The agent does what it understood. A well-stated task saves both time and tokens.
What to write
| Technique | Example |
|---|---|
| Name the chart | "on BTCUSDT 5m on the Work tab" instead of "on the chart" |
| State conditions in numbers | "volume twice the 20-bar average" instead of "big volume" |
| Say what must not change | "don't change the charts", "leave my old tabs alone" |
| Ask to see it first | "don't do anything yet", "check it without saving" |
| Ask for the source | "tell me where you got the levels" |
| Ask it to compute with code | "compute it with code", when it comes to statistics |
| Ask it to list assumptions | "list everything you assumed" |
| Say what to do with a small sample | "if there are fewer than 30 trades, say so plainly" |
New chat or this one
A new task works better in a new chat. In a long chat, the agent may pick up something you said earlier and start applying it to the new task without asking.
Follow-ups and "undo", on the other hand, belong in the same chat: there the agent remembers what it did.
Skill or prompt
| What you have | What to use |
|---|---|
| A one-off task | Prompt |
| Repeated work in a fixed format | Skill |
| The skill did not save | Paste its text at the start of the chat, and the agent follows it in that conversation |
Don't ask for data you don't need
Long lists of bars use up the most tokens and slow down the answer. Ask only for the period you need, and don't ask for the whole table in the chat if you only need the total.