ManuelConejo

VanishingPoint DeepResearch Skill — Development skill for Claude Code

Development community

Deep-research skill for Claude Code and Codex: date-bounded search over open scholarly, legal, financial and code sources, argued report with numbered bibliography, separate verification with the exac.

How to install VanishingPoint DeepResearch Skill

This entry records only its repository, not the path inside it, so there is no exact command to give. Open ManuelConejo/VanishingPoint-DeepResearch-Skill and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What VanishingPoint DeepResearch Skill does

Alternatives in Development

Opencode Claude Auth — OpenCode plugin that uses your existing Claude Code credentials — no separate login needed 903 ★
Discussion Follow Up.V1 — At the end of your response, provide 3-4 numbered candidate next 272 ★
Bug Health Check — Execute the bug-health-inline skill for inline orchestration 172 ★

README

VanishingPoint (VP) - a deep-research skill with discipline plugins

VanishingPoint takes a question (or a project), searches the literature and the code through a date-bounded gateway, triangulates, stops on saturation, writes a report in argued prose with a numbered bibliography, derives a machine annex from it, then verifies the claims that carry a decision in a separate, fresh-context call. Then it stops. It writes no code and tracks no project. Disciplines are text plugins (`plugins/.md`): cs-ai, code, finance, biology, law, plus a template to add one.

**This repository is the product and the source of truth.** It is what a user installs; nothing else is needed. The benches that measured it are not part of this repository and are not required to use the skill.

Status: v0.6.0, research preview. Measured on small private benches (3 to 50 cases); no public benchmark yet; see "What is measured, what is not" below and CHANGELOG.md.

In two minutes

You give: a question and, optionally, a date T ("what was known before T").
It does: 3 to 5 searches through a gateway that only returns documents dated before T (arXiv, OpenAlex, Crossref, PubMed, Europe PMC, OpenReview, code hosts... no key required), stops on saturation, writes the report, then a separate fresh-context call re-reads the sources and checks the claims that carry a decision.
You get, in one directory: REPORT.md / REPORT.html (eight sections, numbered bibliography, each checked claim shown with the exact source passage), card.json / result.json / verdicts.json (machine annex and verdicts: FOUND, DISTORTED, ABSENT, INVALID, SOURCE-UNAVAILABLE, UNVERIFIED), studies.csv and references.bib / .ris.
It costs: about 3 USD of model usage per academic question on Claude Code (2.8 to 3.0 USD measured, verification included), up to about 10 USD on long legal texts, more with the optional full-text verification (see the example); 0 USD metered on a