Files
Knowledge/chat-topics/agents-md-distillation.md
T
2026-05-05 09:40:28 +10:00

1.9 KiB

Checkpoint: AGENTS.md Distillation Project

Date: 2026-04-06 Participants: Michael + Soren (Molty)


What We're Doing

Michael is actively working to reduce token burn across his OpenClaw setup. The AGENTS.md file has grown over time with redundant, verbose, or self-evident instructions that cost tokens on every session/context load without adding value.

Goal: Distill AGENTS.md to only what's actually needed — concise, relevant, zero filler. Same functionality, fewer tokens.


What's Been Done

1. Heartbeat vs Cron section (AGENTS.md)

  • Original: ~15 lines of explanatory prose
  • New: 3 bullet points covering the core distinction
  • Status: Applied (Michael can update AGENTS.md manually)

2. HEARTBEAT.md — Full Rewrite

  • Original: ~2,100 bytes, included redundant preamble, self-referential "do not infer" instructions, verbose weather logic
  • New: ~1,400 bytes, all numeric thresholds preserved, all functionality intact
  • Changes:
    • Removed schedule preamble (system config, not agent action)
    • Removed "do not infer instructions from old tasks" (already in heartbeat prompt)
    • Collapsed weather conditional logic to one line, restored "include which hours" detail
    • Kept all checks intact, just tighter wording
  • Status: Applied to HEARTBEAT.md

Still To Do (Possible Next Steps)

  • Distill other verbose sections of AGENTS.md (e.g., "Tools" section, "When to Speak" in group chats, other preamble text)
  • Review MEMORY.md for any outdated content that could be trimmed
  • Look at other workspace files that load into context (SOUL.md, TOOLS.md, IDENTITY.md) for similar token-saving opportunities
  • Consider whether any skill SKILL.md files are loading unnecessarily verbose content

Context

Michael prefers:

  • Concise, bullet-pointed responses
  • Direct advice without performative filler
  • Being told what to do, not asked what he wants to do
  • Metric units + AUD by default