Knowledge Work Plugins: give a work assistant a shared task file
How productivity memory and task edits move through the system
Trace hot memory, deeper lookup and dashboard serialization while keeping external service access outside the local file loop.
What you will learn
- Resolve meaning before execution
- Share a file, not a transaction log
- Draw the external boundary explicitly
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
- Memory lookup is a host instruction policy.
- The dashboard writes a task-specific representation.
- External actions cross a separate permission boundary.
Resolve meaning before execution
The memory skill first looks in CLAUDE.md, then in memory/glossary.md and deeper people or project files. If a term is still unknown, it calls for asking the user. This is an instruction-level lookup policy, not a database index with guaranteed retrieval behavior.
Frequently used context is promoted into working memory; stale context is moved out of that compact file. The source’s coverage percentage is a design target. We have not measured the proportion of real workplace requests resolved at each tier.
Share a file, not a transaction log
The task skill and dashboard both operate on TASKS.md. The browser parses headings and checkbox lines into sections and cards, then serializes that internal structure when saving. The agent can separately edit the same file through host tools.
This arrangement makes the state visible, but it does not establish atomic coordination between writers. The dashboard skips external reloads while local edits are pending or saving. A team should test overlapping edits before relying on it for shared commitments.
Draw the external boundary explicitly
Connectors let the host obtain additional context or perform authorized actions in work systems. Those requests leave the local file workflow and inherit the connected account’s permissions. Keep an extracted action item distinct from a confirmed external update.
For troubleshooting, inspect the stored meaning, task text and connector result separately. A correctly decoded project name does not prove a task was created remotely. This decomposition gives each failure a next check without guessing that all problems come from the model.
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
Trace one abbreviation through working and deep memory.
- 2
Follow one board edit into TASKS.md.
- 3
Record external action responses separately from local task state.
Copy-ready example
CLAUDE.md → memory/glossary.md → detailed files → ask
TASKS.md → parseTaskMarkdown → board → toMarkdown → TASKS.md
Authorized connector → external service resultFrequently asked questions
Is memory/ automatically a vector database?
The inspected skill describes Markdown files and a lookup sequence, not a vector database.
Does a local task prove a remote task exists?
No. Check the external service response and identifier separately.
Sources
- Knowledge Work Plugins / productivity/skills/memory-management/SKILL.mdSource checked 2026-09-18
- Knowledge Work Plugins / productivity/skills/task-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