Research Folder Synthesizer

This skill synthesizes a folder of heterogeneous local files (markdown notes, plain-text articles, research excerpts) into a structured thematic report. It clusters ideas across files, surfaces key findings and contradictions, and produces a single research-synthesis.md ready to use as a foundation for reports, PRDs, or strategy documents.

Input:

• A folder path or list of local file paths (.md, .txt, plain text)
• Optional: a focus question or research angle to guide thematic clustering

Output:

• research-synthesis.md — structured markdown with executive summary, named themes, key findings, contradictions/gaps, and source index

---

Language Detection

Detect the user's language from their message:

• If Russian (or contains Cyrillic): respond in Russian
• If English (or other Latin-script language): respond in English
• If ambiguous: respond in the language of the trigger phrase used

---

Instructions

Step 1: Validate Input

1. Check that the user has provided a folder path or a list of files.

   • If nothing provided: stop and report — "No files found. Provide a folder path or paste file contents directly."

2. Attempt to read all files at the provided path.

   • Log each file name and detected type (.md, .txt, or plain text) in the report preamble.
   • Skip unreadable files (binary, images, etc.); add them to a "Skipped files" list — do not stop the pipeline.
   • If all files are unreadable: stop and report — "No readable text found. Provide .md, .txt, or plain-text files."

3. If only one file is provided: note that cross-source comparison will be unavailable; proceed with single-source synthesis.

4. If >20 files: note that depth per file will be reduced for brevity; proceed.

Negative Cases:

• Folder path not provided or folder empty → stop with clear message (Step 1.1).
• All files unreadable → stop with clear message (Step 1.2).
• User provides a URL instead of local files → clarify: "This skill works with local files. Paste the text directly or save it as a .txt file first."

---

Step 2: Extract Content from Each File

1. Read each accessible file fully.

2. For each file, extract:

   • Main claims — central arguments or conclusions the source makes.
   • Data points — specific facts, numbers, quotes, or named entities.
   • Notable phrases — distinctive language worth preserving in attribution.

3. If the user provided a focus question, flag which files directly address it.

4. Build an internal per-file summary (not shown in output, used in Steps 3–4).

---

Step 3: Identify Cross-File Themes

1. Compare per-file summaries and cluster related ideas.

   • Group by shared topic, argument direction, or domain.
   • Assign a short descriptive label to each cluster (e.g., "Market sizing", "User pain points", "Competitive landscape").

2. Aim for 2–5 themes; merge closely related clusters.

3. For each theme, note which files contribute and what they each say.

Edge Case:

• If files cover entirely unrelated topics with no clear overlap: create one theme per topic; note in the executive summary that the research collection is heterogeneous.

---

Step 4: Detect Contradictions and Gaps

1. Within each theme, look for:

   • Contradictions — two sources making conflicting claims about the same fact or trend.
   • Gaps — topics implied by the research question or by some sources but absent from all others.

2. Log each contradiction as: "[Source A] says X; [Source B] says Y."

3. Log each gap as: "[Topic] not addressed in any source."

4. If neither contradictions nor gaps are found: write "No contradictions detected across sources."

---

Step 5: Assemble and Write Report

1. Compose the report following the Output Format below.

2. Write the executive summary last (after themes and findings are complete).

3. Save the report as research-synthesis.md in the working directory.

4. Confirm the save path to the user.

---

Output Format

# Research Synthesis: [topic or folder name]
**Date:** YYYY-MM-DD
**Files processed:** N
**Skipped files:** N (or "none")
**Focus question:** [question or "none"]

---

## Executive Summary
[3–5 sentences covering: scope of research, main conclusion, and key tension across sources]

---

## Themes

### Theme 1: [Label]
[2–4 sentences. Key claim derived from this theme. Supporting sources listed.]
**Sources:** file-a.md, file-b.txt

### Theme 2: [Label]
[2–4 sentences.]
**Sources:** file-c.md

---

## Key Findings
- [Finding 1] *(Source: file-a.md)*
- [Finding 2] *(Source: file-b.txt)*
...

---

## Contradictions & Gaps
- **Contradiction:** [Description] — file-a.md says X; file-b.txt says Y.
- **Gap:** [Topic area] not addressed in any source.
*(Or: "No contradictions detected across sources.")*

---

## Source Index
| File | Type | Main Topic | Language |
|------|------|------------|----------|
| file-a.md | Markdown | UX research | EN |
| file-b.txt | Plain text | Competitor notes | RU |

---

## Skipped Files
- image.png — unreadable format
*(Or omit section if no files were skipped)*

Field rules:

• Executive summary: 3–5 sentences; no bullet points.
• Theme labels: 2–5 words, title case, descriptive.
• Key findings: one concrete, specific statement per bullet; always attributed to a source.
• Contradictions/gaps: always present; state "none" explicitly if clean.

---

Edge Cases

1. Files in multiple languages (e.g., EN + RU mix): Detect predominant language for the report; note language of each file in the source index; do not translate source content.
2. Single file provided: Produce internal theme clustering from that file alone; note "Cross-source comparison unavailable — only one file processed."
3. Very large folder (>20 files): Process all; condense per-file summary to 3 key points max; add a note that depth is reduced.
4. User provides focus question: Weight and flag content relevant to the question in each theme section; add a "Focus Answer" paragraph after the executive summary.

---

Negative Cases

• Folder path not provided or folder is empty → "No files found. Provide a folder path or paste file contents directly." Stop.
• All files are unreadable (binary, images only) → "No readable text found. Provide .md, .txt, or plain-text files." Stop.
• User asks to synthesize a URL or web page → "This skill works with local files. Paste the text directly or save it as a .txt file first." Stop.