EVO CAPABILITIES / Intent, clarification and completion
Turn a rough request into a useful outcome.
EVO starts with what you are trying to achieve, asks only when a missing answer matters and keeps the final result separate from a successful tool call.
Mac development · See current scopeThe real problem
Make this better is a normal human request. It might mean clearer writing, stronger conversion, a simpler mobile layout or fixing a broken control. Acting on the wrong interpretation can produce polished work that misses the point. Asking ten questions before doing anything creates a different failure: the user becomes the project manager for every small decision.
What chat history leaves you to do
An answer can sound responsive while solving a neighbouring problem. Conversation alone also makes it easy to confuse three milestones: the assistant understood a request, a tool returned successfully, and the desired result was achieved. Useful assistance needs room to clarify the first and evidence to assess the last, without making every exchange feel like a questionnaire.
How EVO approaches it
Start with the request and what is already known
EVO forms the task from the current instruction and relevant prior user facts before considering a question. An explicit audience, output format or constraint should not be asked again merely because a new answer is being prepared. Small reversible preferences can use a reasonable default; uncertainty should be visible when it changes the usefulness of the result.
Ask when the answer changes the work
A clarification card is optional. It is intended for decisive missing information, conflicting requirements or interpretations that would produce materially different outcomes. The card offers unselected choices and a free-text answer. A greeting, complete calculation or clear translation request should not open a card simply because the assistant has more questions it could ask.
Keep independent work moving
While a question is open, the supported flow can continue authorised reading or calculation that does not depend on the answer. It can return a provisional result with assumptions and missing inputs clearly identified. A timely answer can enter the next existing stage; an answer that arrives later is saved for you to bring into a new request.
Judge the deliverable against the goal
The delivery review puts recorded results beside the work's requirements and shows what remains unknown, pending or invalidated. A successful calculation is evidence for the calculation, not proof that the business decision is settled. A file being written is not proof that the page looks right. Necessary checks remain part of completion.
A concrete example
ILLUSTRATIVE WORKFLOW · NOT A CUSTOMER RESULT
Illustrative workflow: improve a product page
You provide a page and say it feels wrong. EVO can inspect the material while a genuinely important question remains: is the primary audience individual users or purchasing teams?
Review the existing content and identify concrete gaps that matter under either audience.
Ask one focused audience question if the answer would change the proposed message.
Deliver a preliminary recommendation, label the assumption and leave the audience-dependent wording open for refinement.
What you take away
The intended output is useful progress plus a precise remaining decision. Your answer guides the next revision without turning silence into an agreement or presenting the preliminary draft as final acceptance.
Your choices
- Select an answer, write your own, answer later or skip; questions do not take over the whole workspace.
- Correct EVO's interpretation in plain language and review the next request before it runs.
- Keep approval separate: an answer cannot grant file access, cloud sharing, sending or purchasing permission.
Current scope
Implemented foundation
- The Mac development source separates conversation, memory proposals and requested actions.
- Recent real local-model answers still included unwanted advice, incorrect facts and unfinished responses. A valid structure does not establish a useful or complete answer.
Next milestones
- Public rollout and real-model evaluation of when to ask are still pending. A tested card mechanism does not prove the assistant always recognises the right intent.
- Forward suggestions are hypotheses about useful next steps, not mind reading. Consequential action still depends on its own supported workflow and required approval.
These are development capabilities, not a public release. See platform availability before requesting a trial.
Platforms and availabilityRelated questions
How does EVO decide whether a task is finished?
EVO separates producing an answer from meeting the Work’s agreed success conditions, so remaining checks can stay visible after a response ends.
Where supported, completion review connects the requested outcome to actual evidence and asks you to inspect unresolved limits before accepting it.
A supplier shortlist may be written while one important source still needs confirmation; a website file may be saved while its mobile behavior remains unchecked.
Define the result you need, review the evidence and decide whether the remaining limitations are acceptable.
Not every task has an automatic outcome check. Real-world effects and subjective quality can still require your judgment or additional testing.
What if a tool succeeds but my goal is not achieved?
A successful tool result should count as evidence of that operation, not proof that the whole task is complete.
EVO’s supported workflows distinguish a proposed change, an applied change and an outcome that has been checked against the original request.
Saving a new button label proves the file changed; it does not prove visitors understand it or that the button works on mobile.
Inspect the saved result, identify the unmet condition and request the next focused check or correction.
The system can preserve this distinction, but model judgment may still miss a requirement. Important outcomes need direct verification.
How does EVO work out what I really want to finish?
EVO starts with your current request, relevant earlier facts and explicit corrections, then relates them to the outcome you are asking for.
It should distinguish the goal from one suggested method, use known constraints and identify missing information that could materially change the result.
“Make this page clearer” may call for a better product explanation before visual decoration; EVO should keep that interpretation provisional until supported.
Correct the goal, provide a constraint or choose an answer when a decisive question is offered.
This is task interpretation, not mind-reading. Real-model intent accuracy still needs further evaluation, and inferred motives must not replace your explicit request.
When will EVO ask a question instead of proceeding?
Only after interpreting the task, when a missing answer could materially change correctness or usefulness. Clear requests should not trigger a questionnaire.
The development flow checks for supplied facts and useful low-impact defaults before allowing an optional clarification card.
An unknown workshop audience may change the whole agenda; an unspecified minor formatting preference usually should not stop an initial draft.
Choose an option, enter your own answer or leave the question for later. A question card cannot authorize an action.
The flow is implemented, but live-model question quality remains under evaluation. It may still ask unnecessarily or miss a useful question.
Must I keep the conversation open while EVO works?
Reliable background execution across a closed conversation is still an internal experiment. Do not rely on it for unattended work in a public release.
Saved results and full background execution are different capabilities.
See release status before planning a long-running task.
What happens if EVO guesses my intention incorrectly?
Your explicit correction should take priority over the conflicting interpretation while preserving requirements you have not changed.
EVO can revise the brief and produce a more appropriate next result. A tentative interpretation should remain separate from a confirmed user fact.
If “simplify” meant fewer steps rather than shorter text, say so; the next result should address the workflow instead of repeatedly trimming words.
State the intended outcome or answer the relevant clarification. You can also review a captured correction separately.
A correction does not guarantee the model will reason perfectly next time, and it cannot undo an already completed external action.
Try the workflow
Bring a task you want to move forward.
Explore the Mac release, or tell us which workflow you would like to evaluate. An application does not guarantee an invitation.