, local extensions.
• Cowork vs Code — setup is now a one-line install, so choosing Claude Code is about wanting your own file system and configuration, not coding skill.
• Sharing with teammates — Claude Tag needs one org-Owner setup and then a whole Slack channel can use it; Team-plan artifacts and projects share org-only, so a public page needs external hosting.
• Cowork sandbox — Cowork runs in a cloud sandbox or a local Linux VM; outputs are only real files when you work in a connected folder.
• Permission modes — one ladder on both surfaces: manual approval, auto (a classifier reviews actions), skip (no checks); org approvals and connector write prompts survive every rung.
• Account migration — one-way only: personal accounts can migrate into a Team/Enterprise org; between personal accounts there’s a JSON export but no import.
• Memory vs files — instruction files you write and audit beat automatic memory; the inclusion test is staleness, not importance.
• Subagents in chat — Cowork dispatches parallel subagents natively (isolated context, summary back); plain chat only via Research.
• Skills vs prompts — Skills auto-load when a task matches their description; they beat loose markdown prompts once a procedure repeats.
• Claude Design — Claude Design pairs chat with an editable canvas for prototypes, decks, and one-pagers; beats an HTML artifact for click-and-drag refinement and PPTX/PDF export.
• Web fetch limits — fetches fail on robots-blocked, bot-detecting, JavaScript-rendered, or private pages, and a failed fetch can silently become search-snippet fabrication; demand a verbatim quote.
• Biosecurity refusals — bio-writing refusals are false positives from weapons-acquisition classifiers Anthropic acknowledges; retry another model or send feedback.
• Plugins for teams — a plugin is a repo folder of skills plus an .mcp.json of connectors that Cowork and Claude Code both install from; pushes reach users on the next marketplace refresh — manual unless an org Owner provisions it and bumps the version.

Agents

Coding and browser agents across vendors — what they can reach and how they’re contained.

• Browser agents — browser agents can do anything clickable via your signed-in Chrome or a credential-less embedded browser; site permissions contain them, and blocking a site doesn’t touch connectors.
• Agent isolation — Cowork isolates in a VM, Codex fences your real machine by default, Claude Code runs in your real shell; the isolation model is a setting, not the product.
• Cross-model review — both agents are shell commands (claude -p, codex exec), so either can call the other; an AGENTS.md rule makes the review non-optional.
• Codex settings — pick the top model on medium effort, and check the sandbox’s network-access setting: off by default, it silently breaks web research.
• Codex training privacy — personal ChatGPT plans train on Codex content unless you flip the opt-out; business plans don’t by default — and opting out stops training, not the upload.
• GUI to protocol — many GUI-only apps sit on a scriptable protocol, daemon, or API an agent can drive instead of the screen.

Documents

Getting words and formatting between agents and Office/Google documents.

• Markdown vs Office — agents edit plain text directly but manipulate Office files as zipped XML through code, blind to the rendered result; draft in markdown/HTML and convert early.
• Writing to Docs — save Claude-created files to Drive for new docs; in-place Doc edits with history intact need the Docs/Sheets APIs, since the connector only uploads files.
• Comments & suggestions — the Drive connector strips comments and suggestions, but Google’s APIs expose both.
• Comment-mode editors — Roughdraft reviews agent-written markdown with comments and suggested edits, Lex does it in its own editor, and agents can now comment on Google Docs directly.
• Unsaved-note backups — “unsaved” editor notes live in real session-backup folders (Notepad++, Windows Notepad, VS Code) that a file-access agent can read.

Privacy

Who owns, sees, and trains on what you feed an AI.

• Team account data — on a Team plan the org owns the data, only the Primary Owner can export members’ conversations, Anthropic doesn’t train on them, and teammates see nothing.
• GDPR personal accounts — personal Claude accounts have no Article-28 processor contract for other people’s data; commercial plans carry the DPA that fixes it.

Security

What actually protects you when an agent has your files.

• Agent safeguards — approval prompts, scope limits, and version history protect you, not the agent’s promises; prompt injection is heavily demonstrated but has no documented real-world personal-agent harm yet.
• Credentials via 1Password — put op:// references in configs and launch agents under op run from a shared vault; attribution needs one API key per person.

Pricing

What Claude plans and API routing actually cost.

• Plan economics — overage credits bill at full API list rates while subscriptions discount steeply; upgrade tiers (or buy a second subscription) before enabling usage credits.
• Cheap-model routing — cheap API tiers price tokens at ~1/10th–1/30th of frontier rates, turning a $25 run into ~$1 for bounded tasks.

Evaluation

Judging text — human or AI — with something better than gut feel.

• AI-text detectors — detectors hold up on mostly-AI text (~0.01% false positives in independent tests) but collapse on human-drafted, AI-polished hybrids; flag, never auto-reject.
• Grader calibration — have the LLM grade a human-labeled sample blind, score it with a separate script, and require the same agreement threshold as human graders.

Models

How model behavior itself constrains a workflow.

• Context rot — degradation is gradual and starts well below the window limit; restart on quality drop or task switch, not a token count.

Meetings

Capturing calls as AI-readable transcripts, and who else can see them.

• Granola auto-transcribe — no auto-join setting exists; Granola only transcribes notes you click into. Privacy claims vs Fireflies verified.
• Granola agent access — four read-only surfaces: hosted MCP server, Business-gated API, the desktop app’s local cache, and a once-daily CSV export.
• Transcript cleanup — both fixes work and stack: load your jargon into the tool’s custom-vocabulary setting before the call, and run an LLM cleanup pass after.

Notifications

Making the right things ping you — and only those.

• Slack pings — channel reminders and plain bot posts don’t ping by default; fix via “All new posts”, an @mention, or a keyword.
• Sender identity — who an automated Slack message appears to come from, and whether anything can send as you.

Tools

The desk-level tooling around the AI itself.

• Wispr Flow — Wispr Flow dictates cleaned-up text into any app; macOS dictation is verbatim-only and Claude’s Caps-Lock dictation only talks to Claude.
• Raycast vs Spotlight — Spotlight now has basic clipboard history, but Raycast adds formatted long-retention history and extension hotkeys, at the cost of the Accessibility permission.
• Superconductor — Superconductor runs unlimited Claude Code/Codex agents in parallel cloud sandboxes; repo-shaped work only, never your local files.
• Mac-first hardware — Mac-first is a dated, real pattern (Windows lags 1–13 months, shrinking); any recent Mac meets Cowork’s specs, Windows also needs hardware virtualization.
• AI page publishing — ChatGPT Sites and Claude artifacts host a page for you, but business plans default to workspace-only viewers; outsiders need public publishing on or outside hosting.
• Sheet-backed dashboards — Host the HTML on a free static host and read the Sheet through a read-only endpoint (publish-to-web, gviz, opensheet, Apps Script); viewers can’t edit, anyone with the URL can read.