benchr

What do you want to do with AI?

Start from the goal. Every route below ends in something benchr has read the documentation for, with the limits attached.

Most asked

Turn a screenshot into a working page

You have a picture of the thing you want and no design file, and the gap between the two is a day of measuring padding by eye.

Also called make website from screenshot · screenshot to html · image to ui code · design to code

Ways to do it

One screenshot in, a working page outFastest route. One image, one file back, then correct by sending a screenshot of your own render.
Work through an entire repository, not a pasted fileWhen the page has to land inside an existing codebase and match its conventions, not stand alone.
Drive a real browser, not a scraperWhen the source is a live page rather than an image: read the accessibility tree instead of guessing from pixels.

Screenshot to a reviewed pull request

Understand a codebase nobody has explained to you

The code is too big to read and too important to guess at, and the person who wrote it left.

Also called analyze a huge repository · understand a large codebase · onboard to a new codebase · read my whole repo

Ways to do it

Work through an entire repository, not a pasted fileThe agent reads the repository itself, follows imports, runs things, and answers from what it found rather than from a paste.
Fit a whole codebase in one prompt, and find its limitCheaper for one specific question over a corpus that fits. Ask one thing per request - multi-fact questions are where accuracy drops.
Hand off a task and walk away from the machineFor an audit that takes hours: start it, close the laptop, review the branch.

An undocumented repository to a documented test suite

My coding agent fixed the bug and broke three other things

Automate something on a site that has no API

The data is on the screen, the vendor has no API, and the export button does not exist.

Also called browser automation · scrape a site that blocks scrapers · automate a website with no api · fill a form automatically

Ways to do it

Drive a real browser, not a scraperBest fit when everything happens inside webpages. Element references beat coordinates, and the accessibility tree beats screenshots.
Operate a desktop applicationOnly when the task leaves the browser - a file dialog, an installer, a native client.
Give a model your own tools and data through MCPThe durable answer. If you will do this weekly, wrap it once as a server instead of driving a browser every time.

A site with no API to a daily feed you own

Get clean data out of a stack of PDFs

The numbers you need are in charts and tables inside documents nobody will retype.

Also called best ai for pdf analysis · extract tables from pdf · read a scanned contract · invoice data extraction

Ways to do it

Read a 600-page PDF including its chartsStart here. Text and page images together, so a figure that only exists in a chart is still answerable.
Get JSON that always matches your schemaAdd this the moment the output goes into a database instead of a person's eyes.
Run the whole job overnight at half priceFor a backlog rather than a document. Nobody is waiting, so pay half.

A pile of PDFs to a structured dataset

Get current answers you can actually check

An answer without a source is a rumour, and a training cutoff is not a research method.

Also called ai with citations · research assistant with sources · current information from the web · check the latest docs

Ways to do it

Answer from the live web with citations attachedCitations come back with the URL, the title and the quoted span - and those fields do not count as tokens.
Drive a real browser, not a scraperWhen the source needs a login, a click path, or a page that search will not surface.

Let a model use your systems, not just your prompts

The useful context is in your database, your tracker and your drive, and pasting it in does not scale.

Also called mcp server · connect ai to my database · give claude access to my tools · custom connector

Ways to do it

Give a model your own tools and data through MCPOne server, every client. Start read-only and keep credentials on the server side.
Teach the model a procedure once and reuse itWhen what you are sharing is a procedure rather than a data source.

A site with no API to a daily feed you own

Cut the API bill without changing the answer

Most bills are large because the same context is paid for repeatedly, not because the model is expensive.

Also called reduce llm cost · cheaper ai api · lower token spend · prompt caching

Ways to do it

Stop paying twice for the same fileFirst move, always. Reads cost a tenth of base input, and the usage block tells you whether it worked.
Run the whole job overnight at half priceSecond move: anything with no human waiting on it goes asynchronous at half price.
Run a capable model on your own machineThe structural answer for high-volume, low-difficulty work - if you have the memory for it.

A cost audit to a cheaper route, without changing the answer

Turn a recording into something you can search and quote

An hour of audio is unsearchable. An attributed transcript is evidence.

Also called transcribe a meeting · speaker diarization · audio to text with speakers · interview transcript

Ways to do it

Transcribe a recording and label who said whatSplit the file under 25 MB, pick the model by whether you need speakers or word timestamps, then summarise in a second pass.

Raw meeting audio to decisions and owners

Hand off work and get a result, not a chat

The tasks worth automating are the ones that take hours, and you cannot sit and watch them.

Also called long running agent · background ai task · scheduled ai job · overnight code review

Ways to do it

Hand off a task and walk away from the machineCloud sessions, scheduled routines, and handoff between terminal, desktop and phone.
Run the whole job overnight at half priceWhen the work is many small independent requests rather than one long task.

An undocumented repository to a documented test suite

Make a video ad for a product

The models can generate the shots. What breaks a real ad is that the product looks different in each one, so the work is continuity and selection, not prompting.

Also called make an ad · make a video ad · create a commercial · product commercial with AI

Ways to do it

Generate video from a prompt or a stillThe shots themselves. Generate from a still of the real product rather than from a description of it, or the subject drifts between cuts.
Generate an image with legible text inside itEnd cards and on-screen type. If the wording has to be correct, set it in an editor over generated art rather than generating the letters.

A brief to a finished commercial

Keep one subject the same across a set of images

Each generation is a fresh sample, so nothing forces the second image to show the same face or the same product as the first. Identity has to be carried, not described.

Also called same character in every image · consistent character AI · keep the face the same · product photos that match

Ways to do it

Generate an image with legible text inside itWhere the set carries type as well as a subject, keep the wording out of the generation step and the subject in a reference image.
Generate video from a prompt or a stillThe same discipline extends to motion: image-to-video from one accepted frame rather than text-to-video per shot.

A brief to a finished commercial

Summarise a long document without missing the end

A partial read does not announce itself: the summary of the first forty pages looks exactly as confident as a summary of all four hundred.

Also called summarise a long PDF · summarize a report · read a long document · summarise a contract

Ways to do it

Read a 600-page PDF including its chartsThe documented route, with the page and size ceilings stated. Check the end of the file arrived before trusting anything the summary says.
Fit a whole codebase in one prompt, and find its limitWhen the document has to sit alongside a long conversation, the window is the binding limit rather than the upload.
Let the model write and run real code mid-answerFor tables and figures, have it parse the source rather than read a rendering of it. Reading a value off a chart is the measured weak point.

Study from material I was given

The risk is not that it refuses. The risk is a confident summary of the half of the material that arrived, which reads exactly like a summary of all of it.

Also called study from a PDF · study for an exam with AI · help me learn this chapter · make notes from a lecture

Ways to do it

Read a 600-page PDF including its chartsThe documented route for a set of readings. Check the end of the file arrived before you revise from anything it tells you.
Answer from the live web with citations attachedFor anything you will be marked on, ask in a way that returns citations you can open and check yourself.
Transcribe a recording and label who said whatWhen the material is a recorded lecture rather than a document.

It says it read my PDF, but it only answered about the first few pages

Explain something I cannot follow

This is the thing models are genuinely best at, and the thing people check least, because a clear explanation feels like a correct one.

Also called explain this to me simply · I don't understand this concept · explain like I'm five · break this down for me

Ways to do it

Answer from the live web with citations attachedAsk for the explanation and the source together. If it cannot cite, treat the explanation as a hypothesis.
Let the model write and run real code mid-answerFor anything quantitative, have it work the example rather than describe it. A worked calculation is checkable; a paragraph about the method is not.

Check whether the answer is actually right

Fluency and accuracy are produced by the same process and look identical on the way out. Checking is a separate step, and it is the one people skip.

Also called is this AI answer correct · how do I know if AI is lying · check AI for hallucinations · verify what ChatGPT told me

Ways to do it

Answer from the live web with citations attachedThe strongest single move: make the answer carry links you can open, then open two of them.
Let the model write and run real code mid-answerFor any number, have it compute rather than assert. An arithmetic claim it ran is a different kind of claim from one it wrote.
Get JSON that always matches your schemaIn a pipeline, force a structured answer with an explicit field for 'not found', so the model has a legal way to decline instead of filling the gap.

It described the chart correctly and then got the number wrong

Build a working app from an idea

Getting to something that runs is now the easy part. Keeping it working while it grows is where the time goes, and where an agent with no boundary costs you a weekend.

Also called build an app with AI · make a website with AI · vibe coding a project · build a prototype fast

Ways to do it

One screenshot in, a working page outFastest start when you can already picture the interface: one image in, one self-contained file back.
Work through an entire repository, not a pasted fileOnce it is more than one file. Write the boundary down before you hand over write access, not after the first large diff.
Let the model write and run real code mid-answerSo it runs what it wrote before you do. Code that has never executed is a draft, whoever typed it.

Automate something I do every week

The weekly job is the right first automation: you already know the correct answer, which means you can tell immediately when it produces a wrong one.

Also called automate my weekly report · stop doing this manually · automate repetitive work · automate data entry

Ways to do it

Teach the model a procedure once and reuse itWrite the procedure down once so the same steps run the same way, instead of re-explaining it every Monday.
Give a model your own tools and data through MCPWhen the job needs your data or your systems rather than the open web.
Run the whole job overnight at half priceAnything that does not need an answer now belongs on the batch rate. Weekly work almost never needs an answer now.
Hand off a task and walk away from the machineOnce the steps are stable enough that you would let it run unattended — and not one week before that.

Everything else

Get output a program can consume, every single time

One malformed response in a thousand is a production incident, and retry loops are not a design.

Also called force json output · json schema output · stop the model returning prose · structured output

Ways to do it

Get JSON that always matches your schemaConstrained decoding against your schema. Handle refusals and token-limit cutoffs separately - the shape is guaranteed, completion is not.
Let the model write and run real code mid-answerWhen the output must also be correct, not just well-shaped: let the model compute and check it before answering.

A pile of PDFs to a structured dataset

Make a team procedure run the same way every time

You have explained the same release checklist four times this quarter and it came out differently each time.

Also called agent skills · skill.md · reusable prompt for a team · package a workflow

Ways to do it

Teach the model a procedure once and reuse itMetadata always loaded, instructions on trigger, reference files only when read. Costs almost nothing until it is needed.
Work through an entire repository, not a pasted filePair it with hooks so the non-negotiables run automatically instead of being restated.

Analyze a data file and get a chart you can trust

A model that describes your data is guessing. A model that runs the numbers is not.

Also called analyze a csv with ai · make a chart from data · data analysis assistant · excel analysis

Ways to do it

Let the model write and run real code mid-answerUpload the file, ask for the chart and the code that made it, and check the code.
Get JSON that always matches your schemaWhen the analysis feeds another system rather than a slide.

It described the chart correctly and then got the number wrong

Build something people can talk to

A stitched transcribe-answer-speak pipeline sounds like a walkie-talkie, and callers hang up.

Also called voice ai agent · realtime speech to speech · ai phone answering · sip voice bot

Ways to do it

Talk to it out loud and hear it answerOne session, audio in and audio out. Pick the transport by where the audio lives, and design interruption before the prompt.
Transcribe a recording and label who said whatFor the record afterwards rather than the conversation itself.

Generate visuals where the words are spelled right

Legible text was the thing image models could not do, and it is the thing every marketing asset needs.

Also called ai image with text · generate a poster · infographic generator · product image ai

Ways to do it

Generate an image with legible text inside itQuote the exact copy in the prompt, attach brand references, and correct in the same conversation.
Generate video from a prompt or a stillFor motion. Start from a still you already like - composition control is the hard part.

An idea to a narrated product clip

Use a model without sending anything anywhere

Some documents cannot leave the building, and some places have no reliable connection.

Also called local llm · offline ai · run a model without the cloud · private ai for sensitive data

Ways to do it

Run a capable model on your own machineStart from the memory arithmetic, not the leaderboard. Check the fit label before downloading anything.
Stop paying twice for the same fileIf privacy is really about cost, this is cheaper than buying hardware.

When it takes more than one step

Screenshot to a reviewed pull request4 steps
A pile of PDFs to a structured dataset4 steps
A site with no API to a daily feed you own6 steps
An undocumented repository to a documented test suite5 steps
Raw meeting audio to decisions and owners4 steps
A support inbox to answers you have actually graded6 steps
A cost audit to a cheaper route, without changing the answer4 steps
An idea to a narrated product clip3 steps
A brief to a finished commercial9 steps

See all