finance-research-agent: Skill for Claude Code

Sci-Write

Purpose: produce scientific text a tired expert reviewer trusts. Two failure modes kill drafts: shallow content (sentences that sound analytical but assert nothing checkable) and AI-pattern prose (the surface tics reviewers now recognize instantly). This skill guards against both, in that order — content first, surface second, because de-slopping a hollow paragraph just yields a hollow paragraph.

The user's standard (non-negotiable, applies to every draft)

1. Precise flow of thought. Each sentence opens with what links back (old information) and ends on its new point; the next sentence picks that point up. This stress-to-topic chaining is what "flow" mechanically is — if two adjacent sentences could be swapped without loss, the flow is broken. Full mechanics: references/writing-craft.md §A.
2. Deep technical engagement — never shallow. Every paragraph must carry a checkable claim: a named method, dataset, quantity, or citation, and what exactly it shows. When reviewing literature, state what each work actually did and where it stops, not that it "highlights the importance of" something. If you don't know the specifics, say so and ask or research; do not pad with generalities.
3. No AI slop. Run every draft through the checklist in references/ai-writing-signs.md before calling it done.
4. No em-dash unless essential. Essential means a genuine interruption or appositive that commas, parentheses, a colon, or two sentences would mangle. Default to those alternatives.
5. Clarity above cleverness. Plain verbs, subject next to verb, one point per unit, the same term for the same concept every time (a synonym implies a different concept). If a sentence needs rereading, rewrite it.

Workflow

1. Identify the document type and venue first. Paper vs. proposal changes the architecture (below). For the four fellowship schemes, read references/funding-schemes.md — each has hard page limits, mandated section orders, and named rejection causes; text is scored against the funder's template bullets, not against abstract quality.
2. Outline before drafting: one sentence per intended paragraph, in order. Check the outline carries the argument by itself. Fix structure here — vague feedback on a draft almost always means broken story structure, not bad wording.
3. Draft to the architecture below, writing from the reader's expectations (references/writing-craft.md).
4. Revise in two separate passes: first content (every paragraph's checkable claim), then the de-slop pass (references/ai-writing-signs.md, which ends with the full 5-step revision procedure).
5. Verify every citation against the actual source — existence, metadata, and that it supports the exact sentence citing it. Fabricated or mismatched citations are fatal in review, and they are the most common LLM citation failure.