Menu
Run a two-proposer deep-research workflow with an aggregating referee and an evidence-graded report.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
Build, launch, and drive the mindwalk web UI end-to-end for verification.
Set up or troubleshoot Yassimba's curated agent skills, Pi packages, Herdr, and Herdr plugins through the ai-setup CLI. Use when the user asks to install this collection, configure Herdr, add one of its capabilities, update the setup, or diagnose installation problems.
Backlog management: use when the user mentions a backlog, asks what's next, wants work recorded before implementation, or wants queued ideas or specs prioritized, transitioned, completed, or removed; also use when another skill needs to record lifecycle changes.
Use when the user wants to brainstorm or explore an idea — a feature, product direction, or "what if" — before deciding whether it deserves a plan. Ideation only — ends in an idea brief, not a design.
Review the changes since a fixed point (commit, branch, tag, or merge-base) along two axes — Standards (this repo's documented coding standards) and Spec (what the originating issue/PRD asked for). Use when the user wants to review a branch, a PR, work-in-progress changes, or asks to "review since X".
Use when creating a git commit — the user asks to commit, or a unit of work is complete and ready to commit.
| name | deep-research |
| description | Run a two-proposer deep-research workflow with an aggregating referee and an evidence-graded report. |
| disable-model-invocation | true |
Start a deep-research run for a focused question with:
subagent({ action: "deep-research.start", task: "<research question>" })
What it does:
The final Markdown report is saved collision-safely under the project's research directory (ai-docs/research by default) with inline source links, publisher/type, primary status, confidence, fact-vs-inference labels, rejected weak sources, and remaining gaps.
Capability safety: the run requires exact capability grants for every role/workflow/executable pair. In the TUI you will be asked to confirm and record the grants; headless runs without a pregrant fail closed. All fetched web content is treated as untrusted (prompt-injection boundary).
Limitations: the run currently executes in the foreground and blocks the parent turn (default budget: 10 minutes per proposer, 10 minutes for the aggregator, 20 minutes total). Do not start it automatically; it is user-invoked only.