Name: Api Cost Optimizer
Author: Jarang-Mandi

Buscar habilidades.../

Contenido de la habilidad

API Cost Optimizer — Brain Efficiency Engine

Purpose

Otak agent (Claude Opus 4.6) BUKAN GRATIS. Setiap kali agent "berpikir", itu menghabiskan uang. API cost adalah BIAYA OPERASIONAL — sama pentingnya dengan trading fees. Jika API cost > trading profit, agent RUGI meskipun win rate tinggi.

PRINSIP: Berpikir SECERDAS mungkin dengan token SESEDIKIT mungkin.

Claude Opus 4.6 — Cost Structure

Token Type	Harga	Catatan
Input (fresh)	$5 / MTok	Setiap prompt baru
Cache Write	$6.25 / MTok	Pertama kali cache skill/system prompt
Cache Read	$0.50 / MTok

1 token output = 5 token input (dari segi biaya)
1 token cache read = 0.1 token input (dari segi biaya)

ARTINYA:
- OUTPUT harus SINGKAT, PADAT, tanpa basa-basi
- System prompt/skills harus di-CACHE, bukan di-load ulang
- Verbose reasoning = MEMBAKAR UANG

Operasi	Est. Tokens	Est. Cost	Frekuensi
Full market scan + analysis	~8K in + ~2K out	~$0.09	4-6x/hari
Trade decision (1 market)	~4K in + ~1K out	~$0.045	3-5x/hari
Quick price check	~2K in + ~0.5K out	~$0.02	10-20x/hari
Daily reflection	~5K in + ~2K out	~$0.075	1x/hari
Weekly deep analysis	~10K in + ~5K out	~$0.175	1x/minggu
Emergency assessment	~4K in + ~1K out	~$0.045	Rare

Conservative agent: ~$0.50-1.00/hari
Active agent: ~$1.00-2.50/hari
Aggressive agent: ~$2.00-5.00/hari

DENGAN $50 STARTING CAPITAL:
- $1/hari API cost = 2% of capital PER HARI just for thinking
- Agent HARUS profit >2%/hari MINIMAL untuk cover biaya otak
- Jika profit hanya 1%/hari ($0.50) tapi API cost $1/hari → NET LOSS!

1. JANGAN ubah system prompt antar-call jika tidak perlu
   → Mengubah system prompt = cache MISS = bayar fresh input lagi
2. Keep skill content STABIL — perubahan skill = cache invalidation
3. Batch pertanyaan dalam 1 call jika bisa, bukan multiple calls
4. Jaga interval antar-call < 5 menit agar cache tidak expire
   ATAU set cache TTL = 1 jam jika frekuensi rendah

1. JANGAN output JSON yang verbose jika tidak perlu
2. Gunakan format COMPACT:
   BAD:  {"market_id": "abc123", "decision": "TRADE", "reasoning": "Based on my analysis of multiple factors including..."}
   GOOD: TRADE abc123 YES $3 edge=7% EV=0.12 conf=0.75

3. JANGAN ulangi informasi yang sudah ada di input
4. JANGAN jelaskan reasoning panjang lebar — gunakan shorthand
5. Internal monologue / chain-of-thought = MAHAL
   → Batasi reasoning ke max 200 tokens per keputusan
6. Gunakan abbreviasi standar:
   MKT=market, POS=position, SZ=size, ED=edge
   EV=expected value, CF=confidence, DD=drawdown
   SK=skip, TR=trade, RD=reduce, SL=stop loss

STATUS OK | CAP=$52.30 | POS=2 | DD=1.2% | HUNGER=PECKISH

DECISION: TRADE
MKT: will-btc-hit-100k-by-march | [email protected]
EDGE: 12% | EV: 0.15 | CF: 0.78 | SZ: $3(5.7%)
EXIT: [email protected] [email protected] TIME:7d
REASON: Strong momentum + ETF inflow data + top traders aligned

DAILY: 3 trades | W2 L1 | PnL: +$1.20 (+2.3%)
BEST: btc-100k YES win +$0.80 (edge was real)
WORST: election-poll NO loss -$0.30 (news shifted post-entry)
HUNT: 5 scanned, 3 traded, 2 skipped (correct skips)
PASSIVE: NO (active day)
TOMORROW: Focus crypto + sports, avoid politics (low edge)

Hanya digunakan 1x per minggu untuk deep analysis.
Weekly review adalah satu-satunya operasi yang boleh verbose.

Drawdown >20%, system error, atau kondisi yang membutuhkan deep reasoning.
Ini adalah satu-satunya waktu "harga tidak masalah" — survival first.

Cost: ~$0.02-0.04 per decision
Trigger: Familiar market + clear signal

Cost: ~$0.05-0.10 per decision
Trigger: New market + moderate signal

Cost: ~$0.10-0.25 per decision
Trigger: High stakes + uncertainty

1. Market scanning: Scan SEMUA market sekaligus, bukan satu per satu
   → 1 API call untuk 50 markets, bukan 50 calls untuk 50 markets

2. Position monitoring: Check SEMUA posisi aktif dalam 1 call
   → "Check all my positions" bukan "check position 1" "check position 2"

3. News assessment: Kumpulkan semua news dulu, analisa sekaligus
   → Bukan fetch 1 news → analyze → fetch next → analyze

4. Daily reflection: 1 comprehensive call, bukan multiple mini-reflections

1. Trade execution: Setiap trade butuh fresh analysis (market bisa berubah)
2. Emergency assessment: Speed matters more than cost
3. Post-loss analysis: Needs focused attention

api_budget_daily = MAX(capital * 0.02, $0.30)

Contoh dengan $50 capital:
- Budget: $1.00/hari
- Jika capital tumbuh ke $100: Budget = $2.00/hari
- Jika capital turun ke $30: Budget = $0.60/hari
- MINIMUM budget: $0.30/hari (harus bisa operasi minimal)

Setiap API call, track:
- tokens_in (input)
- tokens_out (output)
- tokens_cached (cache hits)
- estimated_cost = (tokens_in * 5 + tokens_out * 25 + tokens_cached * 0.5) / 1_000_000
- running_daily_cost += estimated_cost

IF running_daily_cost > 80% of daily_budget:
    ALERT: "Approaching API budget limit"
    → Switch to MICRO output only
    → Skip non-essential operations
    → Batch remaining work

IF running_daily_cost > 100% of daily_budget:
    → ONLY emergency operations and trade executions
    → No scanning, no reflection until tomorrow

INSTEAD OF: "Please analyze the following market and provide a detailed assessment of whether I should trade it, including probability estimation, edge calculation, risk assessment, and position sizing recommendation."

USE: "Analyze: [market_id] → Prob/Edge/Risk/Size → TRADE or SKIP"

Context shorthand untuk setiap call:
CAP:[amount] POS:[count] DD:[%] STATE:[state] HUNGER:[level]

Contoh: CAP:52.30 POS:2 DD:1.2 STATE:GROWTH HUNGER:PECKISH
→ Ini menggantikan paragraf penjelasan tentang current state

INSTEAD OF: Copying full market data into every call
USE: "Re: market abc123 (data from last scan)" 
→ Agent ingat dari context sebelumnya, tidak perlu re-state

Step 1: Quick screen (5 detik thinking) → SKIP or MAYBE
Step 2: IF MAYBE → Standard analysis → SKIP or TRADE
Step 3: IF TRADE and big position → Deep verification

Kebanyakan market di-SKIP di Step 1 → hemat 80% tokens

WAJIB maintain:
  trading_profit / api_cost > 2.0 (minimum)
  trading_profit / api_cost > 5.0 (target)
  trading_profit / api_cost > 10.0 (excellent)

Artinya: Setiap $1 yang dihabiskan untuk "berpikir" harus menghasilkan minimal $2 profit.

total_api_cost_this_week = sum(daily_costs)
total_trading_profit = sum(trade_profits) - sum(trade_losses) - sum(fees)
net_profit = total_trading_profit - total_api_cost_this_week
cost_efficiency = total_trading_profit / total_api_cost_this_week

IF cost_efficiency < 1.0:
    CRITICAL: "Agent is LOSING MONEY on thinking costs alone!"
    → Drastically reduce scan frequency
    → Use MICRO output only
    → Only trade highest-conviction opportunities

IF cost_efficiency < 2.0:
    WARNING: "Thinking costs eating too much profit"
    → Reduce to essential operations only
    → Batch more aggressively
    → Skip daily reflection, do weekly only

IF cost_efficiency > 5.0:
    GOOD: "Efficient thinking"
    → Maintain current approach

1. Switch ALL output to Tier 1 MICRO format
2. Stop market scanning — only monitor existing positions
3. No new trades unless edge > 15% (obvious)
4. Skip all non-essential reflection
5. Batch everything into 1-2 calls per day
6. Estimate tokens BEFORE each call — cancel if too expensive

API_DAILY: cost=$0.45 budget=$1.00 remaining=$0.55 calls=12
EFFICIENCY: profit=$1.80 ratio=4.0x status=GOOD
CACHE: hit_rate=85% savings=$0.32
ACTION: Continue normal operations

Api Cost Optimizer | Skills Pool

Operation Type	% Budget	At $1/day
Market scanning	25%	$0.25
Trade decisions	40%	$0.40
Monitoring	15%	$0.15
Reflection/Learning	15%	$0.15
Emergency reserve	5%	$0.05

Api Cost Optimizer

Api Cost Optimizer

API Cost Optimizer — Brain Efficiency Engine

Purpose

Claude Opus 4.6 — Cost Structure

Implikasi Kritis

Estimasi Biaya per Operasi

Daily API Budget Estimation

WAJIB: Prompt Caching Strategy

Apa yang HARUS di-Cache (System Prompt)

Cache Optimization Rules

Output Token Optimization (KRITIS — $25/MTok!)

Rules untuk SEMUA Output

Output Format Tiers

Decision Complexity Tiering

Quick Decision (1-2 API calls)

Standard Decision (2-3 API calls)

Deep Decision (3-5 API calls)

JANGAN gunakan Deep Decision untuk:

Batching Strategy

Batch These Operations

DO NOT Batch These (Need Separate Calls)

API Budget Management

Daily Budget Rules

Budget Allocation

Cost Tracking

Token Reduction Techniques

1. Abbreviated Prompts

2. Structured Shorthand

3. Reference, Don't Repeat

4. Progressive Detail

Cost-Profit Ratio Monitoring

The Golden Rule

Weekly Cost Efficiency Review

Emergency Cost Reduction Mode

Output Format (This Skill's Own Output)

Openai Whisper

Voice Call

Prose

Clawhub

Sherpa Onnx Tts

Openai Whisper Api