HABv4 Token-Efficient Coding Agent
Activation
This skill activates automatically for all tasks in this repository.
Instructions - Strict Workflow
- •
Plan First (always):
- •Create a brief 3-6 bullet plan.
- •Decide which files/directories are truly needed.
- •Summarize relevant files before full read.
- •
Progressive Context Loading:
- •Use
ls,tree,grep -ror partial reads first. - •Only request full file content if the summary confirms it's required.
- •Maintain and update a Persistent Session Summary (intent, decisions, changes, next steps).
- •Use
- •
Model & Reasoning:
- •Default: Fast model + low/medium reasoning for analysis and small edits.
- •Switch to high-reasoning only when generating patches or modifying boot logic.
- •
Output Format (Token-Saving):
- •Show only changed code + minimal explanation.
- •Use concise diffs when possible.
- •End every response with:
Current Session Summary: [2-4 bullets]
Success Criteria
- •Task completed with ≤ 40-60% of baseline token usage.
- •All changes verified via
diagnose_iso_repodata.shor equivalent. - •Persistent summary updated.
Never
- •Load entire directories or large files without summarization.
- •Give long explanations unless explicitly asked.