Knowledge Work Plugins: give a work assistant a shared task file
Protect workplace memory and recover from conflicting edits
Set boundaries around people data, connector actions and the shared task file before allowing broad workplace scans.
What you will learn
- Keep memory appropriate to the workspace
- Require approval for external effects
- Recover from a file discrepancy
Before you start
- Basic command-line and configuration reading
- Ability to work in a disposable authorized environment
Design an independent parser preview that exposes unsupported Markdown and conflicting headings before a dashboard save.
Key takeaways
- Readable local files may still contain sensitive data.
- Connector grants have their own lifecycle.
- A rendering check is narrower than a security audit.
Keep memory appropriate to the workspace
People and project files can accumulate sensitive workplace details. Store only information needed for the authorized task and choose who may read the folder. Avoid adding credentials, private HR notes or unrelated personal facts simply because a connector can retrieve them.
CLAUDE.md is also a common host instruction filename. Inspect an existing file before any initialization or update, and preserve the user’s project rules. A memory exercise must not silently replace the instructions that govern a real repository.
Require approval for external effects
The task skill explicitly asks for confirmation before adding commitments extracted from conversations. Keep that review step and inspect proposed owners and dates. An external message, ticket update or broad scan needs its own authorized scope.
Treat retrieved documents as information rather than instructions to expand access. Restrict connector accounts to the intended work and revoke grants through the service when the trial ends. A local plugin setting alone is not proof that remote access has been removed.
Recover from a file discrepancy
If the board and agent disagree, stop both writers and copy the current task and memory files before diagnosing. Compare the last known good snapshot with the current Markdown, then inspect unsupported headings, serializer changes and modification times.
The dashboard escapes task text in the inspected rendering paths, but that observation is not a complete security audit of the HTML application. Test untrusted strings and browser permissions in your environment. Restore only the affected demo or workspace files after reviewing the diff.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Limit the memory folder to approved workplace context.
- 2
Confirm extracted commitments and external actions before applying them.
- 3
Pause concurrent writers and preserve files before recovery.
Copy-ready example
{
"policyExample": true,
"externalWrites": "explicit approval",
"concurrentWriters": 1,
"storeCredentialsInMemory": false,
"recovery": "snapshot and reviewed diff"
}Frequently asked questions
Should I run a comprehensive scan first?
Start with a bounded file-only trial, then approve any connector and scan scope separately.
Can I recover by reinitializing everything?
Preserve the current files first. Reinitialization may overwrite evidence or existing workspace instructions.
Sources
- Knowledge Work Plugins / productivity/skills/task-management/SKILL.mdSource checked 2026-09-18
- Knowledge Work Plugins / productivity/skills/memory-management/SKILL.mdSource checked 2026-09-18
- Knowledge Work Plugins / productivity/skills/dashboard.htmlSource checked 2026-09-18
- Knowledge Work Plugins / productivity/.mcp.jsonSource checked 2026-09-18