context-efficient-navigation

name	context-efficient-navigation
description	Use token-efficient codebase/log navigation workflows. Load when inspecting large repos, logs, generated files, unfamiliar code, structural relationships, or any task where broad file reads would waste context.

Context-Efficient Navigation

Use this skill when accuracy needs targeted evidence, but full-file reads or broad dumps would waste context.

Core Rule

Save tokens without skipping needed verification. Prefer narrow, evidence-first inspection.

Search-First Workflow

1. Start with targeted discovery:
   • rg --files
   • rg "literal text"
   • find, fd, ls, tree when useful
2. Read only relevant slices:
   • sed -n 'START,ENDp' file
   • head, tail
   • wc -l
3. Expand only when needed:
   • nearby lines
   • callers/callees
   • adjacent modules
   • tests covering same behavior

Logs and Generated Artifacts

Large or fast-changing files need narrowing first.

Prefer:

• rg "ERROR|WARN|panic|Traceback" log
• tail -n 200 log
• head -n 80 log
• sed -n 'START,ENDp' log
• ls -lh to spot huge files
• wc -l to estimate read cost

Avoid dumping whole logs, lockfiles, generated bundles, snapshots, coverage reports, or minified files unless the whole artifact is genuinely needed.

Structural Code Questions

Use structure-aware tools before generic text search when asking about:

• callers / callees
• call chains
• inheritance
• overrides
• symbol ownership
• module dependencies
• dead-code triage
• complexity hotspots

Prefer cgc for these if available. Load/use the codegraphcontext-cli skill for exact commands and caveats.

Use direct text search first for:

• exact strings
• small unambiguous matches
• simple literal lookups
• cases where cgc is unavailable or too noisy

Large Retrieval Tasks

When answer depends on finding a few relevant sections among many candidates:

• prefer available indexing/search workflows
• avoid repeated broad reads
• extract bounded candidate snippets first
• then reason from those snippets

Delegation

When a large body of text needs triage or summarization, delegate a bounded extraction/summarization task to a lightweight, low-cost sub-agent if it reduces main context and does not block critical reasoning. Prefer agents/models configured with low thinking effort for extraction-only work; reserve high-cost reasoning for synthesis or decisions.

Response Discipline

When reporting findings:

• cite exact file paths and line ranges when possible
• summarize only relevant evidence
• mention search commands if they materially affect confidence
• state uncertainty when evidence is partial