I Stopped Pasting Prompts After I Moved Them Into Gemini Skills

I started using Gemini Skills for a boring reason: I got tired of pasting prompts. Every morning I pasted the same 20-line instruction into the chat box, and every afternoon I pasted the same 20 lines again. On weekends it happened twenty-five times. I did that for more than a month and my eyes hurt.

Two weeks after moving the work into Skills, here is the honest result: I no longer paste prompts. Three new problems replaced the old one, and they bother me more than the time I saved. Everything is below.

This is not an introduction to Gemini. It is what happened to me over two weeks of use. Prices are the public rate cards as of 2026-10-04, and anywhere I did arithmetic myself I say so.

A Day Before I Had Skills

You cannot see the change without seeing the before. This is the loop I actually repeated.

That adds up to roughly 100 lines of prompt text pasted by hand every day. It is a 20-second job, so by my own arithmetic it is 10 minutes a day, 40 hours over 240 days. I found out late that I could have called this from a local script or a macro.

What Actually Changed After the Move

1) One slash invokes it

Once a skill is registered, typing a slash in the chat box brings up the list, and typing the name applies it. You are not pasting an instruction — you are calling something by name. Twenty repetitions a day became twenty keystrokes. The 09:10 row above is the clearest example.

The difference is obvious. Previously, when a result looked wrong, my first suspicion was "was my prompt bad?" Now the only thing I check is whether I called the right skill name.

2) Skills get created mid-conversation

I do not shape the form by hand in the settings screen. Skills can be created and revised while I am talking to the model. Here is what I actually did:

"Make a skill out of this prompt. Goal: draft reply to a customer
 question. Input: the question text and the customer tier.
 If you cannot do it, ask me why first."

Then I add one more line — "always add a one-line Korean summary before the reply" — and the skill changes immediately. So there are only three questions I have to answer myself when building a skill.

• Why am I writing this again (how often it repeats)
• What has to go in (inputs)
• What is allowed to fail (failure tolerance)

Once the goal and the inputs are set, the rest gets refined in conversation.

3) You can stack two of them

Calling two or more skills in the same conversation applies both sets of instructions together. The combination I use most is this:

/copywriter  +  /seo  +  /brand-tone

Calling a strong copywriter skill together with an SEO skill produces promotional copy where both sets of instructions are already merged. Adding the brand tone check on top reduces how much a human has to re-read.

There is something to check when you stack. I will write down what I got wrong first.

Merging is merging, but conflict is still conflict. The more you stack, the cheaper it is to say the direction out loud one more time.

4) Documents stay attached to the skill

Company profiles or templates (CSV, MD, text) do not need to be uploaded every session. You can attach reference documents permanently to a given skill, so the output is always grounded in the same material.

This was the quietest saving. It was never really about the 30 seconds of uploading — it was about the time lost to "which file was it this time." Because the document version rides along with the skill, I no longer have to keep track of whether the description doc is the latest one.

Managing Skills as an Asset

Once I knew they worked, naming and maintaining them turned out to be the hard part. These are the four habits I picked up in two weeks.

Of the four, the failure log was worth the most. Looking back at why a skill ended up shaped a certain way, I forget — but when the failure conditions are written beside it, the next skill starts from the same note.

What It Is Worth in Money

This is my arithmetic. I used public API prices as of 2026-10-04 and pulled only the models I tend to use for summary work. I did not separate input from output — I multiplied a rough token count by the input rate. This is not a measured bill.

The assumption: 100 lines a day, 30 days a month, roughly 20 tokens per line. Entirely my arithmetic. I have never actually counted the tokens. The numbers look attractive and the evidence behind them is thin, so do not take them at face value.

The table does establish one thing: bundling repetitive prompts into a skill is cheaper than switching models. Whatever model you use, the repeated cost disappears once the repetition is packaged.

People Who Have Tried to Sell This

Skills are not only a personal filing system. Packaging document drafting or question answering as a skill has become a sellable service. I would split it into three lanes.

1) Specialist document services. Legal, real estate, clinics. Many repeated document types with fixed formats. Three conditions matter: a template exists, the demand repeats, and the client checks the output for errors themselves. If any one is missing, the price for a skill setup drops.

2) Bulk SEO content delivery. A solo delivery model producing dozens of posts a day. The copy + SEO stack I described above is used exactly here.

3) Internal knowledge skills. Meeting notes, customer segmentation — internal repetition. Instead of selling it, this one gets attached to a pipeline.

Prices and competition look like this. Per-row figures cover only the range I have traded in directly.

Three skills stacked become a pipeline. Mine looks like this:

/collect   -> gather material (research)
/transform -> turn it into a draft (copy + SEO stacked)
/review    -> check facts, format, tone (rules + explicit failure conditions)

Keeping review as a skill is what lets me relax. Human hands are reserved for the last judgment call. If you leave the review skill out, the first two can run perfectly and still fail at the end.

Six Things I Stopped Doing

After two weeks, the ways I break these are also settled. Written down as they are.

1. Do not collect skills. A skill with a name and no reuse is trash. If there is no second use, do not build it.
2. Do not grow the stack. Three skills merge, but conflict comes along with the merge.
3. Do not drop instructions. Requirements that never made it into the skill have to be pasted every time. The requirements you skip out of laziness show up in the output.
4. Do not assume permanent documents are safe to attach. Attached reference documents travel with every request.
5. Do not let document versions drift. Once the attached file is stale, every answer is stale.
6. Do not give judgment to the skill. Skills reduce repetition; they do not replace decisions.

The Checklist I Have Now

• [ ] Did I repeat any instruction twice today? If so, pull it into a skill
• [ ] Does the skill name contain the audience?
• [ ] Did I write the failure conditions next to the skill?
• [ ] Did I check the date on the attached reference document?
• [ ] Did I say the conflict direction out loud once for the stacked skills?
• [ ] Do I have a review skill?

Two of the six are still empty: failure conditions and the review skill. I expect to fill them.

One Line

Deleting the prompt is what a skill is. Once I deleted them, new things showed up that needed my hands, and that is the whole two weeks.

Related Reading

• Gemini basics and 12 core features — baseline Gemini usage
• Agent Skills 2.1 fundamentals — the general theory of skills
• The token cost bomb — how repetition builds a bill
• Catching agent memory leaks — why context needs managing