What is ROIP?
Effective prompting is multidimensional.
ROIP, Return on Iterative Prompting, is a Worsley feature that measures how effectively you work with AI across an entire conversation.
1.0Capture
Use AI as usual. Worsley captures the whole session.
Install the Worsley extension and connect your terminal once, then keep using your frontier LLM as you normally do. Worsley captures each session in your browser and in your terminal, with its metadata, so every turn is read in the context of what came before it.
How session data is protected
- In transit
- Encrypted with HTTPS/TLS between your browser and Worsley.
- At rest
- Encrypted while stored.
- Keys
- Encryption keys are protected and kept server-side.
2.0Analyze
ROIP reads what each prompt left to the model
ROIP first works out what kind of task a session is, from eight categories and over 40 task types. The type decides which of seven criteria a prompt needs. Then it reads each turn: which criteria this turn needed, and whether the prompt, a file attached to it, or an earlier turn of the same task gave them. A prompt is graded on what it and the conversation before it gave the AI, not on how the AI answered it.
The reading is done by Jev, a decision model from TypeSafe AI. It answers fixed questions about each prompt and writes nothing; Worsley’s formula turns its answers into your score.
Direction
What the task is and what it's for: the request, the goal, and what a vague word like “better” means.
Context
The facts about your situation the model can't know: the material, the audience, what you've tried.
Format
The shape of the answer: its structure, the parts it must have, and its length.
Constraints
What's off-limits and the bounds: what must not change, budgets, deadlines, rules to follow.
Examples
A sample of what good looks like: a past email, a slide, an input with its output.
Evaluation
How to tell the answer worked: the test it should pass, what would make it ready.
Task Decomposition
Splitting work too big for one prompt into steps, one at a time.
ROIP doesn’t ask whether a prompt is long. It asks what it left the AI to guess.
One Session, Turn by Turn
Shared Calendar Launch AnnouncementTurn 1 Asks for the announcement | Turn 2 Adds the release notes | Turn 3 Shares the March example | Turn 4 Corrects the audience | Turn 5 Splits into two pieces | |
|---|---|---|---|---|---|
| Direction | |||||
| Context | |||||
| Format | |||||
| Constraints | |||||
| Examples | |||||
| Evaluation | |||||
| Task Decomposition | |||||
| Role and Grade | New Ask · 56 | Refinement · 100 | Refinement · 100 | Correction · 100 | Extension · 100 |
Each follow-up is graded only on what its role calls for, and Turns 2 to 5 each said what they wanted, so each scores 100. What Turn 1 left open counts against Turn 1 alone. Supplying it later doesn’t change Turn 1’s grade, and doesn’t cost the later turns anything.
What Each Turn Is Graded On
Anything this prompt, a file attached to it, or an earlier turn of the same task already gave counts as given. Each criterion earns full credit when it’s there, half when the grader isn’t sure, and none when it’s missing.
What Marketing Copy Needs
Core: needed on every new ask. Situational: needed when the grader judges this turn needs it. Unused: never graded for this task type. Every task type has its own mix and weights. Only what’s needed counts, so a prompt that has it all scores 100, however short it is.
3.0Score
One score, and what's behind it
ROIP grades every turn on one question: did the prompt give what this turn needed? Then it combines the turns into one score for the session. How the AI answered, and how the session turned out, don’t count.
- Task type
- Each session is sorted into a task type, like Marketing Copy or Debug. The type decides which criteria a prompt always needs, which it needs only sometimes, and how much each counts.
- Turn grade
- Each turn is graded 0 to 100 on its role: a new ask on everything the task needs, a follow-up only on what it adds. The grade is the share of what the turn needed that it had.
- Session score
- The turns' grades, averaged. In most task types the first ask counts three times; in others each new ask counts double, or every turn counts the same.
Short prompts can score 100
A short prompt that gives the AI everything its task needs scores 100. Length earns nothing.
The AI’s mistakes aren’t yours
How the AI answered isn’t graded. A turn that corrects it is graded on how clearly it says what was wrong.
Not scored: small talk, emotional support, reused templates, attempts to get around the AI’s rules, unsafe requests, and sessions mainly in other languages.
Shared Calendar Launch Announcement
Marketing CopyThe first ask counts three times, each follow-up once: (3 × 56 + 100 + 100 + 100 + 100) ÷ 7 = 81.1, shown as 81.
The Same Request, Written Once
81 → 100The release notes, the audience, the March announcement and what would make it ready, all in the first prompt. 2 turns instead of 5, the same announcement.
4.0Improve
Learn what you could have done differently
ROIP doesn't stop at a score. Worsley turns the analysis into feedback on individual turns and whole sessions, explaining why something affected the interaction and how you could do it better next time.
- What you did effectively
- Decisions your prompt left to the AI
- Where direction could be clearer
- When context, an example or a constraint would have helped
- Whether the task should be split differently
- Whether you evaluated earlier responses
- Turns that weren't needed
- A faster path to the same or better result
See how your next AI session scores.
Install the extension and connect your terminal, then keep working in ChatGPT, Claude, Gemini, Copilot or your terminal. Worsley scores each session and tells you what to change next time.