メニュー
Plan the next step as a brief grounded in vision, principles, and experience — then carry it into the spec
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Reconstruct a knowledge base from git as an encoder-decoder — a bidirectional structure pass over the whole history, parallel causal-masked sub-agents that recover each slice's why, a merge, then rumination against git until it converges
Dispatch a discussion or idea into the knowledge base — split by perspective, sort by maturity
Cross-check the knowledge base for coherence and alignment, detect rot, and tidy — proposing, never committing on its own
Human-initiated — consolidate a ripe draft out into the long-term tier (the symmetric OUT to capture's IN)
Create the knowledge base through layered, low-pressure conversation
SOC 職業分類に基づく
| name | knowie-next |
| description | Plan the next step as a brief grounded in vision, principles, and experience — then carry it into the spec |
| user-invocable | true |
| argument-hint | [a direction or feature; empty = infer from vision + recent commits] |
Turn the knowledge base into a structured plan for the next step, and carry the why into the spec tool so it's obeyed during implementation.
draft/ for ripe items first. Before planning, scan draft/ for any topic mature enough to graduate (recognition chain: a design settled enough to commit → its exit is the vision roadmap; an experience/insight verified enough → consolidate). Surface these — never silently build a design straight from draft/; a design must become a roadmap item (the human-commit gate) before it's implemented.knowledge/skills/ — a learned skill that does this step → recommend using it, don't re-plan it from scratch. Missing a relevant one is worse than one extra (retrieval recall is the bottleneck for adherence).Every line cites where it comes from; skip a line if empty.
The next step is the best move, not necessarily the roadmap's sequential item. next may re-prioritize, jump roadmap order, or flag a re-route (a committed route found wrong → say so and make admitting it cheap). Justify the jump from the three perspectives like any recommendation. Guardrail: a new route still passes the gate (draft → roadmap) before it's built — jumping ≠ bypassing the commit gate.
Next: [name] — [one-line] · roadmap position: [phase / milestone]
After the brief, if step 1 found ripe draft/ items, end with the option:
"Before we start, promote these first? (design → roadmap, then build it through the roadmap; insight → consolidate.)" — list them. The human decides; you propose. This is the gate that keeps a design from being built straight out of draft/. (For a focused, human-initiated consolidate outside planning, that's /knowie-consolidate.)
Promoting is a dispatch, not a relocation (see core: Consolidating is dispatch) — direction → roadmap/experience, recurring concept → concepts/, brainstorming scene → episodes/, transition → history/. A new roadmap item carries acceptance criteria (verifiable) and a two-way link to its design draft. The draft is not deleted — it stays as the in-flight design rationale until the item is done (then it reflows + retires, judge §4). Only the pieces already dispatched elsewhere (concept, transition) leave the draft.
If the planned step is a repeated manual operation with no skill yet → flag it a skill candidate (prevention: skill it before doing it manually again; capture records it, the human confirms). next is the earliest, cheapest catch of repetition — before the next manual redo. (judge catches what slips through, later.)
Every recommendation traces to vision / principles / experience. Nothing from thin air — if there's no relevant knowledge for a point, say so explicitly.