Controlling the context a coding agent carries into a task

Why context costs twice

Claude Code's cost documentation states that token costs scale with context size, and that a session sends the full conversation with every request. Prompt caching makes the repeated part cheaper, not free, and a first message after the cache expires reprocesses the whole context.

The second cost is quality. Anthropic's article on context engineering describes context rot: as the number of tokens grows, a model's ability to recall information from that context decreases. It recommends the smallest set of high-signal tokens that makes the desired outcome likely. More instructions can mean worse adherence to the ones that matter.

Know what loads before the first prompt

Start with an inventory, per tool. In Claude Code, CLAUDE.md files, auto memory, MCP tool names and skill descriptions load before you type, and /context shows a live breakdown. MCP tool definitions are deferred by default, so only names and server instructions enter context until a tool is used. In Cursor, always-applied rules sit at the start of the context, a CLAUDE.md is always applied, and a file referenced with @ in a rule is read only when needed, not inlined.

Sort every instruction into a tier

• Always loaded: constraints that apply in every session and that the agent cannot infer from the code, such as build and test commands, paths it must never edit and security rules. Keep this layer short.
• Path-scoped: conventions for one package or file type, in nested instruction files or rules with paths or globs that load when the agent touches matching files.
• On demand: procedures such as writing a migration or preparing a release, as skills that load only when invoked or judged relevant.
• Nowhere: anything the agent can read from the code itself, and anything a linter or CI already enforces.

A constraint that must hold regardless of context belongs in permissions, hooks or CI as well. Claude Code treats instruction files as context, not configuration, so a cut that drops a rule from context should never be the only thing between the agent and a protected path.

The trap in long sessions

Moving instructions to on-demand scopes changes when they are present. Claude Code's documentation spells out what happens at compaction: the project-root CLAUDE.md and unscoped rules are re-injected from disk, but path-scoped rules and nested CLAUDE.md files are summarised away and return only when a matching file is read again. Invoked skill bodies come back capped at 5,000 tokens per skill and 25,000 in total, keeping the start of each file.

So a rule that must hold for a whole session should not be path-scoped, and the most important lines of a skill go at the top. The documentation says the same: if a rule must persist across compaction, drop its paths or move it to the root file.

Levers beyond instruction files

• Filter tool output before the agent sees it. Claude Code's documentation shows a hook that reduces test output to failures.
• Send verbose work such as log analysis to a subagent, which returns a summary. Its own requests still draw on usage, so this shrinks the main context but does not necessarily lower the bill.
• Disable MCP servers nobody uses, and prefer a CLI where one exists, which the documentation describes as more context-efficient.
• Clear the session between unrelated tasks, since old conversation is resent with every message.

Measure the cut on the same tasks

A hypothetical check for one repository: replay 20 merged tasks with the current setup and with the trimmed one, three runs each. Every run starts in an isolated copy whose history ends before the change, with no later refs, remotes, caches, memory or files from earlier trials and no network route to the merged solution. The agent sees the existing tests; the acceptance tests added with each change are held out for grading. Record cost per passing task, the held-out pass rate and a scripted count of constraint violations, such as edits to generated paths. Adopt the trimmed setup if violations do not rise and the pass rate stays within a threshold agreed beforehand. If the sample is too small to tell, the result is inconclusive and the current setup stays.

LLM cost optimisationKeeping Claude Code team instructions useful