Case 3 — From raw data to a research plan

By Tan Haosheng, MD, PhD · Last reviewed 2026-07-31

Task type · forward planning~15 min wall-clockWorkBuddy Hy3One prompt

Cases 1 and 2 had me checking work that already existed. This case flips it: no draft to defend, no manuscript to audit. The prompt is "here is the data, what can I realistically write from it?" A different kind of test — origination, not validation.

⚠️ About this page

The underlying dataset is unpublished. This page reports only the agent's observable behaviour: what kinds of directions it proposed, how it scoped the dataset's ceiling, and how long it took. The science stays out of frame.

The task, paraphrased

Here is the dataset. Don't try to validate any specific claim — tell me what could realistically be written from this, and where it could go. Be honest about what this data can and cannot support.

One prompt. No pre-structured output format. The agent had to invent the framework, then fill it in.

What the agent did (behaviour only)

Observable sequence

Opened with the dataset's ceiling before proposing anything: single patient, single batch, limited resolution — name that first.
Returned several differentiated directions rather than one "best" answer — different scopes, different ambition levels, different data demands.
For each direction, named what additional data or work would be required before it could be written, and roughly what tier of venue would fit.
Did not inflate the scope to look more useful. Marked which directions would not be publishable without external data and said so.
Refused to round the single-case dataset into a cohort claim.

No direction's substance, tier or venue is disclosed on this page. The figure below is illustrative only.

A blurred snapshot of the agent's output window, shown only to indicate that structured output was produced — **Why this image is blurred.** Same principle as the other cases: every label, axis, value and word has been removed at the pixel level. What you are looking at is only the *shape* of the agent's output — that it produced structured text — not what it said.

Time vs. my normal workflow

For me, "what should I write next?" is the most expensive question of all — usually a week of half-formed thoughts before any commitment. The agent produced a ranked short list in one prompt.

Bar chart comparing manual workflow time vs. one WorkBuddy Hy3 agent run for three research tasks — **Manual workflow vs. one WorkBuddy agent run, per task.** The Task 3 column is the forward-planning run from this case. Manual time for this category is the most variable of the three; the figure shows a typical-week estimate.

What this case is — and isn't

Is: a record that an AI agent can produce differentiated, honestly-scoped forward plans from raw data without inflating them.
Isn't: a substitute for editorial judgment, statistical consultation, or a domain mentor who knows the literature.

Read the methodology behind all three cases →