Text
| Task | Tier | Reasoning | The catch |
|---|---|---|---|
| translate, summarize, “make it shorter” | small or mid-tier | not needed | translates terms differently in different places — give it a glossary |
| an email, a post, a description from your facts | mid-tier | low | adds facts of its own — ask it to flag anything that wasn’t in the materials |
| editing someone’s text to a style | mid-tier | low | a style rule plus two “before and after” examples (An example over a description) beat a description |
| proofreading: typos, agreement | mid-tier | low | changes meaning along with the typos — ask for a list of edits, not a new text |
| summarizing a long document | mid-tier | medium | loses the middle — ask for a summary by section with page numbers |
Code and configs
| Task | Tier | Reasoning | The catch |
|---|---|---|---|
a 50-line script, a regex, jq, a one-liner |
mid-tier | low | mixes up GNU and BSD flags — state the OS and version (Put the date, versions and environment in the prompt) |
| explain someone’s code or config | mid-tier | medium | explains from the names, not the code — ask for line references |
| find a bug from a trace | flagship | medium or high | a trace without code leads to guessing — attach the file where it failed |
| a change across several files | flagship in an agent | high | edits too much — commit before the run and git diff after (Clean git status before the agent) |
| architecture, a migration, a week’s plan | flagship or the strongest | high or max |
agrees with your idea — ask for two options and the downsides of each (Review in a new chat) |
Dockerfile, compose, traefik.yml |
flagship | medium | keys from old versions — give the program version and a piece of documentation |
Analysis and extraction
| Task | Tier | Reasoning | The catch |
|---|---|---|---|
| a 5,000-line log: what broke | mid-tier | medium | latches onto the first error rather than the cause — ask for a timeline |
| an email, an invoice, a contract — fields into JSON | small | not needed | invents a field that isn’t there — a schema with null for empty (JSON by schema) |
| label thousands of records by category | small, batched | not needed | borderline categories — an example for each, one-word answers (Non-urgent API requests go as a batch at half price) |
| compare two documents | flagship | medium | misses small differences — run diff first and give the model only the meaning |
| a table or a column of numbers | — | — | leave the arithmetic to a script; the model writes the script or the formula |
Images, PDFs, audio
| Task | What to use | The catch |
|---|---|---|
| a screenshot of an error or UI | any model that takes images | misreads small text — better to copy the error text |
| a PDF with scans and tables | a flagship, the whole PDF | page numbers in the answer — check them against the file |
| a diagram, a chart, a photo of a whiteboard | a flagship | exact numbers off a chart are approximate |
| draw a picture | an image generator, not a language model | text in images and fine details are the weak spot |
| transcribe a recording | a speech recognition model (Whisper and the like) | names and terms — give a vocabulary hint |
Finding facts
| Task | What to use | The catch |
|---|---|---|
| a version, a price, a release date | a model with search, an answer with links | without search — the knowledge cutoff (Put the date, versions and environment in the prompt) |
| “how to set up X properly” | the official documentation in the context | generic advice from the internet five years old |
| compare three tools | a flagship with search | the one people wrote about most wins the comparison |
When a higher tier doesn’t help
- The data isn’t in the context. A strong model guesses as confidently as a weak one; the cure is materials, not a tier.
- The task is formal: a sum, a sort, a count. A script gives the same correct answer on a million lines; a model gives a plausible one.
- The prompt is ambiguous. A strong model will guess a different reading, but it still guesses — clarify the task (Questions and a plan first).