coaching
The agent field manual.
Nine short guides on getting more out of coding agents. Each one maps to a coaching insight PromptConduit computes from your own sessions — the editor extension links here when it spots a pattern worth fixing.
Direct the work
- Use plan mode for non-trivial work
Planning before editing catches wrong approaches before they cost tokens.
- Reduce interruptions
Stopping the agent mid-task burns tokens and breaks its plan. Front-load context instead.
- Adopt skills and slash-commands
Capture repeatable work as skills instead of retyping long prompts.
Scale the work
- Batch your tool calls
Grouping independent reads and edits cuts round-trips and output tokens.
- Delegate with subagents
Parallel agents cover more ground and keep your main context clean.
- Isolate risky work in a worktree
Experiment freely on a branch without touching your main checkout.
Cut the waste
- Improve your cache hit rate
Reuse context across turns instead of re-pasting it — cached reads are ~10× cheaper.
- Raise your tool success rate
Failed tool calls waste a full round-trip each. Give the agent what it needs to get them right first time.
- Right-size your MCP servers
Connect the MCP servers you actually use and prune the rest.