Synthesis Map

Purpose

After a round of discovery interviews, the temptation is to leap straight to solutions: "we should build X, Y, Z." That move skips the synthesis step — clustering what users were trying to do and what was getting in their way — and the resulting backlog ends up being a list of solutions to vaguely-articulated problems. The strongest product organizations synthesize opportunities first, then map solutions under each.

This skill produces the opportunity layer. It clusters the interviews into the underlying jobs and pains that, if addressed, would move the named outcome — each cluster traceable to specific quotes in specific interviews. The PM then decides which opportunities to pursue and what solutions to consider under each.

It does NOT produce solutions, features, recommendations, prioritization, or any version of "and here's what to build." That step happens in a separate working session, with the PM and the team.

Procedure

1. Read graph/graph.json for context. If it does not exist, treat the

batch as standalone and note that the synthesis cannot draw on prior graph structure.

1. Read meta.health (if present). Cold-start branch: on a thin

corpus, even a fresh batch may not be enough to produce a stable opportunity map. State this and produce a sparse map honestly — do not generate opportunities a single interview hints at as if they were established.

1. Identify the outcome the synthesis is mapping toward. Sources of

this:

• The Outcome nodes in the current graph (e.g. trial-to-paid

conversion, week-1 activation).

• The strategic memo or PRD intent, if one is in sources/.
• If unclear, ask the PM to name the outcome before proceeding. An

opportunity map without a target outcome is unmoored.

1. Read every file in the fresh batch (typically the most recently

added interviews; the PM may list them explicitly). Extract candidate opportunities — underlying jobs, needs, or pains that, if addressed, would plausibly move the outcome.

1. Cluster opportunities. Rules:

• Cluster only when ≥2 sources independently describe the same

underlying job or pain. A single-source candidate stays as a Single-source opportunity so its evidentiary thinness is visible — do not promote it into a cluster of one.

• Cluster by the user's job, not by the surface mechanic.

"Importing comments fully", "preserving deal history", "keeping audit trail" might cluster as "preserve historical context when switching tools."

• Distinguish opportunities from solutions. "Decide admin

questions before inviting the team" is an opportunity. "Add a four-question setup wizard" is a solution to that opportunity. This skill only emits the former.

1. For each cluster, attach provenance: 2–4 verbatim quotes drawn

from the interviews, with file paths. The provenance must include at least two distinct source files.

1. For each cluster, note its plausible link to the outcome —

one short sentence on the causal story connecting this opportunity to the named outcome. If the link is not plausible, the cluster is noise; cut it.

1. Append a Single-source opportunities section listing standalone

candidates with their one source. These are not clusters; they are leads for the next round of interviews.

1. Emit the output contract.

Output contract

Emit exactly these four sections, in this order.

Outcome

One sentence: the outcome this map is targeted at (drawn from the graph, the memo, or the PM's stated target). If thin or ambiguous, say so.

Opportunities (multi-source clusters)

For each cluster:

• Label — one line naming the underlying job or pain.
• Provenance — 2–4 verbatim quotes with file paths, drawn from at

least two distinct interviews.

• Link to outcome — one sentence on the causal connection.

Order clusters by how much support they have (most sources first).

Single-source opportunities

A short list of single-source candidates that did not cluster. For each: one-line label, the source, and one specific second interview that would test whether this is a real cluster or a one-off.

Open contradictions

Any opportunities that conflict with existing graph nodes or with each other across the batch. Each contradiction stated as a short pair of positions and the files behind them. Do not resolve.

End with one line: "this is the opportunity layer. pick what to explore next; the solutions discussion happens with the team."

Guardrails

• NEVER emit solutions, features, recommendations, prioritization, or

anything resembling "and here's what to build." If you find yourself writing "we could add…" or "the team should ship…", stop. That belongs in a different session.

• NEVER cluster on a single source. The whole point of clustering is

triangulation across sources. A "cluster of one" is a single-source opportunity and lives in its own section, where its thinness is visible.

• NEVER soften a contradiction. If two interviews say opposite things,

the contradiction is the finding. Surface it and let the PM decide.

• Every multi-source cluster MUST include verbatim quotes from at

least two distinct files. If you cannot, demote the cluster to single-source.

• Do NOT invent the link to the outcome. If you cannot articulate a

plausible causal story in one sentence, the cluster does not belong on the map.

• On a thin corpus or a small fresh batch, produce a small honest map.

Padding the opportunity list to seem comprehensive violates principle 3 and produces a worse downstream solution discussion.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Waddling-Penguin
• Source: Waddling-Penguin/mogkit
• License: MIT
• Homepage: https://mogkit.com

Install and usage instructions live in the source repository linked above.