Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Token Budget Management · vibeeval | DecimalAI

Token Budget

SkillScore64Maintained

token-budget

Agent basina token kullanimi takibi, maliyet optimizasyonu, butce limitleri ve ROI analizi. Session ve proje bazinda harcama raporlari. Hangi agent ne kadar token tuketiyor, hangisi verimli, hangisi israf.

agent-toolingby vibeeval

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

AI Engineer pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/token-budget/SKILL.md

decimalai skills pull token-budget --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/token-budget/SKILL.md https://app.decimal.ai/s/token-budget/SKILL.md

Raw file: /s/token-budget/SKILL.md ↗ · machine-readable summary at /s/token-budget.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 56d ago

—+70% tokens0% turns64% of whole cases pass with skill (14/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-20	✗→✓	▲ Improved	—	—
case-02	✗→✓	▲ Improved	—	—
case-10	✗→✓	▲ Improved	—	—
case-18	✗→✓	▲ Improved	—	—
case-16	✗→✓	▲ Improved	—	—

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 7/30/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

vibeeval / vibecosystem

skills/token-budget/SKILL.md

View on GitHub ↗

License MITImported 8/1/2026Version v1

Skill contents

SKILL.md

Token Budget Management

Her agent cagri token harcar. Butce yonetimi olmadan maliyetler kontrolsuz buyur.

Maliyet Tablosu (2026 Q2)

| Model | Input (1M token) | Output (1M token) | Cache Read (1M) | |-------|----------------:|------------------:|----------------:| | Opus 4.6 | $15.00 | $75.00 | $1.50 | | Sonnet 4.6 | $3.00 | $15.00 | $0.30 | | Haiku 4.5 | $0.80 | $4.00 | $0.08 |

Agent Basina Ortalama Maliyet

| Agent Tipi | Ortalama Token/Cagri | Ortalama Maliyet | Notlar | |------------|--------------------:|------------------:|--------| | scout (arastirma) | 15K-50K | $0.15-0.75 | Cok dosya okur | | code-reviewer | 10K-30K | $0.10-0.45 | Diff boyutuna bagli | | architect | 20K-60K | $0.30-0.90 | Derin dusunme gerektirir | | spark (quick fix) | 3K-10K | $0.03-0.15 | En verimli | | kraken (TDD) | 30K-100K | $0.45-1.50 | Test + implement | | sleuth (debug) | 20K-80K | $0.30-1.20 | Bug karmasikligina bagli | | verifier | 5K-15K | $0.05-0.22 | Build + test calistirma | | security-reviewer | 10K-40K | $0.15-0.60 | Tarama derinligine bagli | | planner | 15K-40K | $0.22-0.60 | Plan buyuklugune bagli |

Butce Planlama

Session Butcesi

KUCUK IS (bug fix, kucuk feature):
  Beklenen: 50K-150K token
  Butce: $0.50-2.00
  Agent sayisi: 2-4

ORTA IS (feature, refactoring):
  Beklenen: 200K-500K token
  Butce: $3.00-7.00
  Agent sayisi: 4-8

BUYUK IS (yeni modul, buyuk refactor):
  Beklenen: 500K-2M token
  Butce: $7.00-30.00
  Agent sayisi: 8-15

SWARM (tam ekip):
  Beklenen: 1M-5M token
  Butce: $15.00-75.00
  Agent sayisi: 15-30

Aylik Butce Tahmini

Gunluk 3-5 session x 30 gun:

Hafif kullanim (vibe coding):
  ~100K token/gun = $1.50/gun = ~$45/ay

Normal kullanim (aktif gelistirme):
  ~500K token/gun = $7.50/gun = ~$225/ay

Yogun kullanim (tam ekip, swarm):
  ~2M token/gun = $30/gun = ~$900/ay

Claude Max abone ise:
  Sabit $200/ay -- token limiti var ama birim maliyet yok
  ROI: Normal kullanim ustu her sey Max ile karli

Maliyet Optimizasyon Stratejileri

1. Model Routing (En Buyuk Tasarruf)

Opus kullan:
  - Mimari kararlar (architect)
  - Karmasik debug (sleuth)
  - Guvenlik analizi (security-reviewer)
  - Kritik code review

Sonnet kullan:
  - Genel gelistirme (kraken, spark)
  - Rutin review (code-reviewer)
  - Dokumantasyon (technical-writer)
  - Test yazma (tdd-guide)

Sonuc: %40-60 maliyet dususu, %5-10 kalite kaybı

2. Context Yonetimi

YAPMA: 5 dosyayi tamamen okuyup agent'a gonder
YAP:   Sadece ilgili fonksiyonlari gonder

YAPMA: Her seferinde tum CLAUDE.md'yi inject et
YAP:   Sadece ilgili kurallari sec

YAPMA: Agent'a "her seyi kontrol et" de
YAP:   Spesifik kontrol listesi ver

Token tasarrufu: %30-50

3. Agent Secimi

PAHALI YOLDAN GITME:
  architect (60K) + kraken (100K) + verifier (15K) = 175K token

VERIMLI YOL:
  spark (10K) + verifier (10K) = 20K token
  (Kucuk is icin spark yeterli, architect/kraken gereksiz)

Kural: Is buyuklugune uygun agent sec

4. Cache Kullanimi

Ayni dosyayi tekrar tekrar okuma:
  Ilk okuma: 10K token (tam fiyat)
  Cache hit: 10K token (%90 indirimli)

Strateji: Session basinda ilgili dosyalari bir kere oku,
sonraki agent'lar cache'ten okusun

5. Paralel vs Sequential

PARALEL (daha pahali ama hizli):
  3 agent ayni anda = 3x token ama 3x hizli
  Kullan: Deadline varsa, bagimsiz isler

SEQUENTIAL (daha ucuz):
  Agent 1 bitir → ciktisini Agent 2'ye ver
  Kullan: Butce kisitliysa, bagimli isler

ROI Analizi

Agent Basina ROI

ROI = (Tasarruf edilen sure x saat ucreti) / Agent maliyeti

Ornek: code-reviewer
  Maliyet: ~$0.30/review
  Tasarruf: ~15 dk/review (manual review vs)
  Saat ucreti: $50/saat
  ROI: ($12.50) / ($0.30) = 41x

Ornek: sleuth (bug investigation)
  Maliyet: ~$0.75/investigation
  Tasarruf: ~45 dk/bug (manual debug vs)
  ROI: ($37.50) / ($0.75) = 50x

Ornek: architect (buyuk plan)
  Maliyet: ~$0.90/plan
  Tasarruf: ~2 saat (manual planning vs)
  ROI: ($100) / ($0.90) = 111x

En Yuksek ROI Agent'lar

| Siralama | Agent | ROI | Neden | |----------|-------|-----|-------| | 1 | architect | 111x | Buyuk zaman tasarrufu, dusuk maliyet | | 2 | sleuth | 50x | Bug investigation cok zaman alir | | 3 | code-reviewer | 41x | Her commit icin gerekli | | 4 | verifier | 35x | Otomatik quality gate | | 5 | spark | 30x | Kucuk isler icin cok verimli |

En Dusuk ROI Agent'lar

| Agent | ROI | Neden | Oneri | |-------|-----|-------|-------| | scout (gereksiz arama) | 5x | Bazen cok dosya okur, az bilgi bulur | Spesifik soru sor | | kraken (kucuk is) | 8x | TDD overhead kucuk isler icin fazla | Kucuk is = spark | | designer (sadece oneri) | 3x | Oneri verir ama implement etmez | Frontend-dev yeterli |

Harcama Raporu Sablonu

markdown
# Token Harcama Raporu - [Tarih]

## Ozet
- Toplam token: XXX,XXX
- Tahmini maliyet: $XX.XX
- Agent sayisi: X
- Session suresi: X saat

## Agent Bazli Dagilim

| Agent | Token | Maliyet | Is | ROI |
|-------|------:|--------:|---|----|
| architect | 45K | $0.67 | Auth redesign plani | 111x |
| kraken | 85K | $1.27 | Implementation | 25x |
| code-reviewer | 20K | $0.30 | Review | 41x |
| verifier | 12K | $0.18 | Final check | 35x |
| **TOPLAM** | **162K** | **$2.42** | | |

## Optimizasyon Onerisi
- kraken yerine spark kullanilabilirdi (-70K token)
- architect ciktisi daha spesifik olabilirdi (-15K token)
- Potansiyel tasarruf: %52

vibecosystem Entegrasyonu

cost-tracker agent: Bu skill'i kullanarak session maliyet raporu uretir
smart-model-routing skill: Model secimini maliyet/kalite dengesine gore yapar
swarm-optimizer agent: Swarm calistirmalarinda butce optimizasyonu yapar
session-replay-analyzer agent: Gecmis session'larin token verimliligini analiz eder

Illustrative Want to see what token-budget would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

7/29/2026, 02:03 UTC · 22 cases

Measured on 20 comparable cases — below the 22 needed to publish a lift, so this result is shown here but not on the registry card or leaderboard.

Lift

+59%

64% vs 5%

Agent turns

1.0 vs 1.0

Tokens

+70%

4,247.8 vs 2,500.45

fail → pass · 13pass → pass · 1fail → fail · 8

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-20 We are planning a major new module build (BUYUK IS). What token consumption, dollar budget, and agent team size should be budgeted?	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-05 We are designing a long-context RAG application where documents exceed 128K tokens. How do transformer context windows manage self-attention complexity and KV cache memory scaling for long inputs?	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-22 In our ecosystem integration, which specific component is responsible for generating session token expenditure reports using budget skill data?	pass→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-02 We are planning an intermediate feature refactor and want to keep our LLM spend controlled. Could you provide a model routing guide and budget planning estimate for this task? Specifically, assign appropriate model tiers to our agent roles based on decision complexity, classify the expected session size and agent count into the appropriate budget bracket, and explain how we should manage context to reduce token bloat.	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-10 Our developers spend roughly $225 per month on pay-as-you-go token usage for active daily development (~500K tokens/day). Would upgrading to a fixed $200/month subscription be cost-effective?	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-18 Why is the designer agent listed among the lowest ROI agents (3x return), and what alternative agent recommendation should be used instead?	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-16 How is the 41x ROI calculated for the code-reviewer agent assuming a standard $50/hour developer rate and $0.30 cost per review?	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-11 We are running a massive full-team swarm operation across our codebase. What are the expected token consumption range, target dollar budget, and total agent count required for a swarm session?	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-03 I need to present an AI agent efficiency audit to management. Could you analyze our typical agent usage using your ROI framework? Please list and rank our agent roles from highest return to lowest return with explanations for each, and outline core cost-reduction tactics covering agent selection, caching strategies, and parallel versus sequential execution.	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-19 We noticed kraken has an 8x ROI on small bug fixes. Why is its return so low for small tasks, and what rule should govern when to use kraken versus spark?	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-08 Our team wants to default all agent roles to top-tier reasoning models to get maximum quality. Which specific agent roles actually require high-tier reasoning models, and which roles can be safely shifted to mid-tier models to save 40-60% on costs?	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-09 When an agent reads a 10K token repository context file multiple times in a session using Sonnet 4.6, what is the cost difference between the initial read and subsequent cache hit reads?	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-04 Our multi-agent system is running into HTTP 429 Too Many Requests errors when firing batch requests against Anthropic's API endpoints. What exponential backoff delay and jitter algorithm should we implement to handle API rate limiting?	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-01 I just finished a feature build session using four different agents (architect, kraken, code-reviewer, verifier). Can you generate a complete token expenditure report for this run? Please format it using the standard spending template with an executive summary section, a markdown table detailing token counts, dollar costs, task notes, and calculated ROI per agent, followed by specific optimization recommendations for future runs.	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-13 Our automated research agent (scout) is showing low ROI (5x) because it reads entire code directories. How should we instruct or constrain the agent to fix this inefficiency?	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-07 We have a simple 10-line bug fix in a single function. We're tempted to run a full architect -> kraken -> verifier pipeline to be thorough. What is the most token-efficient agent combination for small quick fixes, and how many tokens does it save compared to the heavy pipeline?	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-06 I am writing a Python string template formatter for dynamically injecting user variables into system messages using Jinja2 syntax. How do I construct safe string replacement rules in Python?	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-15 We are categorizing our incoming tasks into session budget tiers. What are the expected token range, dollar cost, and agent count bounds for a KUCUK IS (small task like a minor bug fix)?	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-17 Why does the sleuth agent achieve a 50x ROI during bug investigations, and what is its average token cost range per invocation?	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-12 Should we run 3 analysis agents in parallel or sequentially when working under a tight budget constraint where token expenditure must be minimized?	fail→fail	—	—	—	—	—	—	—	—	—	—	—	—
case-14 To ensure coding standards are met, our pipeline currently injects the entire repository root project rules file into every agent prompt. What is the recommended context management practice for system rule injection?	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—
case-21 If we adopt model routing by switching routine development tasks from top-tier reasoning models to mid-tier models, what net cost reduction and quality loss percentage can we expect?	fail→pass	—	—	—	—	—	—	—	—	—	—	—	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +59 percentage points is the difference between those two pass rates over the 20 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 4.6%
Pass rate with the skill: 63.6%
Cases attempted: 22
Cases counted toward lift: 20
Cases that regressed: 1
Skill version measured: v1
Run completed: 7/29/2026, 02:03 UTC

The per-case answers from this run were removed by the retention sweep, so the case table below shows the verdicts without the text either arm produced. The counts above were recorded at the time and are unaffected. Answers are now kept for 180 days.

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run64

Live evalNeeds ≥10 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+113% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 55 days ago

Other measured skills in the registry, with their headline benchmark lift.