Research Conventions

The shared layer beneath every skill in this plugin. Each research skill says what to build; this one says what a figure has to carry, how far the evidence lets the language go, what is worth asking before starting, and how periods, windows and returns are computed. It holds those rules so no skill re-derives them: on evidence, judgement, intake and market-data mechanics this layer is the source of truth, and the calling skill owns what its deliverable contains.

Reference files

• Any deliverable that prints a figure, a number with no source, two sources disagreeing, a figure that may be stale, or how densely to cite: .agents/skills/research-conventions/references/evidence.md.
• A valuation, a thesis, a recommendation, a scenario set, or any section that ends in what to do: .agents/skills/research-conventions/references/judgment.md.
• Before asking the user anything at the start of a task, and when deciding whether to ask at all: .agents/skills/research-conventions/references/intake.md.
• Before the first market-data pull of anything that prints a period label, a multiple, a growth rate or a return: .agents/skills/research-conventions/references/market-data-rules.md.
• Choosing how long a deliverable runs, a draft over its length band, or a request for a short version of something already built: .agents/skills/research-conventions/references/depth.md.

Readiness posture

Every deliverable states one posture, once, near the top, in the reader's language. Posture is about inputs, not conviction: a high-conviction view resting on stale inputs is not decision-grade, and saying so is what makes the rest of the artifact usable.