What counts as a good task description?

What the task field signals to the analysis

The task description is injected into the Phase 2a prompt alongside your role and level. It shifts the pipeline's framing from "what does this page need" to "what does this reader need to accomplish using this page." The task field is the highest-signal input in the survey: it is what unlocked a stably different analysis in a 30-run controlled study, including tighter guidance, higher priority elevation on directly relevant concepts, and a stable additional concept that other conditions did not reliably produce.

Phase 2 requires an analysis credit. The task you enter determines the framing of that analysis.

Strong vs. weak task descriptions

The difference between a weak and a strong task is the difference between a goal and a workflow. Three matched examples:

Weak: "improve my content." Strong: "identify LLM retrieval gaps in technical documentation." This exact task description was used in the ContentGrapher study. It produced the tightest, most targeted results in the dataset.

Weak: "understand the tool." Strong: "evaluate whether this tool fits our content workflow before a vendor review."

Weak: "content audit." Strong: "find structural gaps before a quarterly content audit and prioritise fixes by impact."

The first pair comes directly from the study. The second and third are illustrative examples of the same pattern, not study-tested conditions.

The pattern: verb, object, context

A strong task follows a three-part structure. Verb: what the reader will do (identify, evaluate, find). Object: what they are operating on (LLM retrieval gaps, structural gaps). Context: where or under what constraint (in technical documentation, before a vendor review).

Goal states such as "understand," "learn," or "improve" are weak because they do not define an output. A strong task has a concrete deliverable: a list, a decision, an evaluation, a document. The task "identify LLM retrieval gaps in technical documentation" breaks down as: "identify" (verb), "LLM retrieval gaps" (object), "in technical documentation" (context).

If you entered a weak task

Re-run Phase 2 with a stronger task description. Phase 1 does not need to re-run: the concept map and boundary classifications it produced are reused. Re-running Phase 2 uses one analysis credit.