Core concepts
Connect prompts, subtasks, scores, and usage to understand your Prism results.
Prism brings together what you asked Claude Code or Codex to do, the work that followed, and the activity it used. Understanding how these fit together helps you choose a screen and interpret its results.
Imagine asking Claude Code to fix a login error. You add reproduction steps, inspect the proposed change, and ask it to run a test. Later, in the same conversation, you ask how to build the project. This example contains several prompts and more than one goal.
From a prompt to a piece of work
Section titled “From a prompt to a piece of work”| Term | What it means in the example |
|---|---|
| Prompt | An input you send to an agent, such as the request to fix the login error or the added reproduction steps. |
| Turn | A unit counted in the Turns column of Sessions. Fixing the error can take several turns. |
| Request | A request sent to the model. Claude Code can make several requests while responding to one prompt. |
| Session | The Claude Code or Codex conversation that contains the bug fix and the later build question. |
| Subtask | The smallest unit of work that gets a status and a score. The bug fix and the build question can appear as separate subtasks. |
| Task | The item that Sessions lists. Opening a Task shows its subtasks and the prompts of the Task. |
The example illustrates these relationships; it does not prescribe an exact number of subtasks. Open Sessions to inspect how your activity was grouped and which prompts belong to a Task. Realtime counts the sessions and subtasks of the day, so one prompt, one session, and one subtask are not interchangeable counts.
A subtask ending and its goal being completed are separate things. A subtask can end as Completed, Stopped, Abandoned, or Failed, or its result can still be waiting for judgment. See Scores and work status for the statuses.
Records, estimates, and evaluations
Section titled “Records, estimates, and evaluations”The same bug fix can produce three kinds of information:
| Kind | What you can learn |
|---|---|
| Recorded activity | The sessions, requests, tokens, and active time recorded for the work. |
| Estimates | An estimated cost converted at API prices. It is never a bill. |
| Evaluations and suggestions | A Prompt Score for the prompting in a subtask, and Coaching that ties the missed items to what happened. |
An activity record can arrive before its evaluation. A subtask that is waiting for judgment has no score yet, and an unavailable cost is not zero. Check Periods and data states when a result is calculating, partial, or unavailable.
The Prism CLI and the two plugins connect your Claude Code and Codex activity to Prism; the web app provides the scores and the review screens. Read Data collection and visibility for what is sent and who can see the results.
Read the measures together
Section titled “Read the measures together”Each measure answers a different question about the recorded work.
| Measure | Question it helps answer |
|---|---|
| Prism Score and Prompt score | How was the counted work or the prompting graded over the selected period? |
| Prompt | What is the Prompt Score of each subtask, and how often was each of the seven items met? |
| Speed | How long did subtasks take to finish, how long did they actually run, and how often did you step in? |
| Token Usage | What does one completion cost, and where did the tokens go? |
| Skill | How consistently were the seven prompt items met? |
For the login example, a subtask’s status concerns how the work ended; the Prompt Score concerns the prompting used in that work. A completed fix need not receive the highest score. A subtask below the workload floor gets no score even if its answer was useful.
Use these measures to choose work to review. They do not by themselves establish overall ability, business value, or the cause of a change. Estimated cost is converted at API prices, so fewer tokens alone do not establish a lower bill. Scores and work status explains the scores, the statuses, and which subtasks count.
Follow the work across screens
Section titled “Follow the work across screens”| When | Where to look | What to do with the result |
|---|---|---|
| While fixing the error | Realtime | Inspect today’s average prompt score, the sessions and subtasks, and the Realtime Advisor. |
| When reviewing the fix | Sessions | Open the Task and compare its status and score with the original prompts and result. |
| When checking usage | Telemetry | Review requests, tokens, durations, errors, and estimated cost. |
| At the end of the day | Daily Recap | Review the day’s ended subtasks. Work in progress is added once it ends. |
| When reviewing a longer period | Report | Open the report for a period and read its narrative. |
Use a suggestion when it helps the next task, and verify that task’s actual result. You can continue asking clarifying questions or correcting a fix without trying to shape the conversation around a score.
For a quick definition while reading another page, use the Glossary. To connect Claude Code or Codex, continue to Quickstart.