I discovered a easy solution to cease Claude expertise from quietly going stale

Just like our personal expertise, Claude expertise and the context in them can quietly degrade over time with out updates. Anthropic itself recommends updating expertise quite than leaving them as completed markdown information. Think of them as full toolkits or mini-experts who carry on bettering with time and expertise.

I’m nonetheless getting a deal with on using skills for workflow automation correctly. Continuously patching a talent every time one thing goes flawed appears tiresome and the beginning of bloated directions, regardless that Claude expertise helps variations. So I’m attempting a workaround. I’m utilizing NotebookLM as a extra goal “upkeep man” for the Claude expertise.

I began with a easy meal talent

Even a secular activity can change shortly

My take a look at talent creates a 21-day meal plan based on my preferences (although I’ve simulated the dishes right here), accessible substances, cooking time, and leftovers. It additionally produces a purchasing listing so I haven’t got to work out what to purchase individually. The directions inform Claude to keep away from repeating the identical principal dish too usually, reuse substances throughout meals, account for leftovers, and maintain the purchasing listing sensible.

Maybe, a bit easy and foolish for an AI experiment. But that simplicity permits me to see how my way of life and even random occasions in a single week can throw off the plan in a SKILLS.md file. It’s all the time attainable to tweak and scale it up for extra advanced Claude expertise.

The essential half is that meal planning has so many variables. An ingredient might not be in inventory. A busy week can throw off the whole plan Claude makes for me. Even a time out can have a cascading impact on the meal prep. So, the kitchen is not a foul Petri dish for the Claude expertise upkeep take a look at.

One week can go flawed, whereas the following runs completely. So it is higher to have 3–4 weeks minimal of knowledge earlier than working the talent.

I log corrections as a substitute of patching

Real-world use exposes the lacking guidelines

Rather than instantly altering the talent every time one thing would not work, I maintain a easy Google Doc recording what occurred throughout every week’s meal planning. As I’m utilizing NotebookLM for this take a look at, I report the log in Google Docs, and never as an uploaded markdown file. NotebookLM syncs Drive sources periodically, so it may well self-update the supply. It’s not all the time an on the spot real-time sync. Often, you may need to manually open the NotebookLM “Sources” panel and click on the Refresh icon on the Google Doc supply to power it to tug the newest entries.

I enter issues resembling meals I skipped, substances I could not discover, recipes that took too lengthy, leftovers that weren’t used, and strategies I needed to right. I additionally report issues that labored notably properly. For instance, the talent may schedule a lentil curry on Monday and counsel utilizing the leftovers in a pancake for lunch on Tuesday. But if I uncover that the leftovers aren’t sufficient for 2 folks, that is helpful details about how the talent ought to plan parts.

The essential mindset shift right here is to not create or change guidelines with each miss. Simply recording them as observations (as scientists do in precise lab experiments) is an train to find patterns. One meal gone flawed does not imply inserting one other rule into the Claude talent.

Editing the SKILL.md file is okay for fixing real errors like typos or a rule you forgot to incorporate. Just spotlight textual content in a talent file and click on Edit.

I let NotebookLM discover the patterns

It compares my guidelines with precise expertise

After a number of weeks, I add the present SKILL.md and my meal-planning log to a NotebookLM pocket book. NotebookLM can then examine and analyze the talent alongside my precise observations.

I then ask NotebookLM to match what the talent says with what actually occurred. Here’s the precise immediate I exploit:

Act as a upkeep reviewer for my Claude weekly meal-planning talent.

I've supplied two sources:

1. The present SKILL.md file, which describes how Claude is meant to create my weekly meal plans.
2. A working meal-planning log containing what truly occurred after I used the Skill.

Compare the Skill file in opposition to the meal-planning log. List each place the place what occurred would not match what the talent says. Cite the log entry for every hole and kind by how usually it occurred.

Classify each advice as:

PROMOTE — Strong proof suggests the Skill needs to be modified.

WATCH — Interesting sample, however there is not sufficient proof to alter the Skill but.

REJECT — This might be a one-off incident or doesn't justify altering the Skill.

Important guidelines:

- Do not rewrite the Skill.
- Do not invent preferences that are not supported by the meal-planning log.
- Do not flip a single disliked meal right into a everlasting rule.
- Look for patterns throughout a number of weeks.
- Prefer small, exact modifications over including a lot of new directions.
- Distinguish between a real downside with the Skill and a one-off downside attributable to uncommon circumstances.
- Pay consideration to profitable meals too. Repeated successes might reveal helpful preferences that are not at present encoded within the Skill.

End with a bit known as "Most necessary modifications" containing solely the strongest PROMOTE suggestions.

This is the half I like most concerning the workflow. The talent is the baseline for my splendid system, whereas the meal log catches my precise habits throughout the weeks. NotebookLM helps me see the distinction between the 2. I might have skipped NotebookLM by dropping each information right into a Claude Project and asking Claude to evaluate its personal talent.

That works, however a contemporary reviewer works higher. Anthropic’s personal steering splits the job between one Claude that refines a talent and one other that makes use of it. NotebookLM solutions solely out of your sources, and you’ll hint all the pieces to its actual log entry.

I flip patterns into suggestions

Not each meal mistake wants a brand new rule

Once NotebookLM identifies recurring issues, I ask it to show the strongest findings right into a “changelog.” NotebookLM may counsel making the rule extra particular. Thanks to NotebookLM’s grounded sources, the responses come from my real-world use quite than from a mannequin’s or my very own assumptions about what seems like a very good meal-planning rule.

For occasion, it may well counsel reusing substances the place sensible, however keep away from requiring unusually massive portions of perishable substances merely to cut back the variety of objects on the purchasing listing. But I’ve additionally made the error of fixing a rule that broke one other rule. That led to errors in Claude’s response.

That’s why the Claude immediate really helpful the Promote, Watch, and Reject classes. This prevents the talent from rising each time one thing goes barely flawed. Otherwise, sustaining it might ultimately turn out to be an train in over-fixing by including exceptions on high of exceptions.

I take a look at the up to date talent once more

The new rule has to outlive the longer term weeks

After reviewing the proposed modifications, I open the talent in a Claude chat and apply the modifications. I pay specific consideration to the meals and conditions that compelled the modifications. I test whether or not the following meal plan truly solves that downside with out taking the remainder of the plan off-track.

It’s an everyday upkeep cycle. Maybe it is overkill for a easy train like meal planning, but it surely’s essential for extra necessary duties. If an replace backfires, your changelog is the one report of what modified.

The course of additionally makes it simple to roll again something that does not work. I can take a look at the correction log and see the real-world issues that precipitated it.

Pick one talent you employ probably the most

Try this with a Claude talent you already use usually. Don’t spend hours attempting to make the preliminary model excellent. Use it, report what goes flawed, and let a number of real-world examples accumulate. For occasion, you may maintain a working log of company jargon or stylistic habits you retain having to manually edit out. Coders can log syntax errors generated by Claude.

Then put the talent and your log into NotebookLM and ask it to search out the patterns. Now, a Claude talent is not a sacrosanct set of directions. You’re treating it as a small system that will get higher as you study what truly works together with your behaviors.