1. LandOffer
  2. Blog
  3. All articles
  4. iMocha AI-Assisted Coding: Attach the Right Context and Keep Chat Separate from Submission

Job search

iMocha AI-Assisted Coding: Attach the Right Context and Keep Chat Separate from Submission

Two original completed context and local-code verification artifacts distinguish permitted built-in hints, chat attachments, actual editor code, and final assessment submission.

9 min readLandOffer team

iMocha’s AI-Assisted Coding assessment places a hint-only assistant beside the code editor. In that configured assessment, using the built-in assistant is allowed and assessed. The important boundary is context: it already knows the problem statement, but it cannot see your code, run results, or screen unless you share relevant material in the chat. A chat attachment is not your code submission.

This guide uses public iMocha documentation inspected on October 21, 2026. No private invitation, assessment, account, assistant conversation, attachment, or submission was opened or performed. LandOffer publishes this guide and offers job-search tools. Keep preparation relevant to the role; browse recent roles on LandOffer.ai if you need to define that target.

A developer separates a code worksheet from an assistant-context note at a desk.

Confirm which assessment you have

The candidate AI-Assisted Coding guide describes three areas: the question on the left, the editor with run and submit controls in the center, and the collapsible assistant panel on the right. The panel belongs to the assessment when enabled; it is not a recommendation to use an unrelated external AI service.

The administrator guide says the setting is per test and applies at invitation level, with the default off. Enabling it does not retroactively rescore past attempts or enable assistants everywhere. Its availability on one assessment therefore says nothing about another employer’s rules.

The assistant helps explain errors, clarify concepts, discuss approaches, and suggest investigations. The candidate guide says it will not write a passing solution, judge whether your answer is right, explain scoring, or answer unrelated questions. A refusal to provide the solution is consistent with this documented design. Do not try to bypass that boundary.

If the panel is absent, the public candidate guidance says to treat the assessment as a standard coding test. Read the actual invitation and section instructions. Ask the assessment provider through its permitted channel if the instructions are inconsistent; this review did not contact anyone or inspect an individual invitation.

Attach the current evidence, not everything you have

The candidate guide describes typing @ in the assistant composer to attach My current code or Test cases, including their actual output. The documented attachment choices provide context to the built-in chat. They do not establish permission to upload arbitrary employer files or a private repository.

Because the assistant cannot see your editor automatically, “Why does this fail?” may omit the thing it needs. Supply the relevant current version, what input you used, what output you expected, and what actually happened. Describe changes made after an earlier message instead of assuming the assistant can infer them.

The documentation says attachments go to the assistant as chat context. They do not alter the problem statement or count as a submission. That means there are two separate checks: whether the chat has the evidence needed for useful guidance, and whether the editor contains the code you intend to submit.

This feature should also be distinguished from iMocha’s separately documented Prompt Engineering question type. Its workflow and attempt rules should not be imported into Coding. The Coding candidate guide currently describes no message-count cap and a 15,000-character limit per message. The actual question and assessment instructions still determine what work you are being asked to complete.

Worked asset: a completed context-selection brief

Here is an original fictional preparation scenario for Anika. It is not an iMocha question or a prompt sent to its assistant. Her local exercise asks for the most recent status for each item, using the largest integer event tick, with results sorted by item identifier. Input order is arbitrary; ticks for a given item are distinct. The function must not mutate its input.

Anika’s draft keeps whichever event it encounters last for each item. It appears correct when events arrive in increasing tick order, but fails when newer events appear first. Her completed context brief chooses exactly what a hint request would need:

Material available in the fictional scenario Completed decision Reason
Current draft implementation Include the relevant code through the assessment’s documented code-context mechanism if available. The error depends on how an existing value is replaced.
Input A,9,ready followed by A,2,queued Include this small invented failing case and actual local output. It exposes the difference between arrival order and event tick.
Expected result A,ready State explicitly with the rule “largest tick wins.” Otherwise the assistant could explain the wrong interpretation of “latest.”
An unrelated real production log Exclude entirely. It is unnecessary, potentially confidential, and not part of this exercise.
A later editor change not yet run Do not present it as the tested version. The attached code and reported output must describe the same state.

Her finished, unsent request is: “This exercise defines latest by the largest tick, not by list position. My attached draft returns queued for the two-event case, but I expect ready. I suspect that replacing the stored value on every iteration ignores the tick. Give me a hint about what comparison I should inspect, without writing the solution.”

This is a context brief for the permitted hint-only workflow. It does not ask for a hidden answer, an evaluation grade, or confidential test cases. In a real assessment, Anika would use the actual allowed context controls and instructions; the fictional request was not pasted into any service.

Anika leaves attachment refresh after a later editor change explicitly unverified in her brief. Before a follow-up, she would check the visible context and describe the new version rather than assume a previous attachment silently refreshed. That is a review precaution, not an observed iMocha refresh defect.

A candidate reviews separate chat, editor, and submission records on paper, with both hands on the desk.

Worked asset: a local correction and a submission-state record

We executed the original toy exercise locally in standard Python for this article. It uses only invented tuples, not an employer’s question, assessment data, or an iMocha runtime. The first implementation simply replaces each item’s stored event on every iteration. The revision replaces it only when the incoming event has a larger tick, then sorts the final item identifiers.

The actual local results were:

Authored local case Draft result Revised result
Empty input Passed Passed
One event Passed Passed
Two events in increasing tick order Passed Passed
Newer event followed by older event Failed Passed
Interleaved items with out-of-order ticks Failed Passed

The saved local run records five passing revised cases and confirms that each input remained unchanged; these are authored fixture results, not platform results. The draft passed three. Equal ticks for the same identifier were outside the declared exercise contract; this does not demonstrate a solution for unspecified tie handling, every possible input, or an employer’s hidden tests. No iMocha assistant response was obtained, so the correction is not attributed to an actual iMocha hint.

Anika’s completed reasoning note is: “The stored event should be the maximum-tick event seen so far for that identifier. Reading an older event later must not replace it. Sorting the final identifiers establishes the requested output order; it does not decide which event is newest.” This explains the correction instead of treating a passing count as sufficient evidence of understanding.

Her separate state record deliberately keeps four statuses apart:

Work artifact Status in this completed example
Chat request Authored but never sent; no iMocha attachment or response exists.
Revised code Authored locally and checked against the five declared cases.
Assessment editor Not opened; no code was entered there.
Final assessment submission Not performed; local test success is not a submission receipt.

In a real enabled assessment, an informative conversation would not move the last two rows forward by itself. The candidate still needs to review the editor’s actual code and use the assessment’s own run and submission controls as instructed. Do not describe an answer as submitted merely because it appears in chat or because an attachment carries the right code.

Tracking the version behind each artifact helps prevent a checked local revision from being confused with the code actually in the assessment editor. A hint may lead you to change your implementation, but the current chat context, the last run, and the editor’s final contents may represent different versions unless you check them. No claim about automatic saving of unsubmitted changes is established by our local exercise.

Understand the two evaluations without optimizing for message volume

The public coding-question authoring guide describes code scoring through the assessment’s defined test cases. The AI-assisted guides separately describe evaluation of the candidate’s context, reasoning, and use of hints. Good code and effective AI collaboration are separate readings; one does not prove the other.

The administrator documentation says the AI report can include the conversation, written summaries, token usage, and grades for Context Quality, Critical Thinking, and Collaboration. It also says not engaging the enabled assistant produces no collaboration evidence and is assigned the lowest AI-fluency level, while coding remains separately scored. Do not infer that avoiding the built-in assistant is required for this particular configured question type.

Use follow-ups to advance the investigation rather than to increase the message count. A useful follow-up reports what you tried after a hint and whether the evidence changed your hypothesis. Repeatedly asking for the full solution is outside the documented assistant’s purpose and does not demonstrate that you evaluated its guidance.

These descriptions are not an employer’s hiring decision, a guaranteed grade, or a complete disclosure of every scoring weight. No report was obtained and no grading or employer decision was observed, so the documented scoring categories remain descriptions rather than validated predictions.

Chat permission does not establish private or unproctored work

The admin guide says the candidate–assistant conversation is available in the report. Treat it as assessment material, not a private scratchpad. Keep personal credentials, unrelated company files, and confidential examples out of messages and attachments.

It also says AI-Assisted Coding does not add or change the test’s existing proctoring configuration. The report’s Proctoring Details reflect the settings already present. This does not establish that the test has no camera, screen, browser-event, or other monitoring. Review the actual assessment’s notices and permitted actions rather than inferring privacy from the presence of an allowed AI panel.

The test-settings guide documents employer configuration of the AI option and other assessment settings. Candidate access to the assistant does not authorize changing settings, using external tools, or moving assessment content into another service. We did none of those things.

If the assistant fails, keep the code state separate

The public guides say coding work can continue when the assistant is unavailable: candidates can still write, run, and submit code. They advise checking connectivity and, if a problem persists, reporting the test name and time through support. They also refer to saved submitted answers. That is not a blanket guarantee that every unsent message or unsubmitted editor change has been preserved.

Use the actual permitted recovery instructions and keep track of which work was submitted. Do not start another private assessment attempt or duplicate an invitation merely because an assistant panel failed. This review did not simulate an outage, refresh a live test, contact support, or establish any personal extension or retake entitlement.

The useful preparation habit is concrete: share the relevant current evidence, reason about hints, verify changes, and check the editor’s submission state independently. Apply those habits only within the actual assessment’s permissions. Keep the broader job search focused by browsing recent roles on LandOffer.ai.

Sources and Further Reading