Start from the job, not the model

Why this comes up: 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.

Documented approaches

1. One screenshot in, a working page outQuickPartly documented

   Fastest route. One image, one file back, then correct by sending a screenshot of your own render.

2. Work through an entire repository, not a pasted fileMediumDocumented

   When the page has to land inside an existing codebase and match its conventions, not stand alone.

3. Drive a real browser, not a scraperDeepDocumented

   When the source is a live page rather than an image: read the accessibility tree instead of guessing from pixels.

Related tasks: Understand a codebase nobody has explained to you

Why this comes up: The code is too big to read and too important to guess at, and the person who wrote it left.

Documented approaches

1. Work through an entire repository, not a pasted fileMediumDocumented

   The agent reads the repository itself, follows imports, runs things, and answers from what it found rather than from a paste.

2. Fit a whole codebase in one prompt, and find its limitQuickPartly documented

   Cheaper for one specific question over a corpus that fits. Ask one thing per request - multi-fact questions are where accuracy drops.

3. Hand off a task and walk away from the machineMediumDocumented

   For an audit that takes hours: start it, close the laptop, review the branch.

Related tasks: Hand off work and get a result, not a chat, Turn a screenshot into a working page

Why this comes up: The data is on the screen, the vendor has no API, and the export button does not exist.

Documented approaches

1. Drive a real browser, not a scraperMediumDocumented

   Best fit when everything happens inside webpages. Element references beat coordinates, and the accessibility tree beats screenshots.

2. Operate a desktop applicationDeepDocumented

   Only when the task leaves the browser - a file dialog, an installer, a native client.

3. Give a model your own tools and data through MCPMediumDocumented

   The durable answer. If you will do this weekly, wrap it once as a server instead of driving a browser every time.

Related tasks: Let a model use your systems, not just your prompts

Why this comes up: The numbers you need are in charts and tables inside documents nobody will retype.

Documented approaches

1. Read a 600-page PDF including its chartsQuickDocumented

   Start here. Text and page images together, so a figure that only exists in a chart is still answerable.

2. Get JSON that always matches your schemaQuickDocumented

   Add this the moment the output goes into a database instead of a person's eyes.

3. Run the whole job overnight at half priceQuickDocumented

   For a backlog rather than a document. Nobody is waiting, so pay half.

Related tasks: Get output a program can consume, every single time, Cut the API bill without changing the answer

Why this comes up: One malformed response in a thousand is a production incident, and retry loops are not a design.

Documented approaches

1. Get JSON that always matches your schemaQuickDocumented

   Constrained decoding against your schema. Handle refusals and token-limit cutoffs separately - the shape is guaranteed, completion is not.

2. Let the model write and run real code mid-answerMediumDocumented

   When the output must also be correct, not just well-shaped: let the model compute and check it before answering.

Related tasks: Get clean data out of a stack of PDFs

Why this comes up: An answer without a source is a rumour, and a training cutoff is not a research method.

Documented approaches

1. Answer from the live web with citations attachedQuickDocumented

   Citations come back with the URL, the title and the quoted span - and those fields do not count as tokens.

2. Drive a real browser, not a scraperMediumDocumented

   When the source needs a login, a click path, or a page that search will not surface.

Related tasks: Automate something on a site that has no API

Why this comes up: The useful context is in your database, your tracker and your drive, and pasting it in does not scale.

Documented approaches

1. Give a model your own tools and data through MCPMediumDocumented

   One server, every client. Start read-only and keep credentials on the server side.

2. Teach the model a procedure once and reuse itMediumDocumented

   When what you are sharing is a procedure rather than a data source.

Related tasks: Make a team procedure run the same way every time, Automate something on a site that has no API

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

Documented approaches

1. Teach the model a procedure once and reuse itMediumDocumented

   Metadata always loaded, instructions on trigger, reference files only when read. Costs almost nothing until it is needed.

2. Work through an entire repository, not a pasted fileMediumDocumented

   Pair it with hooks so the non-negotiables run automatically instead of being restated.

Related tasks: Let a model use your systems, not just your prompts

Why this comes up: Most bills are large because the same context is paid for repeatedly, not because the model is expensive.

Documented approaches

1. Stop paying twice for the same fileQuickDocumented

   First move, always. Reads cost a tenth of base input, and the usage block tells you whether it worked.

2. Run the whole job overnight at half priceQuickDocumented

   Second move: anything with no human waiting on it goes asynchronous at half price.

3. Run a capable model on your own machineMediumPartly documented

   The structural answer for high-volume, low-difficulty work - if you have the memory for it.

Related tasks: Use a model without sending anything anywhere

Why this comes up: A model that describes your data is guessing. A model that runs the numbers is not.

Documented approaches

1. Let the model write and run real code mid-answerQuickDocumented

   Upload the file, ask for the chart and the code that made it, and check the code.

2. Get JSON that always matches your schemaQuickDocumented

   When the analysis feeds another system rather than a slide.

Related tasks: Get clean data out of a stack of PDFs

Why this comes up: An hour of audio is unsearchable. An attributed transcript is evidence.

Documented approaches

1. Transcribe a recording and label who said whatQuickDocumented

   Split the file under 25 MB, pick the model by whether you need speakers or word timestamps, then summarise in a second pass.

Related tasks: Build something people can talk to

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

Documented approaches

1. Talk to it out loud and hear it answerDeepDocumented

   One session, audio in and audio out. Pick the transport by where the audio lives, and design interruption before the prompt.

2. Transcribe a recording and label who said whatQuickDocumented

   For the record afterwards rather than the conversation itself.

Related tasks: Turn a recording into something you can search and quote

Why this comes up: Legible text was the thing image models could not do, and it is the thing every marketing asset needs.

Documented approaches

1. Generate an image with legible text inside itQuickDocumented

   Quote the exact copy in the prompt, attach brand references, and correct in the same conversation.

2. Generate video from a prompt or a stillMediumPartly documented

   For motion. Start from a still you already like - composition control is the hard part.

Why this comes up: Some documents cannot leave the building, and some places have no reliable connection.

Documented approaches

1. Run a capable model on your own machineMediumPartly documented

   Start from the memory arithmetic, not the leaderboard. Check the fit label before downloading anything.

2. Stop paying twice for the same fileQuickDocumented

   If privacy is really about cost, this is cheaper than buying hardware.

Related tasks: Cut the API bill without changing the answer

Why this comes up: The tasks worth automating are the ones that take hours, and you cannot sit and watch them.

Documented approaches

1. Hand off a task and walk away from the machineMediumDocumented

   Cloud sessions, scheduled routines, and handoff between terminal, desktop and phone.

2. Run the whole job overnight at half priceQuickDocumented

   When the work is many small independent requests rather than one long task.

Related tasks: Understand a codebase nobody has explained to you