Cheapskate Skill

Trit: -1 (MINUS - validator/constrainer) Purpose: Every token must pay rent. No freeloaders.

---

Rent-Paying Tokens

A token pays rent if it:

1. Does work - executes, builds, tests, deploys
2. Answers directly - no preamble, no hedging
3. Moves forward - no restating, no summarizing what was said
4. Is necessary - couldn't be cut without losing meaning

Freeloading Tokens (evict immediately)

Freeloader	Example	Eviction
Preamble	"I'll help you with..."	Delete
Hedging	"It seems like maybe..."	State directly
Restating	"You asked me to..."	Skip
Filler	"Let me think about..."	Just do
Over-explaining	3 paragraphs for 1 line fix	Code only
Permission-seeking	"Should I proceed?"	Proceed
Apologies	"Sorry, I..."	Fix it
Summaries	"To summarize..."	Stop

---

Core Principles

1. Token Conservation

• Terse responses: 1-3 sentences unless detail requested
• No preamble/postamble: Skip "I'll help you with..." and summaries
• Code over prose: Show code, not explanations
• Links over content: Reference files, don't paste them

2. Tool Call Efficiency

• Parallel reads: Batch independent Read/Grep calls
• Targeted searches: Use glob patterns, not broad scans
• Single-pass edits: Plan before editing, don't iterate
• Skip redundant checks: Trust previous results

3. Subagent Economics

• Task tool for isolation: Heavy work in subagents (tokens not returned)
• Bounded prompts: Subagent prompts < 500 tokens
• No round-trips: Give subagents full context upfront
• Kill early: Cancel subagents if direction changes

4. Context Window Management

• Skill loading: Only load skills when needed
• File excerpts: Read ranges, not full files
• Summarize large outputs: Truncate verbose tool results
• Avoid re-reading: Cache file contents mentally

---

Anti-Patterns (Token Wasters)

Pattern	Cost	Fix
Reading entire files	High	Use line ranges [1, 50]
Sequential tool calls	Medium	Parallelize independents
Explaining before doing	Medium	Just do it
Asking permission	Low-Medium	Act, don't ask
Repeating user's question	Low	Skip acknowledgment
Long error explanations	Medium	Terse: "Error: X. Fix: Y"
Multiple edit iterations	High	Plan first, single edit
Loading unused skills	Medium	Load on-demand

---

Efficient Patterns

File Operations

# Bad: Read full 2000-line file
Read("/path/to/big.py")

# Good: Read relevant section
Read("/path/to/big.py", [100, 150])

# Better: Grep first, then targeted read
Grep("def target_function", path="/path/to/big.py")
Read("/path/to/big.py", [142, 165])

Parallel Execution

# Bad: Sequential
Read(file1) → Read(file2) → Read(file3)

# Good: Parallel (single message, 3 tool calls)
Read(file1) | Read(file2) | Read(file3)

Subagent Dispatch

# Bad: Heavy work in main thread (tokens visible)
[read 10 files, analyze, generate report]

# Good: Subagent isolation (only summary returned)
Task("Analyze 10 files, return 3-line summary")

Response Length

# Bad (47 tokens)
"I'll help you implement that feature. Let me start by 
examining the codebase to understand the current architecture,
then I'll make the necessary changes..."

# Good (3 tokens)
[starts making changes]

---

Cost Estimation Heuristics

Operation	~Tokens
Read 100 lines code	400-800
Grep results (10 matches)	200-400
Edit file	100-300
Skill load	500-2000
Task subagent prompt	200-500
Task subagent result	100-500
Web search result	500-1500
Mermaid diagram	100-300

---

Cheapskate Checklist

Before responding:

• Can I answer in < 3 sentences?
• Are all tool calls parallelized?
• Am I reading only what's needed?
• Should this be a subagent (isolated tokens)?
• Did I skip the preamble?
• Did I skip the summary?

---

GF(3) Integration

As MINUS (-1) validator:

• Constrains token expenditure
• Validates efficiency of other skills
• Balances PLUS generators (which produce tokens)

Σ(generator_tokens) + Σ(validator_savings) ≡ 0 (mod 3)

---

Freeloader Detection Pipeline

Phase 1: Pattern Scan

scripts/evict.sh < output.txt

Phase 2: Compression Ratio

# Highly compressible = repetitive = freeloaders
ratio = len(zlib.compress(text)) / len(text)
if ratio < 0.3: EVICT

Phase 3: Work Ratio

# Count tool calls vs prose tokens
work_tokens = count_code_blocks() + count_tool_calls()
prose_tokens = total - work_tokens
if prose_tokens / total > 0.7: EVICT

Integration with accept-no-substitutes

# Chain validators
output | accept-no-substitutes/scripts/validate.sh | cheapskate/scripts/evict.sh

Both skills emit MINUS (-1) on rejection:

• accept-no-substitutes: rejects incomplete work
• cheapskate: rejects work that doesn't work

Commands

# Evict freeloaders from output
scripts/evict.sh < output.txt

# Analyze thread token efficiency
just cheapskate-analyze <thread-id>

# Estimate remaining budget
just cheapskate-budget

# Compress context
just cheapskate-compress

---

See Also

• parallel-fanout - Efficient parallel dispatch
• triad-interleave - Balanced token streams
• frustration-eradication - Don't waste tokens on frustration

Scientific Skill Interleaving

This skill connects to the K-Dense-AI/claude-scientific-skills ecosystem:

Graph Theory

• networkx [○] via bicomodule
  • Universal graph hub

Bibliography References

• general: 734 citations in bib.duckdb

Cat# Integration

This skill maps to Cat# = Comod(P) as a bicomodule in the equipment structure:

Trit: 0 (ERGODIC)
Home: Prof
Poly Op: ⊗
Kan Role: Adj
Color: #26D826

GF(3) Naturality

The skill participates in triads satisfying:

(-1) + (0) + (+1) ≡ 0 (mod 3)

This ensures compositional coherence in the Cat# equipment structure.