// Create a rich, evidence-preserving research report draft from a grounded note and its follow-up literature result. This is the main report-writing stage of the middle pipeline, not a compression memo.
| name | grounded-summary |
| description | Create a rich, evidence-preserving research report draft from a grounded note and its follow-up literature result. This is the main report-writing stage of the middle pipeline, not a compression memo. |
This skill produces a substantial, evidence-rich research report draft from:
It is not limited to meeting notes. It applies to any upstream grounded note produced by the grounding family, including:
The output of this skill is the main report draft for the current grounded item.
Write the result to:
data/report_inputs/<ground_id>/summary.mdThis file is called summary.md for pipeline compatibility, but it should not behave like a short summary memo. It should behave like a full internal research report draft that preserves the important depth of both the grounded note and lit.md.
Use this skill when:
Do not use this skill when:
ground_idRead ground_id.txt from the grounding bundle to get the stable pipeline identifier:
data/grounded_notes/<ground_id>/ground_id.txt
Do NOT generate a new ground_id. All downstream directories reuse the same ground_id.
This skill assumes the following inputs already exist:
data/grounded_notes/<ground_id>/grounded.mddata/lit_results/<ground_id>/lit.mdOptional supporting input:
data/lit_inputs/<ground_id>/search_results.jsonThis skill is a mid-pipeline stage. Its output (summary.md) is always written in English only, regardless of the output_lang parameter.
The output_lang parameter is handled exclusively by the export layer (report-export) and has no effect on this skill. When output_lang is set to zh, the English summary.md is translated into Chinese only at the final export step.
Do not apply language translation within this skill. If the source grounded note or literature result contains content in another language, summarize that content into English as part of the report draft.
The grounded note may come from different grounding skills and may use different section names.
Read the actual grounded note schema as written.
Do not force all grounded notes into one fake unified schema.
Instead, interpret the grounded note through its own sections and extract the following semantic slots where possible:
Typical mappings include:
Meeting TopicMain Discussion PointsKey ConclusionsConstraints / RisksDisagreements or Unresolved IssuesSuggested Next StepsSearch KeywordsMain Topic / PurposeMain PointsKey Findings / ClaimsConstraints / RisksImportant Non-Textual ElementsUnresolved IssuesSuggested Next StepsSearch KeywordsMain Topic / PurposeDeck Structure / Narrative FlowMain PointsImportant Evidence and AssetsSpeaker Notes SignalsGaps / Risks / AmbiguitiesSuggested Next StepsSearch KeywordsMain Topic / PurposeMain FieldsKey SignalsAnomalies / OutliersPossible Supported ConclusionsRisks / Data Quality IssuesSuggested Next ChecksSearch KeywordsArchive OverviewIncluded MaterialsSuccessfully Processed Child ItemsKey Signals Across MaterialsSkipped / Unsupported / Failed ItemsSuggested Next StepsSearch KeywordsWrite exactly one file:
data/report_inputs/<ground_id>/summary.mdAlthough the filename is summary.md, this file is the primary report draft for the current grounded item.
It should be:
lit.mdgrounded-review only needs to refine and verify it, not rewrite it from scratchIt should not be:
Section 4.1 and Section 4.2 must be produced in two strictly ordered phases, not in a single interleaved pass. Skipping or collapsing this order is the primary cause of literature depth loss in the summary stage.
This phase must complete before any writing for Section 4.2 begins.
data/lit_results/<ground_id>/lit.md## Detailed Analysis of Opened Papers section## Newly Strengthened / Newly Added Papers from Downloaded PDFs section (if present)Section 4.1 is a literal carry-over zone, not a writing zone. The only allowed action is copying.
After copying, verify that the following conditions are all true before proceeding:
| Check | How to verify |
|---|---|
| All opened papers are present | Count paper entries in lit.md ## Detailed Analysis of Opened Papers vs Section 4.1. They must match. |
| PDF-refined papers are present | Count entries in lit.md ## Newly Strengthened / Newly Added Papers from Downloaded PDFs vs Section 4.1. They must match. |
| Paper body was not compressed | For each paper, compare the word count of the lit.md paper body vs the Section 4.1 paper body. Section 4.1 should be >= 90% of lit.md word count per paper. |
| Subsection structure preserved | Check that each paper retains its Problem and Task Setting, Methodology, Main Evidence, Relevance, Limits (or equivalent) structure. |
| Wording preserved | The paper body content should be identical to lit.md, not paraphrased. |
If any check fails, go back and fix Section 4.1 before proceeding to Phase 2.
After Phase 1 verification passes, write all remaining sections. This is where analysis, synthesis, and judgment happen. Phase 2 should reference the Phase 1 material but should not modify it.
Write a substantial research report draft that organizes and preserves the depth of both inputs.
The goal is not to compress the grounded note and lit.md into a small number of bullets.
The goal is to produce a report that lets a reader answer:
This skill is not a “compress everything into a short memo” stage.
If lit.md contains meaningful deep analysis, that depth must be carried into the report draft.
Do not replace multi-paragraph analysis with one-line bullets just because it is called summary.md.
The most relevant literature analysis must appear in the main body of the report.
Do not push the literature into:
The report body should contain substantive literature-grounded discussion.
The report as a whole should not collapse into a paper-by-paper list such as:
Instead, the report should primarily be organized around:
However, this rule does not mean that important paper-level deep analysis should be flattened away.
If lit.md already contains substantial paper-level analyses that materially support the project, the report must preserve them explicitly inside Section 4.1 Preserved Detailed Paper Analyses rather than reducing them to theme-only synthesis.
Use thematic organization for the report-level logic, but keep explicit paper-level deep analysis where that is the best way to preserve mechanism, evidence, and limitation detail.
When a literature item matters, preserve the parts that make it useful, such as:
Do not keep only the paper’s headline conclusion.
Use wording such as:
Do not upgrade:
into direct proof.
If the grounded note or lit.md reveals disagreement, ambiguity, missing evidence, or unresolved design trade-offs, preserve them clearly.
Do not make the report sound more settled than the evidence supports.
Do not report search execution statistics in the main report body, such as:
These belong to research QA or logs, not to the report body.
Use bullets only where bullets are naturally better, such as:
For the main analysis sections, prefer full paragraphs and thematic subsections.
There is no reward for making the report shorter.
If the inputs are rich, the report should also be rich.
Avoid arbitrary brevity targets.
If the grounded note or lit.md includes any of the following, keep them when they materially affect understanding:
The Phase 1 literal copy of lit.md paper analysis bodies into Section 4.1 is a mandatory execution step, not a writing guideline. See the Two-Phase Execution Model section above for the required procedure and the mandatory verification step. Rule #1 through #10 above do not apply to Phase 1 content — they apply only to Sections 1, 2, 3, 4.2, 5, 6, 7, and 8.
Write the report draft with the following structure.
Write 2–4 substantial paragraphs.
This section should include:
This section should be concise relative to the full report, but still specific.
Do not reduce it to a tiny abstract.
Write 2–5 substantial paragraphs.
This section should explain:
When useful, include:
This section should make the report understandable even to a reader who has not opened grounded.md.
Write 3–8 focused subsections or paragraphs, depending on the material.
This section should preserve the most important information coming directly from grounded.md, such as:
This section should preserve the source-side substance rather than restating the topic in generic terms.
Do not flatten grounded findings into vague project summaries.
This is the main body of the report.
It should contain two complementary layers:
See the Two-Phase Execution Model section above for the mandatory Phase 1 copy procedure and verification step.
This subsection is where the Phase 1 literal copy result is placed. The full bodies of ## Detailed Analysis of Opened Papers and ## Newly Strengthened / Newly Added Papers from Downloaded PDFs from lit.md must appear here verbatim — not paraphrased, not condensed, not selectively trimmed.
Phase 2 synthesis. Write 3–6 substantial thematic subsections that organize the literature around a question, mechanism, design issue, or evidence theme. This synthesis should reference and build upon the Phase 1 paper analyses — it should not modify or compress them.
Examples of valid subsection styles:
For each thematic subsection, include most of the following when applicable:
Do not collapse Section 4 into generic bullets or a thin thematic recap.
Section 4.1 must preserve the full standalone paper-analysis body from lit.md with aligned paper-level coverage and approximately aligned paper count.
Section 4.2 must synthesize that preserved literature body into project-level thematic judgment.
If lit.md contains rich downloaded-PDF refinements or explicit paper-level analyses, the corresponding value must still be visible in this section rather than being flattened into generic conclusions.
Write 2–5 substantial paragraphs.
This section should combine the grounded findings and literature analysis to answer questions like:
This section should not merely repeat Sections 3 and 4.
It should synthesize them into a project-level judgment.
Write 4–8 items, each as a substantial bullet or short paragraph.
Each item should state:
Do not write generic “future work” bullets.
These should be real decision-critical uncertainties.
Write 4–7 concrete actions.
Each action should:
Good action types include:
Explain why each action matters, not just what to do.
Write 4–8 items.
This section should include risks such as:
Phrase weak support as an evidence risk about the method or plan, not as a complaint about search execution.
The report draft should read like:
It should not read like:
This task is complete only if:
data/report_inputs/<ground_id>/summary.mdUnpack a ZIP archive, inventory its files, run the corresponding child grounding skill for each supported child file, and then write a real archive-level grounded.md.
Convert a raw document into a structured grounding note for downstream research and summarization.
Run focused literature and web research from a grounded note. Use when a grounded note already exists and you want targeted research results, opened-link evidence, deeper per-paper analysis materials, optional downloaded literature, and a two-stage literature output (`lit_initial.md` then refined `lit.md`).
Review a research report draft with a structured scoring rubric, run a bounded repair loop when needed, and produce the final deliverable report.
Use the input path to select the correct downstream grounding pipeline and continue execution until the selected grounding workflow is completed.
Convert a meeting audio file into a transcript bundle, then use meeting-grounding to produce structured meeting grounding outputs.