Cowork Common Mistakes
Eight mistakes that show up in nearly every new Cowork user's first month — and the fix for each, before you make them.
← Back to Reference HubBest for: Mount narrowly — one project folder at a time. Set narrow defaults in user preferences so every session starts scoped.
- Symptom: the agent surfaces files you didn't expect — old tax PDFs, archived chats, env files
- Symptom: the folder approval dialog showed a path higher up the tree than the project you actually opened
- Fix: revoke the broad mount in Settings → Folder access, re-approve the narrow folder you actually need
- Prevention: set Default mounts in user preferences to the two or three folders you use most
Limitations: Scope creep is the most common Cowork mistake by a wide margin. Authority granted is authority used — the agent has no way to know which files inside an approved folder were "really" intended to be in scope.
Best for: Tighten each skill's description to the specific workflow it owns, or uninstall the ones that aren't earning their slot.
- Symptom: the same prompt produces different output structures across sessions
- Symptom: you can't predict which skill will fire before you fire the prompt
- Fix: open each overlapping skill, tighten the description to the workflow it actually owns
- Fix: uninstall skills you haven't used in a month — they're competing for triggers without earning their slot
Limitations: Skill descriptions are the only thing standing between a prompt and the right capability. Vague descriptions create lottery-ticket behavior; sharp descriptions create predictable behavior.
Best for: If the job must run regardless of your sleep schedule, use Routines (cloud-hosted). For personal, desktop-bound briefings, Scheduled Tasks are fine.
- Symptom: the morning briefing you set up yesterday is missing
- Symptom: /schedule list shows the task as queued or skipped, not completed
- Fix: rebuild the job as a Routine (cloud-hosted) if it has to fire on a guarantee
- Fix: if a desktop task is acceptable, set your laptop to never-sleep on AC power and accept the failure mode
Limitations: Scheduled Tasks and Routines are two different products with similar names. Cowork surfaces both; the trap is treating them as interchangeable. Pick the one whose failure mode you can live with.
Best for: Revoke, re-authenticate with read-only. Add write scopes only after the read flow has been working for a week and the specific write workflow demands it.
- Symptom: surprise actions in the connected service — sent emails, modified events, posted messages
- Symptom: OAuth scope on the connector shows write/manage permissions you don't remember granting
- Fix: revoke the connector in Settings → Connectors, re-authenticate, decline write scopes on the consent screen
- Fix: keep a one-line note of which connectors have write scope and why — make the trust grants visible
Limitations: Once granted, write scope persists until you revoke it. There's no per-prompt scope downgrade. The discipline is at the OAuth screen — slow down, decline what you don't need.
Best for: Connect the apps you actually use daily, read-only first. Add new connectors only when a real prompt surfaces the need.
- Symptom: the agent picks unexpected connectors for tasks ("why is it using Jira here?")
- Symptom: you don't remember granting access to half the apps in your connector registry
- Fix: in Settings → Connectors, disable the ones you have not used in two weeks
- Fix: install new connectors only when a specific prompt fails because the connector is missing
Limitations: The marginal connector adds more discovery cost than capability — until you actually have a workflow that uses it. Install reactively, not aspirationally.
Best for: Check the connector registry before reaching for Chrome. Connectors are stable typed contracts; UI selectors change without warning.
- Symptom: tasks that should take seconds take minutes — clicks, waits, scroll, retry
- Symptom: brittleness when the SaaS UI changes a selector or layout
- Fix: search the connector registry for the SaaS you're driving via browser — install it if it exists
- Fix: keep the Chrome bridge for genuinely UI-only workflows (configuring a vendor that has no API)
Limitations: The Chrome bridge is the right tool when there's no typed alternative — vendor onboarding, settings pages, dashboards with no API. It's the wrong tool when an MCP connector already exists.
Best for: Before installing a plugin, list what it ships (skills, hooks, MCP servers) and read the SKILL.md descriptions. The same vetting you'd give an npm package or a VS Code extension.
- Symptom: unexpected actions on certain prompts you can't trace to a skill you wrote
- Symptom: new MCP servers appearing in your tool list after a plugin install
- Fix: before install, expand the plugin's contents in the install dialog — read what's bundled
- Fix: after install, list the plugin's skills and hooks; turn off any auto-firing hook you didn't expect
Limitations: Official doesn't mean inert. Every plugin extends the agent's authority along multiple axes simultaneously — vet accordingly.
Best for: Use the role plugin (sales, finance, legal, marketing, HR, etc.) — it ships the matched set of skills, connectors, and slash commands. Customize after the baseline is in muscle memory.
- Symptom: prompts that should fire a skill don't, and you can't find the broken link
- Symptom: the workflow has gaps you can't trace (missing connector? missing skill? wrong description?)
- Fix: install the role plugin matching your domain (sales, finance, legal, etc.) and run a baseline prompt
- Fix: customize from a working baseline — replace one piece at a time, never start from scratch
Limitations: Hand-assembly works for advanced users who already know the surface. For everyone else, the plugin is the path of least resistance and least missing-piece debugging.
Mistakes scale faster than features
| Capability | Where the mistake hides | How to catch it early |
|---|---|---|
| Folder access | Over-broad mount | Per-session approval dialog shows the path — narrow it if it looks higher up the tree than the project you opened |
| Skills | Overlapping descriptions | Two skills both fire on the same prompt — tighten one description or remove the loser |
| Connectors | Excess scope | OAuth screen shows write/manage when read would do — back out and re-grant read-only |
| Plugins | Unread bundles | You can't list what the plugin ships without checking — read the contents before installing |
| Scheduled Tasks | Desktop-bound assumption | Job didn't fire — check whether the app was open; promote to a cloud Routine if reliability matters |
| Artifacts | Unhandled connector failures | Page shows an error when a connector goes down — build graceful degradation into the artifact prompt |
| Browser bridge | Using Chrome instead of an existing connector | Task is brittle and slow — check the connector registry first; keep Chrome for genuinely UI-only work |
| MCP servers | Trusted by default | Third-party server added without code review — audit it like any other dependency you'd add to a project |
Three patterns under every Cowork mistake
Scope creep (over-broad authority — folders, write scopes, plugin trust). Trust drift (treating new surfaces as zero-risk because they came from a known source). Duplication (hand-assembling what a plugin or connector would do better). Every mistake on this page reduces to one of those three. Catch the pattern and you catch the mistake.