How an AI draft becomes a Marketplace search you still approve

The AI builder turns one sentence into a draft Marketplace search, dry-runs what a sweep would do, and waits. Crawlbench never auto-saves that draft and never messages sellers. You approve every field, then you activate on purpose.

A shorter version of "sentence to search" already lives in Facebook Marketplace AI monitor. This post is the control surface around it: plan limits while drafting, the dry run, and the fact that walking away leaves nothing live. The judgment piece is AI should fill filters, not pick deals.

What is the AI builder, and what is it not?

It is a drafting assistant for the same wizard you would fill by hand.

It is:

• A plain-English sentence in, a structured draft out.
• A handoff into the search wizard, where you can still edit.
• A dry run that shows what a sweep would have pulled, ignored, and alerted on.
• A save button that only you press.

It is not:

• A sold-profit oracle.
• A bot that messages sellers.
• An agent that scans everything and files deals while you sleep.
• A silent search creator. A draft is not a search.

Search filters shows the product line in public: "4Runners under 26k within 100 miles of Dallas, no salvage." The assistant drafts that search, flags what is missing, and hands you the wizard. If you close the tab, nothing is checking Marketplace.

How do you go from a sentence to a saved search?

Use one deal you would actually buy this week. The public example is a truck. An electronics line works the same way: "Open-box ThinkPads under 600 within pickup distance of two cities, no parts-only and no iCloud locks."

1. Say the deal in one sentence. Include identity, a price ceiling, where, and the hard nos. "Cheap phones everywhere" is not a sentence the builder can rescue. It will draft something wide, and you will have to tighten it.
2. Read the drafted fields. Category, includes, excludes, price, cities, and radius should look like the wizard. Category does not change later just because you wish it would. If the draft picked the wrong lane, fix it before save.
3. Read the plan limits it already applied. Price floors, radius limits, and the number of cities show up while it drafts, so save time is not the first time you hear "that city list does not fit." A Day Pass is one search. Make this draft the one.
4. Dry-run before you commit. Look at what a sweep would have pulled, what it would have ignored, and what it would have alerted on. If the sample is junk, fix excludes and price now. Do not activate and hope alerts will filter it.
5. Apply, edit, then save. Apply only seeds the wizard. Nothing is stored as a search until you press save. Change any field you do not believe.
6. Activate, then connect a channel. The first fill lands in your results. After that, add email or Telegram in alert channels if you want a ping. Alerts come after the gate, not before it.

Callout you can tape to the search: if you walk away before save, nothing is live.

What do you still have to approve?

Everything that decides whether a listing enters your results.

• Identity and category. The builder can miss a field you care about. Add it in the wizard.
• Excludes. A draft that forgets "salvage" or "for parts" will look helpful and then flood you. The search excludes pattern still applies after the sentence.
• Cities you will not drive. Every city counts toward your plan. Delete the metro you added to feel covered.
• The dry-run result. Pulled versus ignored versus would-alert is the honesty check. A pretty sentence with a junk dry run is a no.

You are not approving a deal score. You are approving rules. Create the first search is the same form, with or without the assistant.

Which plan limits should the draft already respect?

The builder is supposed to apply the rules of the plan you are on, so save is not a surprise.

On a Day Pass that means one search, up to 20 cities, and a Market Snapshot window of 7 days. A sentence that names eight cities can still fit the cap. It can also be more geography than you will triage today. Cut the list to the cities you will open. The number is a ceiling, not a target.

Radius limits and price floors get applied in the draft for the same reason. If you asked for "anywhere in the state" and the draft pulls the radius back, that is the product telling you the search was wider than the plan. Edit the sentence or accept the tighter radius. Do not discover it as an error after you thought the search was live.

Speed comes from the plan. Once a search is live, Crawlbench checks busy cities every 10 minutes on Starter, every 5 on Growth, every 2 on Pro and every minute on Scale, as pricing shows on 18 October 2026.

Paid plans change the number of searches and cities. Pricing is the table. The builder should reflect the plan you are drafting on. If it does not, fix the draft before save rather than assuming the meter is wrong and the save will be generous.

Where do drafts go wrong?

Three failures show up more than clever prompts.

The prompt is vague. "Good laptops near me" becomes a wide draft. Tighten identity, ceiling, and nos, then dry-run again.

A field is missing. The assistant can skip something you assume is obvious, like a year floor or an exclude. The wizard is where you add it. The readiness meter helps, and your review is the final check.

The dry run is junk and you activate anyway. That is how unused searches happen. People start a pass, leave a draft unsaved or save a pile of keywords, and never reach a queue they would message. Fix the filters, then activate one lane.

Crawlbench will not message the seller at any of these steps. FAQ is the short version. The builder does not create an exception.

Prove one approved search on a Day Pass

The useful test is draft, dry-run, save, activate, and read your results inside 24 hours. A Day Pass is no card and one search. Say the deal once. Approve the search. If the results look like units you would message, connect one alert channel. If they do not, edit and rematch before you add a second idea.

The Day Pass post covers the window. This post is why the builder belongs in that window: it gets a real search saved, which a sentence in your head does not.

There is a longer free trial if you already know you need several days of iteration. Use that instead of a Day Pass. Do not combine them into one offer. For this path, the Day Pass is the point, because the leak is searches that never go live.

What "AI" is not allowed to mean here

AI sped up the draft. Excludes, the price band, and your save click still decide whether an alert is worth opening. No accuracy percentage, no claim that the model picks winners, no autonomous sourcing. If the dry run looks wrong, the sentence was not finished.

Questions people ask before they trust the draft

Does the AI save a search automatically?

No. Apply seeds the wizard. The search exists when you press save. Leave the page and you have a sentence, not coverage.

What is a dry run?

A look at what a sweep would have pulled, ignored, and alerted on before you commit. Use it to fix excludes and price. Once the search is live, new listings that pass arrive as alerts on your plan's check.

Can the builder message sellers?

No. Nothing in the draft, the dry run, or the saved search sends a Facebook message. You do that yourself after you decide a row is worth it.

What if the draft misses a field?

Add it in the wizard. The assistant can skip something you care about. That is why approval is the product, not a courtesy step.