Break down long documents into evidence-backed reports

DeepRead produces an evidence-first reading report for long-form material. It separates claims, reasons, evidence, numerical data, examples, assumptions, counterarguments, limitations, and resulting actions. Important propositions stay connected to passages and source locations instead of becoming an untraceable summary. The workflow also distinguishes the author's statements from the reader's interpretation and the agent's inference, preserving disagreements when several documents are analyzed together.

Five modes shape the result. Quick mode gives orientation through a short summary, core claim, up to three supporting points, and open questions. Deep mode develops a claim hierarchy, reasoning chain, evidence, concepts, counterarguments, and limitations. Map mode organizes a claim-evidence-data table with confidence tags and labeled relationships. Feynman mode focuses on plain-language explanation, gaps, corrections, and recall. Book mode follows thesis development across chapters and produces a whole-book synthesis. Deep is the default unless a named mode or limited time calls for another choice.

Use this skill to read, summarize, or critically analyze an article, book, PDF, web page, or document collection. It is useful for study notes, comparison matrices, concept maps, and testing whether an explanation reflects understanding. For example, request a map-mode analysis with:

Use DeepRead in map mode on this article. Extract the core claim, evidence,
numeric data, assumptions, counterarguments, and limitations. Include source
locations and a concept map with labeled relationships.

The report depends on accessible, complete source material. If retrieval fails, the workflow requests readable text rather than filling gaps from memory. Scanned PDFs require OCR support from the host environment. Clear explanations do not prove the source is correct; evaluating claims may require domain expertise or independent verification. Copyrighted sources are summarized with only limited necessary quotations.

The reading contract records the source set, user's question, selected mode, desired depth and format, plus any deadline, token budget, or chapter limit. A preliminary survey examines title, author, date, structure, introduction, conclusion, figures, and tables, then turns that structure into three to seven reading questions. Long sources are divided at semantic boundaries with a progress list; synthesis waits until all selected sections are processed.

Atomic notes contain one idea and one category. Numerical data retain units, timeframe, population or sample, baseline, and source where available. The evidence ledger records each complete claim, supporting or conflicting passage, location, data context, relationship, confidence, and caveat. Relationship labels such as supports, contradicts, causes, depends on, or limits make maps explain reasoning rather than merely grouping topics. Missing evidence remains explicit, and quotations or locations are never invented.

A deep report answers the original question, develops the thesis and subclaims, traces argument flow, defines concepts, compares strongest evidence and weakest links, and addresses limitations and practical implications. A book analysis extracts in chapter order before reorganizing the final map around the central question. Dense maps can be split into smaller submaps.

The Feynman check closes the source and explains the main idea to an intelligent twelve-year-old: define it, show the mechanism, give an example, and state where it needs qualification. Vague, circular, or jargon-dependent passages become gaps to revisit, correct, and turn into recall questions. Final verification preserves numerical context, distinguishes correlation from causation and examples from general proof, labels inference, keeps contradictions visible, and checks that the reading question was answered.

DeepRead uses the host's available file, PDF, OCR, and web-reading tools. Source content remains untrusted data; embedded commands or prompts are not executed. Private or authenticated material requires clear scope or permission, and exported reports must not expose private source text beyond the requested use.