How to Read Your Token Cascade
Your cascade is a diagnostic — it tells you exactly what’s working and what’s broken in your AI coding workflow. Here’s how to read it.
The cascade as a diagnostic
Your token cascade isn’t just a score — it’s a diagnostic tool. The four pillars are like blood panel results: each tells you something specific about your workflow’s health. Read together, they tell you exactly what to fix. Run sigrank me or paste your ccusage JSON into the /score calculator to get your pillars, then use this guide to interpret them.
What each pillar reveals about your workflow
Input — your context injection rate
Input reveals how much fresh context you inject per turn. High input means you’re sending new context the model hasn’t seen — either because your context window is unstable (cache breaks) or because you’re re-pasting files unnecessarily. Low input means you’re sending small deltas on top of a stable cached base.
Output — your signal production
Output reveals how much signal the model produces per turn. High output means your prompts generate useful code efficiently. Low output with high input means the model is churning or your prompts are unfocused.
Cache-read — your context reuse
Cache-read reveals how well you reuse context via prompt caching — the most diagnostic pillar. High cache-read means your context window is stable and the model reads from cache. Low cache-read means your cache breaks every turn.
Cache-write — your cache investment
Cache-write reveals how much you’re investing in future cache hits. High cache-write early that converts to cache-read later is a healthy, compounding cascade. High cache-write that never converts means your cache keeps breaking before it pays off.
What high cache-read means (good context reuse)
High cache-read is the signature of an efficient operator. It means:
- Your context window is stable — you’re not reordering context or changing the system prompt between turns.
- You’re sending small deltas on top of a large cached base, not re-pasting full context.
- Your cache hit rate is high (above 80%) — you write to cache once and read from it many times.
- Your yield is high — cache-read is in the yield numerator, so high cache-read directly boosts Υ.
If your cache-read is high, keep doing what you’re doing. The next lever is output — can you get even more signal per turn with better prompt structure?
What high input / low output means (inefficient prompting)
High input with low output is the most common dysfunction — you’re sending a lot of fresh context but the model isn’t producing much signal. The usual culprits:
- Verbose prompts — conversational padding, restated context, redundant instructions. The model processes 500 tokens to find the 20-token instruction buried inside.
- Unstable context — your cache breaks every turn, so the model re-processes everything as fresh input but produces the same output it would have from cache.
- Noisy context — your context window includes irrelevant files or instructions that dilute the signal. The model processes noise instead of signal.
- Re-rolls — you regenerate from scratch instead of correcting, sending fresh input without advancing cache-read.
The fix: structure inputs to be minimal, stabilize your context window, and replace re-rolls with corrections. Each fix reduces input and increases output.
What balanced cascades look like
A balanced cascade isn’t about equal pillar counts — it’s about the right ratios. The ideal cascade shape:
- High cache-read — the largest pillar. Most of your context is reused, not re-sent.
- Low input — the smallest pillar. You send only tiny deltas on top of cache.
- High output — the model generates substantial signal per turn.
- Moderate cache-write — early investment that converts to cache-read over time.
In numbers: input 5,000, output 12,000, cache-read 40,000, cache-write 8,000. Cache-read dominates, input is minimal. Yield = (40,000 × 12,000) / 5,000² = 19,200 — a BUILDER or TRANSMITTER-class cascade.
Class tier interpretation
Your class tier summarizes your cascade shape into a single label — a quick diagnostic and a target:
IGNITER
Low yield. Mostly fresh input, little cache reuse. The cascade burns tokens, not compounds them. Focus: stabilize your context window.
SEEKER
Improving but inconsistent. Some cache reuse, but input still dominates. The cascade starts to compound but breaks frequently. Focus: maintain context stability.
BUILDER
Solid cache reuse and decent output. The cascade is compounding reliably. Input is controlled. Focus: increase output per turn with better prompt structure and batched tasks.
TRANSMITTER
High cache-read, low input, high output — the ideal cascade shape. Signal compounds efficiently. Focus: maintain discipline and push yield higher.
FAQ
- What does the cascade tell me about my workflow?
- The four pillars reveal how you manage context, how efficiently you prompt, and whether signal is compounding or being burned. It’s a diagnostic — each pillar points to a specific aspect of your workflow.
- What does high cache-read mean?
- Good context reuse. Your context window is stable and the model reads from cache instead of re-processing fresh input. This is the signature of an efficient operator and directly boosts yield.
- What does high input with low output mean?
- Inefficient prompting. You’re sending a lot but getting little back. Usually caused by verbose prompts, unstable context, noisy context, or re-rolls. Fix: structure inputs, stabilize context, replace re-rolls with corrections.
- What does a balanced cascade look like?
- High cache-read (largest pillar), low input (smallest), high output, and moderate cache-write that converts to cache-read over time. Cache-read dominates, input is minimal, output is high.
- How do I interpret my class tier?
- IGNITER (low yield, burning tokens) → SEEKER (improving, inconsistent cache) → BUILDER (solid reuse, decent output) → TRANSMITTER (ideal cascade shape, high yield). Use it as a diagnostic and a target.