Best practices
These guides collect what works when you use Context day to day. Each one links to the reference pages for the exact steps.
What these guides cover
| Guide | Read it when |
|---|---|
| Write good task instructions | Results come back generic, off target or in the wrong format |
| Build and improve skills | You keep giving the same instructions, or a skill’s output varies |
| Choose a model | You want to trade speed, cost and quality for a kind of task |
Pick the right tool
Start with a task. Turn it into something more only when the need repeats. Choose between tasks, skills, automations, agents and applets compares every option and follows one job as it grows.
A few quick signals help inside a single task:
- You can say exactly what “done” looks like. Use Goal mode, so the agent keeps working until it gets there.
- A wrong first step would be costly, such as changes to many files. Use Plan mode, so you see the plan before the agent acts.
- You want to try another direction without losing this one. Fork the task.
- You typed the same rules into your last three tasks. Move them into a skill.
- The rules describe who the agent is and what it may reach, not how to do one kind of job. Put them in an agent’s Instructions instead.
Habits that pay off
- Write the brief once, well. A clear goal, reader and source list saves more time than any correction later.
- Correct in the same task. The agent keeps the context it has built up. A new task starts from nothing.
- Check the result against your own test. If you can’t say how you would judge the output, the agent can’t either.
- Measure before you change a skill that many people use. Evals let you compare versions on the same examples.
Next steps
- Write good task instructions: the parts of a good brief, and how to steer and iterate.
- Build and improve skills: when to write a skill, how to test it, and how to improve it.
- Choose a model: pick a model for the job.
- Choose between tasks, skills, automations, agents and applets: the full comparison.