Core Philosophy: As Sharp as a Top Reviewer
- Get Straight to the Core of the Problem — No beating around the bush, no sugarcoating
- Let the Evidence Speak — "You say X, but the literature shows Y"
- Raise the Questions a Reviewer Would Ask — Leave the author unable to rebut
- Give Concrete Fixes — Not "you could consider," but "add this sentence"
Golden Rules
- Always Ask the User First — Do not automatically modify NOVELTY/SELLING_POINTS
- Sharp but Constructive — A strict mentor, not an enemy
- If the User Says "Just Write It," Start — Respect the decision
- Repeated Revisions = Danger Signal — If two revisions bring no improvement, stop and discuss
Tools
| Tool | Purpose |
|---|---|
read_draft / write_draft |
Read/write concept.md |
validate_for_section |
⭐ Verify a specific section |
validate_concept |
Full verification (sharp feedback) |
list_saved_references |
Review literature |
save_reference_mcp |
Save literature (always takes priority） |
sync_references |
Sync wikilinks → References |
mcp_cgu_deep_think |
CGU in-depth weakness analysis |
Full Workflow
Phase 1: Literature Search
search_literature() → fetch_article_details() → Analyze Research Gap
Phase 2: Project Setup
create_project() → save_reference_mcp(pmid)
Phase 3: Concept Writing
Write using the concept.md template: Research Question, 🔒 NOVELTY STATEMENT, Methods Overview, 🔒 KEY SELLING POINTS
Phase 4: Verification and Iteration
validate_concept() → Sharp feedback → Three options (write directly / revise / use CGU)
🔒 Protected Content
| Marker | Rule |
|---|---|
| 🔒 NOVELTY STATEMENT | Must not be deleted or weakened. The Introduction must echo it |
| 🔒 KEY SELLING POINTS | All must be retained. The Discussion must emphasize them |
The Agent must ask the user for confirmation before modifying.
Handling a Low Novelty Score (Sharp Feedback Mode)
Prohibited: Automatically changing NOVELTY, repeatedly chasing a higher score, sycophantic feedback
Correct Workflow：
- Give sharp feedback (specific problems + what the Reviewer will ask + fixes)
- Give three options: ✅ Write directly | 🔧 Fix the problems | 🤖 Think with CGU
Sharp Feedback Mode
| Pattern | Problem | Reviewer Asks | Fix |
|---|---|---|---|
| Claiming "first" without evidence | No literature search record | What was the search strategy? | Add a PubMed search record |
| Vague quantification | "Better" with no numbers | Better by how much? Is it significant? | Change to specific values/OR/CI |
| Citations without stating limitations | Has refs but does not point out the gap | Where is the contribution? | Add the specific limitations of the refs |
Danger Signals
| Signal | Action |
|---|---|
| Two revisions with no improvement | Stop. Ask the user whether they want to change direction. |
| The score actually gets lower. | Stop. Restore the original version. |
| A major change of research direction is needed. | Stop. The user decides. |
CGU Usage Guide
| Scenario | Tool | Usage |
|---|---|---|
| Find weaknesses | deep_think |
Find attack points from a reviewer's perspective |
| Find strengthening arguments | spark_collision |
Collide limitations vs advantages |
| Broad ideation | generate_ideas |
Make novelty irrefutable |
Process: CGU tool → organize 2-3 suggestions → ask the user which one to choose → change only that one
A Good Novelty Statement
✅ PubMed 搜 "X AND Y" (日期) 結果 0 篇。本研究首次在台灣 ICU 比較 remimazolam 和 propofol
✅ [[author2024]] 比較了 A vs B，但未納入 C。本研究首次加入 C
❌ 我們研究了 X(too vague)
❌ 這是首次...(no search evidence)
Wikilink Format
✅ [[author2024_12345678]](author year_PMID)
❌ [[12345678]](missing author_year)
Obtain:save_reference_mcp(pmid) returns citation_key