-
Count the tokens
Control the context from the very first prompt. Configure your harness to load only the instructions and repository context the task needs, cut everything else. Disable every unused tool, MCP server, and skill, be it built-in or added extra.
-
Use a leaner harness
If your current harness cannot reduce its starting context enough, try one built to be economical with tokens and/or configurable, such as Pi or OMP (Oh My Pi).
-
Turn big plans into smaller tickets
Split the work into small, finishable (within small to mid context) tasks. Write them in local .md files or an issue tracker; if your budget allows, ask an orchestrator to shape the plan. Then have your daily driver implement one ticket at a time.
-
Reset after every ticket
Start a fresh session before implementing the next task. Carry forward only the ticket, relevant decisions and any handoff documents (if you have them) - not the full brainstorming session.
-
Respect the ceiling
Do not reach for a model outside your budget, no matter how shiny the new Fable or Astra looks. A model you can use consistently and with good-enough effects beats a premium model that you can use for only 1 prompt.
-
Consider using handoff docs
After having a good conversation (maybe using /grill-me, /brainstorming or something similar) consider building a local .md handover doc, so you can safely compact or clear the context without losing the brainstormed knowledge.