Frame the training around output quality
The framing of credit optimization training determines how people behave afterward. If the training is positioned as cost control, people hesitate before using useful agents, builders avoid ambitious workflows, and managers treat high consumption as suspicious before understanding whether the workflow is valuable. Position it instead as workflow quality: getting better results from better-designed interactions with agents.Teach prompting defaults before teaching the dashboard
Most employees do not need to understand tokens, model pricing, or retrieval mechanics. They need a simple mental model: AI consumption is shaped by workflow design. A vague request makes the agent do more guessing. A broad agent may retrieve more noise. A long conversation can carry irrelevant context. A deep workflow does more work than a lightweight first pass. An automation multiplies whatever design choices were made before it launched. Once people understand that, credit optimization becomes a question of shaping the work better — not spending less.What to teach
For new users, teach concrete prompting defaults.
These prompts produce better output. They tell the agent what work matters, where to look, how much depth to use, and when to stop. See How to write effective instructions for more guidance.
Simple user guidance
Teach when to go deep
Deep workflows produce high-value output. They also consume more credits by design. A deep dive should be treated like sending a senior analyst into a problem — not like asking for a quick summary. It is the right choice when the question is complex, the answer depends on multiple sources, the stakes are high, or the user needs synthesis rather than retrieval. Go Deep gives agents the ability to conduct comprehensive research across company data, databases, and web sources. It includes sub-agents, company data exploration, data warehouse queries, and web search and browsing.Good reasons to use deep research or a stronger model
- Preparing for a high-stakes customer meeting.
- Investigating a complex support issue across multiple sources.
- Synthesizing conflicting evidence.
- Drafting executive-ready recommendations.
- Comparing strategic options with tradeoffs.
- Turning a messy set of inputs into a clear decision memo.
Less suitable use cases
- Summarizing a short document.
- Rewriting a message.
- Formatting notes.
- Classifying a list.
- Extracting fields from a document.
- Asking for a first-pass opinion.
The habit to train
First pass first. Deep dive second. Start lightweight, then escalate only if the first pass shows the problem deserves it. A stronger model is a tool for a class of work — if the task is simple, it may produce a slightly more polished answer at a higher cost without changing the business outcome. See What settings and model should I use? for guidance on model selection.Teach when to use a skill, an agent, or a chain
As teams get more advanced, users start chaining things manually. They ask one agent to search, then another to summarize, then another to rewrite, then paste the output into a new conversation for a final draft. Sometimes this is useful exploration. Often, it is a sign that the workflow should be formalized.Three modes to distinguish
Simple rule
If you chain it twice, consider turning it into a workflow. If a team chains it repeatedly, formalize it. Manual chaining often hides cost. Each step can add context, trigger tools, use a strong model, or create another long conversation. A reusable workflow can be designed more carefully: lighter models for simple steps, better source scoping, clearer tool usage, and a cleaner output. Skills in Dust are reusable packages of instructions, knowledge, and tools that you can share across multiple agents. When you update a skill’s instructions, every agent using that skill automatically gets the improvement.Retrain existing users on better habits
Existing users already have habits, and many of those habits were rational when they formed. They use the general agent because it worked. They paste documents because they want to be sure the agent has context. They reuse long conversations because it feels convenient. They ask for deep research because they associate depth with quality. The goal is to help them refine those habits as usage scales.What to do
Use real examples from your workspace to show better alternatives:
People learn faster from patterns than from policy.
Train builders with a deeper curriculum
End-user habits matter, but builder habits scale further. A user can make one conversation inefficient. A builder can make hundreds of future runs inefficient by choosing the wrong model, attaching too many sources, giving an agent too many tools, or leaving instructions too vague.Builder training checklist
For builders, credit optimization is product design. Every agent has a usage profile. Every model choice, data source, tool, and instruction changes that profile. See How to write effective instructions, model settings, and agent troubleshooting for details.
Builder best practices
- Use lighter models for structured or repetitive tasks when the output remains reliable.
- Connect only the sources that matter for the agent’s job.
- Remove tools that are not central to the workflow.
- Write clear, specific instructions — vague instructions make agents do more guessing. See How to write effective instructions.
- Test model changes on real examples before changing an important workflow.
- Use skills to share instructions and tools across multiple agents.
Help managers focus on workflow quality
Managers set the tone for how the team approaches credit optimization. If they frame it as cost control, people become cautious. If they frame it as workflow quality, people become more intentional about how they use agents.Better manager questions
These questions keep the conversation focused on work design and make optimization collaborative.