Bob is IBM's AI coding assistant, and a session is one continuous conversation with it. You don't spend Bobcoins on the work you get done. You spend them on the text Bob has to read and create. The difference between the two is where most waste comes from.
The goal is not to spend less. A large refactor costs a lot and is worth it. The goal is to stop paying for work that isn't worth it.
In this blog, you will learn how to get more out of every Bobcoin by managing context windows, avoiding context poisoning, writing effective prompts, configuring Bob to cut recurring costs, and measuring where your spend goes.
Why long context windows cost more
Every time you send a message, Bob receives the entire conversation again, not just your latest message. A message sent an hour into a context window therefore costs more than the same message you sent at the start. The longer the conversation, the more Bob has to read, and the more Bob has to read, the more Bobcoins you spend.
There is a second effect, and it is possibly even more expensive. The longer a conversation gets, the harder it is for Bob to work with. Answers become less accurate. Less accurate answers need correcting. Each correction makes the conversation longer, which makes the next answer worse. Cost rises and quality falls at the same time.
For example, a long debugging session where Bob re-reads the full history on every turn can end up costing more than the same fix done across a few focused, fresh context windows.
- Learn how to start and manage context windows
- Learn how to avoid context poisoning
Cost saving strategies
Begin a new context window for each unrelated task. Once you complete a task, start a new context window. If you work from the same context window for your next task, Bob still reads all of it every turn. Don't split so often that Bob has to re-learn background you already gave it — that re-explaining costs more than the long conversation would have.
- Learn how and why to create a new context window
State the scope of the request. Use the @ mention as much as possible to point Bob at specific files. Without context mentions, Bob searches the repository to figure out what matters, and that search can cost a lot of Bobcoins.
- Learn how to use context mentions and see best practices for using them.
Write effective prompts. A prompt is a request to Bob. The more precise it is, the less Bob has to guess, and the less you have to correct.
- Learn how to write effective and cost-efficient prompts
Roll back rather than correct repeatedly. If Bob has misunderstood your prompt multiple times, you should either roll back or start fresh with your original request plus what you have learned. Rolling back or starting a new context window can reduce the cost of a long conversation with many corrections.
With a rollback you can:
- Safely experiment with AI-suggested changes
- Easily recover from undesired modifications
- Compare different implementation approaches
- Revert to previous project states without losing work
- Learn how to rollback
- Learn best practices for rollback
Configure Bob to reduce cost
AGENTS.md. Write down your project's conventions and structure so Bob does not have to work them out again in each context window. Keep your AGENTS.md concise. Bob reads the AGENTS.md file constantly, so a bloated file turns a one-time win into a permanent cost.
- Learn to use
/initto tell Bob to create an AGENTS.md file for your project. - Create custom rules for each of your modes.
Skills. A skill holds domain knowledge that loads only when needed, instead of taking up space in the context window. You pay for it when you use it.
- Learn about Skills.
- Learn how to create and use skills in Bob.
Hooks. A hook runs a linter, formatter, or test command directly and costs no Bobcoins. Asking Bob to do the same action of a hook costs Bobcoins. If a tool can do the check, don't pay a model to do it.
- Learn about lifecycle hooks.
Connected tools. Bob reads the description of every connected MCP (Model Context Protocol) server on every request, whether you use those tools or not. A server you switched on for one task last month is still costing you. Turn off MCPs when you're not actively using them.
- Learn about MCP servers
- Learn how to create an MCP server and connect it to Bob.
Save Bobcoins on complex tasks
Plan before you build. A plan is a list of steps to complete a task. It can be a bulleted list, or as complex as a full design document. A plan reduces the number of corrections you need to make during the build phase, and you can reuse it for similar tasks. Once you and Bob create a plan, review it before you implement it.
- Learn how to create a plan and implement complex features.
- Learn best practices for creating plans.
Subagents. Bob assigns work to subagents automatically, so you don't need to ask. You can also directly assign work to a subagent when a task needs heavy research: the digging happens outside your main context window, and only the findings come back. Subagents free up room in your context window, but they do not always cost less, and running several at once can cost more. Reach for a subagent when a task needs heavy, isolated research; keep short, sequential tasks in your main context window instead — coordinating several agents can cost more than doing the task directly.
- Learn about subagents
Measure Bobcoin usage
Bobalytics shows what you actually spent. The total tells you little on its own. What you want to spot is waste: long context windows on small tasks, the same correction made repeatedly, and background you type out every morning that belongs in AGENTS.md.
- Learn how to see your Bobalytics
- Learn how to monitor your Bobcoin usage
Conclusion
None of this limits what you can attempt. It moves Bobcoins away from re-explaining your project, fixing avoidable misunderstandings, and paying for context you never needed, and toward the work itself.
More Bobcoins increase what you can spend. Less waste decides how much of that reaches the actual problem.
