deep-research-article（深度研究文章）

Deep Research Article

Overview (primary artifact)

Produce:

1. article.md: one coherent article with strong reasoning and transitions

This skill is intentionally topic-agnostic. If the user later wants an image-only deck, pass the approved article directly to infographic-powerpoint-deck, which can ingest raw articles or structured slide tables.

In internet-backed mode, also produce a research package:

• source_dossier.md
• evidence_extract.md
• research_notes.md
• claim_evidence_ledger.md

Workflow

Step 0 — Confirm constraints and research mode

Ask only what’s necessary:

• Audience + use case (sermon, lesson, blog, internal memo)
• Desired length (minutes or word count)
• Language (CN/EN/bilingual) and tone (academic / pastoral / simple)
• Success criteria (what a “good final article” must achieve for this user)
• Explicit non-goals/out-of-scope topics to prevent drift
• For Bible: translation/version and whether to include cross-references
• Must-use sources or explicit constraints (paywalled-only, region scope, date range)
• Whether web research is prohibited by user/policy

Default mode is internet-backed deep research. Only skip internet if user explicitly asks for no-web research. If the user does not specify a length, do not default to a short article. Read references/longform_depth_standard.md and use the long-form depth defaults there.

Step 1 — Internet source sweep (required by default)

Follow references/web_research_protocol.md and create source_dossier.md using references/source_dossier_template.md.

Minimum source counts:

• Normal: at least 8 sources
• Deep: at least 12 sources

Coverage requirements:

• At least 2 source types (e.g., primary + secondary)
• At least 2 sources that present meaningful tension/alternative interpretation
• For time-sensitive topics: include recent sources and clearly note dates

Drafting is blocked until this step is complete.

Step 1a — No-web fallback (only when explicitly requested)

If the user explicitly forbids web research:

• Use only provided/local materials.
• Build source_dossier.md from local sources.
• Mark mode as no-web constrained in the dossier.
• Keep the same extraction/ledger/QA flow, but clearly flag evidence limits in the article.

Step 1b — Evidence extraction (deep reading)

Create evidence_extract.md from the dossier using references/evidence_extraction_template.md.

For each major source, record:

• exact claim/data point extracted
• what it supports or challenges
• uncertainty/caveat

For Bible passages:

• Note structure (pericope boundaries), repeated words, contrasts, commands, warnings, comfort.
• Identify “center of gravity” (main burden) and what the text is not saying.

Step 1c — Research notes (synthesis scratchpad)

Create research_notes.md using references/research_notes_template.md:

• key definitions
• observations from sources/text
• competing interpretations
• load-bearing facts with source IDs
• open questions

Step 1d — Claim→evidence ledger (required)

Before drafting, build claim_evidence_ledger.md:

• Use references/claim_evidence_ledger_template.md.
• Every major claim must have at least one evidence anchor and source ID.
• Mark confidence and disconfirmation condition.

Step 2 — Argument outline (the real engine)

Write an outline that is easy to defend:

• Thesis (one sentence)
• Mainline statement (one sentence): what the article must prove for this user, and what is explicitly out of scope
• 3–6 supporting moves (each is a claim → evidence → implication)
• Guardrails: what to omit, what not to overclaim
• Transition logic: why move A leads to move B

Before drafting, run a quick outline challenge:

• For each move, state “if removed, what breaks?”
• Remove or merge moves that don’t break anything important.

Step 3 — Draft article.md

Use the skeleton in references/article_skeleton.md. Also read references/longform_depth_standard.md.

Rules:

• Prefer short paragraphs with explicit signposting (“因此/所以/因为/然而”).
• Separate observation vs application.
• Include explicit citations (source IDs and links) for load-bearing claims.
• Default to a substantial long-form article unless the user explicitly asked for a brief.
• Integrate evidence into the body itself. Do not outsource the reasoning to a references section or tell the reader to “see the sources” instead of explaining the point.
• A 3–6 move outline does not mean a short article. Each move may need multiple subsections or several paragraphs of evidence synthesis, tension, and implications.
• If evidence is mixed, present both sides and state your judgment criteria.

Step 4 — QA gates (logic + evidence)

Run both:

• references/logic_qa_checklist.md
• references/evidence_gates.md
• references/objectivity_checks.md
• references/mainline_coherence_gate.md
• references/final_article_quality_gate.md

If any hard gate fails, return to Step 1 (source sweep) or Step 1b (evidence extraction). In no-web constrained mode, return to Step 1a and narrow claims as needed.

Step 4b — Iterate until it “locks” (human-like drafting)

Deep research writing is normally iterative. Use references/iteration_protocol.md:

• revise thesis/outline if QA reveals gaps
• rewrite sections for clarity and scope
• add/remove evidence so every claim is supported
• trim or rewrite off-mainline sections (anything that does not advance the thesis)
• repeat QA + quality scoring until stop criteria are met

Step 4c — Revision log (recommended for real iteration)

Humans keep track of what changed; do the same to avoid thrash:

• Add a short “Revision notes” section to article.md, or keep a separate revision_log.md.
• Use references/revision_log_template.md.

Step 5 — Stop at the approved article

The normal end state of this skill is an approved article.md plus the supporting research package. Do not create a separate slide-extraction handoff artifact as part of the default deep-research workflow. If the user wants slides after the article is approved, hand the article directly to infographic-powerpoint-deck.

Output files (recommended)

• article.md
• source_dossier.md (required in internet-backed mode)
• evidence_extract.md (required in internet-backed mode)
• research_notes.md (required)
• claim_evidence_ledger.md (required)
• revision_log.md (optional, but helpful)

References

• Read references/article_skeleton.md when drafting article.md.
• Read references/web_research_protocol.md for the required search/deep-reading workflow.
• Read references/source_dossier_template.md to structure source collection.
• Read references/evidence_extraction_template.md for claim-level extraction.
• Read references/source_strategy.md to pick sources and record uncertainty.
• Read references/research_notes_template.md to structure research_notes.md.
• Read references/claim_evidence_ledger_template.md to prevent unsupported claims.
• Read references/logic_qa_checklist.md for the reasoning QA pass.
• Read references/evidence_gates.md for objective evidence sufficiency gates.
• Read references/objectivity_checks.md for neutrality and confidence calibration checks.
• Read references/mainline_coherence_gate.md to catch and remove sections that drift away from the core thesis/user intent.
• Read references/final_article_quality_gate.md for final draft quality scoring before handoff.
• Read references/longform_depth_standard.md so the final article defaults to paper-grade depth rather than a short summary.
• Read references/iteration_protocol.md for iterative improvement and stop criteria.
• Read references/revision_log_template.md to track iterations cleanly.