choplin/my-agent-skills/inception-deepen

Inception — Deepen (深掘り)

Assumes inception-base is loaded (graph model, CLI, dig rule). Phase map: inception-base/references/contracts/phases.md.

Stance: adversarial. Attack premises; be the devil's advocate. Drain the queue one foundational point at a time, but pressure-test before closing. This is where thinking actually advances.

How to run it

• Pick the next point with the CLI's next — let the dependency graph choose the most foundational unblocked node, don't guess.
• Pressure-test: 5 Whys, premise attack ("what must be true for this to hold?"), dialectic (thesis → antithesis → synthesis), multiple perspectives, red-teaming. Record strong objections as Counter nodes.
• Act on the node's nextMove:
• decide → surface enumerable, mutually-exclusive options (use AskUserQuestion) and get a choice
• investigate → name what to find out and who/how
• validate → find the cheapest way to test the assumption
• deepen → keep discussing; not yet resolvable
• Draw the user's own reasoning out via discuss-toolkit-dig — never close on the AI's assumption.

Closing a point (the core rule)

A point is closed only when its outcome is recorded in the graph:

• a choice → a Decision node with chosen, rejected[] (option + reason), and rationale
• work → an Action node
• a learning → an Insight node

Then mark the originating question resolved. Never close with an assumption — it looks decided and silently blocks later work; leave it open instead. Deepening legitimately spawns new Question/Idea nodes — add them; a growing queue mid-discussion is healthy.

Exit

When the remaining open nodes are non-foundational or deferred, propose moving to Converge (inception-converge). The user approves the transition.