Knowledge Work Plugins: give a work assistant a shared task file
Measure productivity’s context cost and correction work
Design a small evaluation of memory lookup and task extraction without repeating unverified coverage or productivity claims.
What you will learn
- Choose a fixed set of requests
- Count work that the user still performs
- Evaluate stale memory deliberately
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
- Clarifying an unknown term can be the correct result.
- Human corrections belong in the cost model.
- Memory coverage claims require a measured sample.
Choose a fixed set of requests
Prepare requests containing known abbreviations, unknown names, deadlines and ambiguous commitments. Label the intended interpretation before the trial. Include a case that should trigger a clarification instead of silently creating a task.
Compare file-only operation with a small, approved connector-assisted run using the same host and model configuration. Record which context was available in each condition. Otherwise a larger information supply can be mistaken for a better memory policy.
Count work that the user still performs
Measure correct task fields, unnecessary tasks, missed commitments and minutes spent correcting the output. For memory questions, record whether the answer came from working memory, deeper files or clarification when this is observable. Leave unobservable paths unknown.
Collect host-reported model usage and connector latency where available. Local Markdown storage does not eliminate model processing, subscriptions or paid external services. The source’s compact-memory targets do not establish a monetary saving.
Evaluate stale memory deliberately
Change a fictional project owner halfway through a controlled sample and update the relevant memory files. Check whether later answers use the new owner and whether the old value remains in another file. This tests maintenance burden as well as initial setup.
Report the sample size, revision and unresolved errors with any adoption decision. We have not executed this benchmark. Blank measurements in the proposed record are intentional and should be replaced only with observed results.
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
Label a fixed request set with expected task fields and clarifications.
- 2
Record model usage and human correction time where observable.
- 3
Repeat selected requests after changing a fictional owner.
Copy-ready example
{
"experimentProposal": true,
"requests": null,
"correctFields": null,
"missedCommitments": null,
"correctionMinutes": null,
"modelUsage": null,
"executed": false
}Frequently asked questions
Does a local dashboard make the assistant free?
No. Host model access and connected services can still have costs.
Should every request become a task?
No. Ambiguous or informational requests may need clarification or no task at all.
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/README.mdSource checked 2026-09-18