Prompt Slimmer

Audit and slim down OpenClaw workspace files (SOUL.md, AGENTS.md, TOOLS.md, IDENTITY.md, USER.md, HEARTBEAT.md, MEMORY.md) to reduce system prompt token usag...

Audit and optimize OpenClaw workspace files to reduce system prompt token overhead without losing information.

Why This Matters

Every workspace file (SOUL.md, MEMORY.md, etc.) is injected into the system prompt on every API call. A 50K-char workspace means ~12K tokens sent with every single message — even a simple "hi". At Opus pricing ($15/1M input tokens), that's $0.18 per message just for workspace overhead.

Quick Start

Run the audit:

1. Measure all workspace files (chars + lines)
2. Analyze each section for frequency-of-use
3. Identify redundancies across files
4. Generate slim-down plan with before/after estimates
5. Execute (with user approval) and verify

The Method: Frequency-Based Layering

Layer 1: Always-On (workspace files)

Content needed in every session: core identity, active relationships, behavioral rules, critical safety constraints.

Layer 2: Searchable (memory/archive/)

Content needed sometimes: completed projects, historical records, detailed technical specs, one-time learnings. Retrieved via memory_search when relevant.

Layer 3: Skill-Embedded

Content needed only for specific tasks: workflow steps, code templates, platform-specific guides. Lives in SKILL.md files, loaded only when the skill triggers.

Audit Procedure

Step 0: Ghost File Scan (Often the Biggest Win!)

OpenClaw injects all .md files in the workspace root into the system prompt — not just the standard files. Scan for "ghost files": old reports, task materials, research notes, or temp files sitting in the workspace root.

cd <workspace_dir>
echo "=== All .md files in workspace root ==="
for f in *.md; do
  [ -f "$f" ] && echo "$f: $(wc -c < "$f") chars"
done
echo "=== Standard files ==="
echo "SOUL.md AGENTS.md TOOLS.md IDENTITY.md USER.md HEARTBEAT.md MEMORY.md BOOTSTRAP.md GARDEN.md"

Any .md file NOT in the standard list is a ghost file candidate. Move completed task files, old research, and temp notes to memory/archive/ or a subdirectory (subdirectories are not injected).

Real example: One instance had 15K in standard files but 31K in two ghost files (an old promotion review + research report) — removing them saved 63% instantly.

Step 1: Measure Standard Files

cd <workspace_dir>
for f in SOUL.md AGENTS.md TOOLS.md IDENTITY.md USER.md HEARTBEAT.md MEMORY.md; do
  [ -f "$f" ] && echo "$f: $(wc -c < "$f") chars, $(wc -l < "$f") lines"
done

Step 2: Section-Level Analysis

For each file, extract ## headings and measure each section's size. Then classify:

Category	Criteria	Action
🟢 Keep	Referenced every session, identity-critical	Keep as-is
🟡 Slim	Useful but verbose, can be compressed	Rewrite concisely
🔴 Archive	Completed/paused projects, historical records	Move to memory/archive/
⚫ Deduplicate	Same info in multiple files	Keep in one place, remove others

Step 3: Cross-File Deduplication Check

Common duplication patterns:

• IDENTITY.md ↔ MEMORY.md (appearance, voice info)
• SOUL.md ↔ MEMORY.md (behavioral rules)
• HEARTBEAT.md ↔ AGENTS.md (task scheduling rules)
• MEMORY.md ↔ Skill files (project details duplicated in both)

Rule: Information lives in the most specific location. If a skill covers it, remove from MEMORY.md.

Step 4: Execute Slim-Down

1. Create archive file: memory/archive/projects.md (or topic-specific files)
2. Move archived sections: Cut from workspace file → paste to archive
3. Replace with pointer: Details: see memory/archive/projects.md or just a 1-line summary
4. Verify memory_search: Confirm archived content is findable via search
5. Measure result: Compare before/after char counts

Step 5: Verify Integrity

After slimming, verify:

• memory_search can find archived content
• No critical behavioral rules were accidentally removed
• Core identity (name, relationships, key rules) still present
• Safety constraints still in workspace files (not just archive)
• Pointers/references are correct paths

File-Specific Heuristics

MEMORY.md (Usually the biggest win)

Typical bloat sources:

• Paused/completed project details → archive, keep 1-line status
• Detailed timelines/changelogs → archive
• Team roster tables → archive
• Cron job indexes → cron list can fetch this live
• Platform account details → slim to name + ID only
• Milestone lists → archive (historical)
• Cross-referenced content → remove if covered by skill or other file

Target: MEMORY.md should be < 5,000 chars for a well-maintained instance.

HEARTBEAT.md

Typical bloat sources:

• Code templates (osascript, shell snippets) → agent already knows these
• Verbose priority descriptions → compress to 1-2 lines per priority
• Paused project references → 1 line max
• Repeated emphasis ("核心！", "每次 HB 必做！") → once is enough

Target: HEARTBEAT.md should be < 3,000 chars.

SOUL.md

Be careful here. SOUL.md is identity-critical. Don't optimize away personality.

• Hierarchy diagrams → can be compressed
• Behavioral rules → review for overlap with AGENTS.md
• Generally leave SOUL.md alone unless it's > 10,000 chars.

TOOLS.md

Usually already lean. Check for:

• Deprecated tool entries → remove
• Detailed port tables for rarely-used services → slim or archive

AGENTS.md

Usually already lean. Check for:

• Redundancy with system prompt (OpenClaw injects its own rules)
• Over-detailed workflow descriptions → reference skill instead

Expected Results

Workspace Size	Before	After	Typical Savings
Light (<20K)	20K	15K	25%
Medium (20-50K)	35K	18K	50%
Heavy (50K+)	60K	22K	63%

Cost Impact

At Opus pricing ($15/1M input, $1.50/1M cached):

Savings	Chars Saved	Tokens Saved	$/message saved	$/day (100 msg)
25%	5K	~2K	$0.03	$3
50%	25K	~10K	$0.15	$15
63%	38K	~15K	$0.23	$23

With prompt caching, savings are ~90% less but still meaningful for cache-miss turns.

What NOT to Slim

• Safety constraints (never archive security rules)
• Core identity (name, key relationships, personality)
• Active behavioral rules that prevent known failure modes
• Credentials management rules (how to handle secrets)
• Cross-instance coordination rules (if running multiple agents)