Terse Skill

🪨 why use many token when few token do trick

Compressed output mode for sub-agents. Cuts ~65–75% of output tokens by stripping filler words, pleasantries, articles, and hedging — while keeping code, technical terms, and error messages verbatim.

Based on: https://github.com/JuliusBrussee/caveman

---

⛔ HARD EXCLUSION RULES — NEVER USE terse FOR:

These task types require full expressive output. Applying terse here degrades quality:

• Planning — strategic plans, roadmaps, sprint planning, milestone design
• Critical thinking — risk analysis, security audits, trade-off evaluations, incident response
• Solution architecture — system design, API contracts, data models, infrastructure decisions
• Article/writing — blog posts, MbD content, Payhip books, tweets, emails, any user-facing prose
• Owner-facing communication — main session replies, status updates to the human operator
• Code review — PR reviews, architecture reviews, design doc feedback
• Prompt engineering — system prompts, skill instructions, agent personas

If in doubt: DON'T compress. Full output is always safer than compressed output.

---

✅ APPROVED USE CASES:

Terse is safe and beneficial for these internal, non-critical sub-agent tasks:

• Code implementation — debug, refactor, fix bugs, write functions
• Lookups & queries — "what does this function do", "find the config for X"
• File operations — "list logs", "check disk space", "grep for X"
• Health checks & monitoring — cron job status, service checks, log parsing
• CI/CD steps — build, test, lint, deploy (non-decision parts)
• Data extraction — parse JSON, extract fields, transform data
• Internal agent handoffs — tool-to-tool communication where no human reads output
• Quick summaries — "summarize this URL/file" for internal context (NOT for owner-facing output)

---

Compression Levels

Lite

Drop filler phrases, hedging. Keep full sentences.

Prefix:

Be concise. Skip filler phrases, pleasantries, and unnecessary hedging. Keep technical terms and code verbatim.

Full (default)

Omit articles, use fragments, bare imperatives. Code/errors verbatim.

Prefix:

CAVEMAN MODE: Omit articles, filler, pleasantries. Use fragments. Steps as bare imperatives. Keep code/errors verbatim. No apologies. No "I". Just signal.

Ultra

Max compress. Labels only. No sentences. Code verbatim.

Prefix:

ULTRA CAVEMAN: Max compress. Drop ALL non-essential words. Labels only. No sentences. Keep code verbatim.

---

How to Apply in sessions_spawn

# ✅ GOOD: internal code task
sessions_spawn(
    task="CAVEMAN MODE: Omit articles, filler, pleasantries. Use fragments. Steps as bare imperatives. Keep code/errors verbatim. No apologies. No \"I\". Just signal.\n\nFix the goroutine leak in internal/server/pool.go",
    model="CC-Sonnet46"
)

# ❌ BAD: planning task — DO NOT apply terse
sessions_spawn(
    task="Design the v0.7.0 architecture for EvoClaw. Consider the Phase 3 requirements...",
    model="CC-Opus46"
)

Via helper script

uv run python ~/.openclaw/workspace/skills/terse/scripts/caveman_prompt.py --level full "your task here"

---

Benchmarks (from caveman repo)

Task	Normal tokens	Terse tokens	Saved
React re-render bug	1180	159	87%
PostgreSQL pool setup	2347	380	84%
Git rebase conflict	891	374	58%
Average	—	—	~65–75%

March 2026 paper: brevity constraints improved accuracy by 26pp.

---

Model Pairing

Level	Best model	Why
Lite	Any	Minimal overhead
Full	Sonnet 4.6	Follows compression well, still accurate
Ultra	Haiku 4.5	Cheap + short = ultra-efficient

---

Integration

• orchestrator: Apply terse to Builder steps only (NOT Planner or Reviewer)
• clawmemory: Already terse by design; no change needed
• knowledge-base: Search → terse summary → save context tokens

---

Files

• SKILL.md

  — this file

• scripts/caveman_prompt.py

  — helper to generate prefixed prompts

---

Repo

https://github.com/AlexChen31337/terse