token-optimization

name	token-optimization
description	Rules to minimize token usage

Concise responses

State the result. Skip the process. Omit what the user already knows.

• One short paragraph or a few bullets. Not both.

Cap output

Every command can produce unbounded output. Always limit it.

• Redirect to file first, then extract: cmd > /tmp/out.txt 2>&1 → grep/head/tail → rm
• Pipe to head -n or tail -n. Use --quiet, --short, --oneline.
• Must cap: find, ls -R, cat on unknown files, test suites, git log, builds.

Search before read

Never read entire files. Locate the relevant lines first.

• glob → grep -n → sed -n '30,50p' → stop.
• Narrow: filter by directory, file type (--include='*.go'), match count (-m 10), or grep -q for boolean checks.

Extract fields, never dump

Do not dump raw JSON/YAML/XML.

• Use jq '.key', awk '{print $2}', cut -d: -f1.
• For build/test output: grep -m1 'ERROR\|FAIL' + tail -30.

Diff before full read

Use git diff -- file first. Read the full file only when diff lacks sufficient context.

No re-ingestion

Check context before any read. Do not re-read files, re-run searches, or verify what is already known.

Batch calls

• Run independent reads/searches in parallel within one turn.
• Chain dependent commands: cmd1 && cmd2.
• Combine patterns: grep -n 'pat1\|pat2' file.

Surgical edits only

Use targeted replacements (sed -i '' 's/old/new/g'). Never rewrite whole files.