Context Pack

Context Pack

Token cost is dominated by what enters context, not by what you do with it. This skill is about reading less while knowing more.

Read known paths, search for unknown ones

A role usually needs specific artifacts — specification.md, architecture.md, tasks.md, analysis.md. Read those directly; there is nothing to discover.

Reach for search only when the question is "where is X". Then make the search structural rather than textual.

ast-grep for code questions

ast-grep matches syntax, not strings, so it returns the construct you asked about instead of every line mentioning it.

# Every class implementing an interface, without reading any file whole
ast-grep --lang csharp -p 'class $NAME : $$$ IResourceRenderer $$$'

# Every call to a method, with its arguments
ast-grep --lang csharp -p 'RegisterRenderer($$$)'

# Every method with a given attribute
ast-grep --lang csharp -p '[Test] $RET $NAME($$$) { $$$ }'

Always invoke ast-grep by its full name. On Linux, sg is util-linux's setgid binary; a script that calls sg expecting ast-grep runs the wrong program silently.

Fall back to rg for prose, configuration and markdown, where there is no syntax tree worth matching.

Delegate what you do not need verbatim

Spawn a cheap-tier subagent for anything whose output you need conclusions from rather than content:

• "Where is the display-name logic for azurerm?" — the answer is a path, not a file.
• Build and test runs where only pass/fail matters.
• "Does anything else already do X?" before writing something new.

Do not delegate a single file read or one rg — the spawn costs more than the call.

Command output is compressed for you

rtk sits in front of Bash calls and compresses the output of git, test runners, linters, gh and docker before it reaches context. You do not invoke it; the hook rewrites the command.

Two consequences worth knowing:

• Its compression is lossy — dedup, truncation, grouping. Where you need the exact bytes, say so; tee mode preserves full output on failure.
• The Code Reviewer's diff is excluded, because a truncated diff produces a confident review of code nobody saw.

What not to do

• Do not cat a 300-line file to answer a one-line question.
• Do not read a file "for context" before knowing what you are looking for.
• Do not paste a role file into a subagent prompt — the subagent reads it itself.