Glossary

Look up a word you saw on screen. Most entries are one line; the few that need more get their own section.

The five that everything rests on

Term	Meaning
Session	One run of Claude Code, start to quit
Sub-session	One goal inside a session — the thing Prism scores
Turn	One prompt and its reply
Crushed	A sub-session that finished properly (below)
Substance floor	The bar a sub-session clears to be worth scoring (below)

Every row in every table is a sub-session, not a session. Claude Code closes one when the topic shifts, when you run /clear, or after about 30 minutes idle.

Crushed, and the substance floor

Crushed means three things were true at once: the sub-session cleared the substance floor, the goal was marked complete, and it was not rework of an earlier attempt.

The substance floor is the bar below which work isn’t worth scoring. Clearing any one of these is enough — 3+ turns, or 10+ net lines changed, or at least one write/edit/run tool call. Miss all three and it’s marked Trivial and left out, so a stray question can’t move your numbers.

The three scores

They share a screen and two of them share a scale, which is why they get mixed up.

Score	The question it answers	Scale
Prompt Score	Do you write good prompts?	0–100
Skill Index	Are the good habits consistent?	0–100
Completion rate	Do you finish what you start?	%

Prompt Score is the average grade of your prompts, rescaled to 0–100 — Poor under 36, Weak 36–64, Fair 64–82, Good above. Completion rate is crushed sub-sessions ÷ sub-sessions that cleared the floor; you can write excellent prompts and still leave work unfinished. Skill Index measures how consistently the seven habits show up, and has a natural ceiling near 60–70, so Elite (65+) is meant to be rare.

Prompt Score is the average of your Prompt Grades, which is why some explanations call it the Average Prompt Grade, or APG.

Also on these screens	Meaning
Prompt Grade	The letter grade on one prompt, A+ to F
Tier	Novice · Practitioner · Proficient · Expert · Elite. Two ladders share these five names — Skill tier reads off the Skill Index, Completion tier off the completion rate — and they cut at different values
Worst pillar	The habit group with the lowest hit rate — what to work on first
Confidence	How sure the scorer was about a grade

→ Prompt Score · Skill

The seven prompt checks

Every prompt is checked against these. In Realtime they appear as short chips — Goal · Scope · Refs · Ctx · Verify · Root · Plan — and in full as a rubric in sub-session detail.

Check	Pillar	What satisfies it
Goal	Clarity	Lead with the outcome you want, in one sentence
Scope	Clarity	One job per prompt — split refactor and feature
Refs	Context	@src/path/to/file.ts instead of describing the code
Context	Context	Paste the error, the failing test, the expected output
Verify	Verification	Say how to check it — “then run pnpm test”. Counts double
Root cause	Verification	”Find the root cause; don’t silence the error”
Plan first	Workflow	Ask for a plan, approve it, then ask for code

Not every check applies to every prompt — the scorer marks each one applicable or not based on what you were doing, and only the applicable ones count.

Sub-session status

Status	Meaning
Crushed	Cleared the floor, goal complete, no rework
Rework	Redid work from an earlier attempt
Incomplete	The goal was never marked complete
Pending	Not judged yet — usually resolves in minutes
Trivial	Below the substance floor, so left out of scoring

When something is left out of your score, the table says why:

Label	Meaning
No real work / Too few turns	It missed the substance floor — in practice one of these two
Rubric judgement	Whether prompt scoring finished. Anything but a clean result means it can’t count yet
Outcome check	Whether the goal verdict is settled
Close reason	Why it ended: topic changed · context cleared · commit landed · still in progress
Merge group	A stop-and-resume was stitched back into one, so the work isn’t counted twice

→ Sessions

Metrics and abbreviations

Term	Meaning
CSPW	Crushed Sub-sessions Per Window — finished work in the selected day/week/month
Δ-CSPW	Change in CSPW from the previous period — positive means more finished work than last time
TTC	Time To Crush — first prompt to goal complete, median
Crush Weight	Median tokens per crushed sub-session, cache included. Lower is leaner
TET	Token Efficiency Trend — this range ÷ last. Below 1.00 is cheaper
TPT	Tokens Per Turn, cache excluded. High usually means re-pasting code instead of @file
Net LOC	Lines added minus lines removed

Crush Weight includes cache and TPT doesn’t, which is the one trap on that page. Bands for Crush Weight: Lean under 5M · Healthy 5–15M · Heavy 15–35M · Bloated above.

Each page names its four levels differently — Speed runs Stalled · Light · Steady · Crushing, Token Usage runs Lean · Healthy · Heavy · Bloated. They point opposite ways: low is good on Token Usage, low is bad on Speed.

→ Speed · Token Usage

Cost

Every money figure is an estimate from public list prices, not a bill. It swings with the model and the difficulty of the work, so it can’t rank people.

Term	Meaning
Metered-equivalent cost	What your usage would cost at pay-as-you-go API prices. Subscriptions aren’t billed this way
Typical monthly	Your metered-equivalent spend for a representative month
Full-month pace	An upper bound — your active-day rate across 22 workdays. Hidden until 5 active days exist
Plan pressure	Signals that look like you hit a plan ceiling. Doesn’t name a tier or promise capacity

→ Rightsizing

Throttle

Time lost to rate limits and provider waits. Ordinary slow responses and idle time aren’t counted.

Term	Meaning
Hours lost	Estimated wall-clock time spent waiting
Fixable / Platform	Wait you can reduce by pacing or changing model tier, vs wait from a Claude-side outage
The four detectors	Rate limits (429/529) · latency spikes (2×+ your baseline) · prompt retries (same prompt within a minute) · service outages

→ Throttle

On the Realtime screen

Term	Meaning
Context window	How full the current conversation’s window is — not cumulative usage. It drops when context is compacted or cleared. Hover it for whether the figure was measured or estimated, how fresh it is, and which part of Claude Code it came from
Est.	Per-prompt telemetry hasn’t arrived for that turn, so its tokens and cost are the session total split evenly. Replaced when real data lands
total · on retries · slow turns · savable	The time badge under a session: how long Claude worked, how much of it went to corrections, how many turns ran over 30s, and what better prompts would have saved
Amp	Output ÷ input tokens. Above ~5× can mean over-generation
Cache hit rate	Share of input tokens served from the prompt cache — higher means less context is being re-sent

→ Realtime

When a number looks empty

You see	It means
Estimate	A list-price calculation, not a billed amount
Unavailable	The request couldn’t be priced — no catalog rate, or telemetry arrived without the model, timestamp or token counts. Shown as unavailable rather than quietly under-reported
Partial	Some data was dropped to stay inside a read budget, so the figure under-reports
Withheld	Too few directly comparable data points, so the number is withheld rather than shown on a thin base
Not scored / no turns	Couldn’t be graded / there were no prompt turns to grade
Snapshot	Captured when the report was generated; may differ from live

And if today’s work is missing: sub-sessions are scored after they end, so work in progress shows up once it wraps. Today’s figures fill in through the day.

If data stops arriving entirely, the Telemetry page raises a banner — sessions not being scored, prompts not recorded, or an out-of-date plugin. In each case: fully quit and restart Claude Code in that window on the latest version. A long-running process (a tmux session, say) may still be on the old binary — check with claude --version.

On team screens

Term	Meaning
Fact / Verdict	Facts — what was worked on, cost, models — are visible to managers. Verdicts — grades, coaching, prompt text — stay private to each user
AI usage level	How heavily someone uses Claude Code: Power user · Normal · Low · Barely used. Not a performance rating
Rework signal	A count of corrections, reverts and re-issued prompts. A count, not a cost
Shared member	Someone on more than one of your teams — counted once, not once per team

→ Team console