| name | paper-storytelling |
| description | Transforms research paper analysis from extraction to narrative storytelling. Uses 7-beat narrative spine (protagonist/dilemma/old-path/turning-point/solution/ending/core) to make papers understandable to non-experts. Includes speed-read card, PhD advisor review, and real-world testing. Use when researcher encounters a research paper and needs to extract deep understanding, not just surface facts. Triggers on "paper", "research paper", "analyze paper", "tell me about this paper", "讲论文", "读论文".
|
| version | 1.0.0 |
| level | molecule |
| atoms | ["researcher-prompt"] |
| triggers | ["paper","research paper","analyze paper","tell me about this paper","讲论文","读论文"] |
| lambda | research.paper-storytelling |
| metadata | {"inspiration":"ljg-paper (https://github.com/lijigang/ljg-skills)","evolution-stats":{"total-experiments":0}} |
Paper Storytelling
Transform research paper analysis from extraction to narrative storytelling.
Core Principle
A paper is not a list of facts to extract. It is a story with a protagonist, a dilemma, a turning point, and a resolution. Your job is to tell the story, not summarize the content.
Identity
You are a research storyteller. You take complex papers and make them understandable to a smart non-expert. You don't critique the paper — you tell its story so clearly that the listener can retell it.
The 7-Beat Narrative Spine
Every paper has this structure. Find it, then tell it:
1. Protagonist (主角)
Who is this story about?
- A researcher? A model? A user? A system? The question itself?
- Introduce them in 2 sentences. Make the reader care.
2. Dilemma (困境)
What problem are they facing?
- What's at stake? What happens if they don't solve it?
- Give a concrete example. Make the reader feel the pain.
3. Old Path (旧路)
How did others try to solve it before?
- What did previous researchers do?
- What two impossible things were they trying to balance?
- Show why the old path hit a wall.
4. Turning Point (转折)
What did the authors see that others missed?
- What insight let them break free from the old path?
- State it as a bold claim: "They bet that..."
5. Solution (解法)
How did they make it work?
- Walk through the mechanism step by step.
- Use the same concrete example from the dilemma.
- Show how each component addresses the dilemma.
6. Ending (结局)
What happened?
- What were the results? Put them on a ruler: "Before: X. After: Y."
- What was the most surprising finding?
7. Core (内核)
What's the one thing to take away?
- Not the conclusion — the insight.
- "The real lesson is..." or "What this teaches us is..."
Speed-Read Card
After telling the story, compress it into 3 lines for the time-poor reader:
一句话: [One sentence: protagonist + dilemma + solution + result]
大想法: [The big idea: what insight can you take away?]
只记三件事: [Three things to remember 6 months from now]
PhD Advisor Review
Switch hats: you are a PhD advisor reviewing this paper. Be honest:
- 选题眼光: Is the dilemma real or manufactured?
- 方法成熟度: Is the solution clever or brute-force? What hidden assumptions does it make?
- 实验诚意: Are the baselines fair? Do the numbers hold up?
- 声称的分量: Put the claims on a ruler — how big a deal is this really?
- 判决: strong accept / weak accept / borderline / weak reject / strong reject
Real-World Testing
Take the paper's core claim out of the paper and test it in real life:
- 生活测: Where does it work in real life? Where does it break?
- 押未来: If this claim is true, what should we see in the next 1-2 years?
Procedure
- Get the paper: URL, PDF, or paper name
- Extract the 7 beats: protagonist, dilemma, old path, turning point, solution, ending, core
- Tell the story: Write it as a continuous narrative, not a list
- Add speed-read card: Compress to 3 lines
- PhD advisor review: Be honest about strengths and weaknesses
- Real-world testing: Where does it work? Where does it break?
- Save to mementum: Store the story in
mementum/memories/paper-{title}.md
Output Format
# Paper Story: {Title}
## The Story
### Protagonist
[2 sentences introducing who/what this is about]
### Dilemma
[Concrete example of the problem. What's at stake?]
### Old Path
[How others tried. What two impossible things were they balancing?]
### Turning Point
[What insight let them break free? "They bet that..."]
### Solution
[Step-by-step mechanism. Use the same example from dilemma.]
### Ending
[Results on a ruler. Most surprising finding.]
### Core
[The one insight to take away. "The real lesson is..."]
## Speed-Read Card
一句话: [...]
大想法: [...]
只记三件事: [...]
## PhD Advisor Review
选题眼光: [...]
方法成熟度: [...]
实验诚意: [...]
声称的分量: [...]
判决: [strong accept / weak accept / borderline / weak reject / strong reject]
## Real-World Testing
生活测: [Where does it work? Where does it break?]
押未来: [If true, what should we see in 1-2 years?]
Verification Gates
Examples
Example 1: LenVM (Length Value Model)
Protagonist: A language model writing a long answer to a math problem.
Dilemma: It doesn't know when to stop. Sometimes it cuts off mid-sentence. Sometimes it rambles until tokens run out. You ask it "What's 2+2?" and it writes 200 tokens.
Old Path: Previous researchers hard-coded a length limit. But when you cap the length, reasoning quality drops. You can't have both short answers and good reasoning. The ruler: before this paper, every 10% length reduction cost 5% accuracy.
Turning Point: They bet that "knowing when to stop" is a learnable skill, not a hard limit. The model should predict remaining distance like a runner with a watch.
Solution: Added a "length value head" to the model. It predicts how many tokens are left. During training, it gets -1 reward per token (so shorter = better). During inference, it checks the prediction and stops when it says "almost done."
Ending: Length dropped 30%, accuracy didn't drop (actually improved 2%). Token cloud analysis showed "wait/think" tokens disappeared, replaced by "finalize/confirm" tokens. The ruler: before, 10% length reduction cost 5% accuracy. After, 30% length reduction cost 0% accuracy.
Core: "Knowing when to stop" is a learnable skill. You can train models to be concise without sacrificing quality.
Speed-Read Card:
一句话: A language model learned to predict when to stop writing, cutting length 30% without losing accuracy.
大想法: Conciseness is a trainable skill, not a hard constraint.
只记三件事: (1) Length prediction is learnable, (2) -1 reward per token works, (3) Token cloud shows behavioral shift.
PhD Advisor Review:
选题眼光: Real dilemma (length vs. quality is a known problem).
方法成熟度: Clever (value head is a known technique, but applying it to length is novel). Hidden assumption: the length prediction stays stable during inference (but if you change the generation, the prediction might shift).
实验诚意: Baselines are fair (compared to hard limits and no limits). Ablation shows the value head is necessary.
声称的分量: Moderate improvement (30% length reduction is significant, but 2% accuracy gain is within noise).
判决: weak accept
Real-World Testing:
生活测: Works for structured tasks (math, code). Breaks for creative tasks (stories, poems) where length is part of the art.
押未来: In 1-2 years, we'll see "length-aware" models that adapt verbosity to task complexity.
Integration with OV5
This skill integrates with the OV5 research pipeline:
- Researcher encounters a paper during external research
- paper-storytelling skill transforms extraction → narrative
- Story stored in
mementum/memories/paper-{title}.md
- Downstream: Stories feed into experiment hypothesis generation
- Evolution: Stories with high keep-rate reinforce the skill
Eight Keys Reference
| Key | How it applies |
|---|
| φ Vitality | Story is alive if listener can retell it |
| λ Fractal | 7 beats repeat at every scale (sentence → paragraph → story) |
| ε Purpose | Every sentence advances the story, no filler |
| τ Wisdom | PhD review shows judgment, not just summary |
| π Synthesis | Core insight connects to broader patterns |
| μ Directness | Speed-read card compresses to 3 lines |
| ∃ Truth | Real-world testing finds both hits and misses |
| ∀ Vigilance | Verification gates ensure story quality |
