October 8, 2026Authority AI Editorial

Prompts vs Skills: Why Rewriting the Request a Third Time Doesn't Work

A tangle of glowing loops circles back on itself on the left, while on the right a single straight path of lit steps runs cleanly to a finished panel.

It's eleven on a Thursday night, and a client proposal is due at nine. You ask your agent for it and get a draft that reads well and commits to nothing. You explain again. The second draft is closer, and still generic. You rewrite the request a third time, adding the client's name, the budget and a line about tone.

The third draft sounds right and says little. You end up doing the thinking yourself, then doing it again in the agent's words, and it's past one before you close the laptop.

Most people who use an agent for real work know this night. The usual response is to write a better prompt. Sometimes that helps. Often it doesn't, because the request was never the part that was missing.

What a prompt can carry

A prompt is the request. Write me a proposal for this client. It says what you want, and with some effort it can say a good deal about this particular job: who the client is, what they asked for, the deadline, the budget.

What it rarely says is how you would do the job yourself. How you open a proposal. Which section the client reads first. What you always check before anything goes out. The mistake you made two years ago and now look for every time.

Nobody types that out at eleven at night. It would take longer than writing the proposal. So the agent gets the what and has to guess at the how, and its guess is the average of every proposal it has ever seen. That average is the draft that reads well and commits to nothing.

Anthropic's own prompting guidance for Claude puts it plainly. It asks you to think of the model as "a brilliant but new employee who lacks context on your norms and workflows." It also offers a golden rule: show your prompt to a colleague with minimal context on the task and ask them to follow it. If they would be confused, the model will be too.

Run the proposal request through that test. A capable new colleague, handed "write the proposal for Henderson, budget is tight, keep it confident," would ask a dozen questions before starting. Your agent usually doesn't ask. It fills the gaps itself.

Why rewriting the request backfires

The natural fix is to add more to the request each time. That's where the third rewrite comes from, and there's evidence it works against you.

In 2025, a team of researchers compared how language models perform when a task arrives fully specified in one message with how they perform when the same details arrive over several turns of a conversation. Across more than 200,000 simulated conversations and six kinds of generation task, every leading model they tested did worse in the multi-turn setting, with an average drop of 39 percent.

The reason matters for the proposal. The researchers found that models "often make assumptions in early turns and prematurely attempt to generate final solutions," then lean on those early attempts. Their summary is blunt: when a model takes a wrong turn in a conversation, it gets lost and does not recover. Most of the loss came from the models becoming less reliable, not less capable.

That describes Thursday night closely. The first draft was built on guesses. Each correction was layered on top of those guesses rather than replacing them.

Length brings a second problem. A study first released in 2023, and later published in Transactions of the Association for Computational Linguistics, found that models use information best when it sits at the beginning or the end of a long input, and that performance "significantly degrades" when the relevant detail sits in the middle. A request that has grown by three rounds of patches puts the important parts wherever they happened to land.

What a skill carries instead

A skill is a written procedure your agent works through before it starts a job. It holds the part of the work that doesn't change from one client to the next:

What a good result looks like, described plainly enough to check the draft against.
The steps in order, with a check between them.
The traps: what usually goes wrong on this kind of job, and what to do when it does.

The request stays short because it no longer has to carry the method. Write the proposal for Henderson; the notes from our call are below. The skill brings everything else, the same way every time, in the same order, before the agent has made a single guess.

Anthropic's engineering team makes a related point about building agents. Its guide to effective agents calls it "crucial" to write up an agent's tools and their instructions "clearly and thoughtfully", and suggests writing those instructions as you would for a junior developer joining your team. Its advice is to "put yourself in the model's shoes" and ask whether the job would make sense with only what's on the page.

What to move out of the prompt

A simple way to sort it: anything you'd say for every job of this kind belongs in the skill. Anything true only of this job stays in the prompt.

Move into the skill:

The structure you always use, and the order the reader meets it in.
Your standard for done: what has to be true before it's finished.
The checks you run, such as every number matched to its source or every claim tied to something the client said.
The traps you've learned to look for, written as instructions.
Your house tone, described with an example rather than adjectives.

Keep in the prompt:

Who this job is for and what they asked.
The material: call notes, figures, the brief, past correspondence.
Anything unusual about this one, such as a tighter deadline or a section the client asked you to skip.

Once that split is made, the prompt gets shorter and the results get steadier. You're no longer writing your method from memory at eleven at night. You wrote it once, when you had time to think.

The same proposal, with a method

What follows is an illustration of what a proposal skill's steps tend to produce, not a logged run.

Without a skill, the agent opens with a paragraph about your company. It lists services. The client's own problem appears on page two, in general terms. The price sits in a table with no reasoning beside it. Every sentence is fluent, and none of it is specific to Henderson.

With a proposal skill, the agent starts from a method. First, restate the client's problem in their own words from the call notes. Then propose one approach, with the reason it suits them. Then the price, with what it covers and what it doesn't. Before finishing, it checks that every promise in the draft is something you can actually deliver, and it flags anything it had to assume.

The agent and the request haven't changed. The second run had your method in front of it before it began, so it didn't have to guess its way there.

When rewriting is still right

None of this means prompts don't matter. A clear request still beats a vague one, and the specifics of each job will always belong in the prompt.

Rewriting is the right move when the request itself was unclear: the wrong client, a missing figure, a misread brief. It's the wrong move when the draft is fluent and generic. That's a sign the method is missing, and no amount of rewording will supply it.

If a conversation has gone badly off course, our reading of the research above is to start a fresh one and give the agent everything in one message, rather than adding a fourth correction to a pile of earlier guesses.

Where to start

Think back to the last request you rewrote three times. Look at what you added each round. Most of it will be method, not detail: put the price after the approach, don't open with our history, check the figures against the spreadsheet.

Those additions are your first skill. Write them down in order, add one line describing what a finished version looks like, and note the trap that caught you last time. Next time the same job comes up, keep the request short and attach the method.

You can also start from methods other people have already written down. The AI Ultimate Skill Collection is 3,029 ready-made skills sorted into 12 packs, one for each kind of work, such as coding, research or finance. The Business pack includes proposal drafts, and the collection works with Claude Code and other agents that use skills.