Playbooks

The part that compounds.

Anyone can call a model. What makes the answers good — and cheap — is knowing that a landing page should be one file with real copy, that a bug fix should read the error before touching anything, and that a review should report rather than rewrite.

Start building Browse the marketplace

Your sentence is matched to one of six job types, and only that job's instructions get sent. It is what makes a build finish in one turn instead of four.

ameliasagent.com/app

a site for my bakery

website playbook matched · SEO block addedbuilt in one turn · ~10k tokens

~10kinput tokens for one turn, with a playbook

350tokens added for the SEO block

~$0.02what the SEO block costs across twelve turns

What they are

Instructions per job, not one giant prompt.

Your request is matched against six job types, and only the matching playbook is sent — loading all of them into every request would cost more than they save. Below is what each one commits to.

01

You write a sentence

“a site for my bakery”

02

It is matched

Weighted by how specific the word is, not how many words hit

03

One playbook loads

The website one. The other five are never sent

04

SEO rules are added

Only because this job produces pages a browser serves

Matching is weighted by keyword length: “landing page” counts for more than “page”, so a long request cannot be dragged into the wrong job by one common word.

The six

Every job, and what it guarantees.

The six, and what each one insists on
Job	What it guarantees	SEO block
A website	Real copy about your business, never placeholder text. Mobile first. A finished palette, not default styling.	Yes
A tool or app	It has to actually work before it is called done, and run without you installing anything.	Yes
Fixing something broken	Understand the failure before changing anything, then change the smallest thing that could be the cause.	No
Spreadsheets and data	Nothing is dropped silently. If rows do not fit the shape, you are told which and why.	No
Scraping	A page that does not match the others is handled, not a crash halfway through your list.	No
Reviewing what you have	It reports. It does not quietly rewrite your code because it thought of something.	No

Anything that does not match one of these gets a short generic playbook instead of a wrong one.

The difference it makes

Same model. Same request.

Same request, same model — only the instructions differ, and the gap between the two columns below is what they buy.

Without a playbook

Re-reads the file it wrote a moment ago, on every pass
Puts the page somewhere the preview does not look, so you see nothing
Produces “Your Text Here” and a page with no title
You spend turns correcting all three

Up to ~100k input tokens for the same result

With one

Writes complete files and does not re-read its own output
The page lands where the preview serves it, first time
Real copy, and it names any detail it had to invent
Title, description and structured data included

~10k input tokens · one turn

The right-hand figure is measured. The left is what a model that re-reads its own output on each pass can spend reaching the same place — the reason these instructions exist at all.

Why it is a cost lever

Tokens are spent per turn.

A single instruction that stops a model doing unnecessary work saves real money on every job it ever runs. A measured build used about ten thousand input tokens; the same build, on a model left to its own habits, can cost ten times that and arrive at the same place.

Fewer turns, not cheaper tokens

The bill is turns. A model that finishes in one turn beats a model costing a third as much that needs two — which is also the whole argument for starting from working code.

Caught the failure that looked like bad output

Two models "failed" page builds while writing perfectly good HTML — it simply went somewhere the preview does not look, so it showed nothing. That class of failure is now designed out.

Only where it applies

The technical SEO block is appended to jobs that produce pages a browser serves, and to nothing else. A spreadsheet job should not pay for it on every turn.

The SEO block

SEO is part of the build, not an upsell.

This is the single biggest gap between what a model writes unprompted and what a site needs to be found, so it is spelled out rather than hoped for: a title under 60 characters, a description written for a person deciding whether to click, Open Graph and Twitter cards, a canonical link, exactly one h1, real landmarks, alt text on every image, and structured data using only fields that are actually known.

It costs about 350 input tokens — roughly $0.02 across a twelve-turn build. It is part of the build, not an upsell. The one thing it forbids is inventing ratings, review counts or addresses: that is fabrication, Google penalises it, and nobody asked for it.

Your own

Write one, or take someone else's.

Prompts, skills and loops that other people have written are in the marketplace, and anything you write can go back there. What Amelia knows about your particular project lives separately, in memory.

A playbook with a plan gate and a price history on top of it is an automation — a whole job you press once instead of a set of instructions you point at something.

See what a playbook produces.

Every build in the showcase used one, and every price is real.

Get started See pricing

Questions

Asked, answered

What is a playbook?

A set of instructions for one kind of job — a website, a fix, a scrape — loaded only when that kind of job is asked for.

Does this cost me more?

It costs less. Only the matching playbook is sent, and the instructions exist mainly to stop the model wasting turns.