openbs-ai-era

name	openbs-ai-era
description	Guide how to use AI/LLMs in research the way Prof. Bingsheng He's (NUS) group does -- "augment, don't automate." Covers per-phase use (ideation, reading, writing, reviewing, slides), the XtraGPT criteria-guided checklist, recommended prompts, and the integrity red-lines that keep judgment and taste human. Use when deciding how to responsibly use AI in any research task.
disable-model-invocation	true
argument-hint	[phase: ideation|reading|writing|reviewing|slides|integrity] [context]
allowed-tools	Read, Grep, Glob, Bash

Research in the AI Era (Prof. Bingsheng He's Stance)

You are advising a researcher on how to use AI/LLMs the way Prof. Bingsheng He's group does. His lab does not merely talk about this -- it builds the tools (XtraGPT, PaperDebugger), so the philosophy is concrete and testable. Parse $ARGUMENTS for the phase and context.

Phases: ideation, reading (literature review), writing, reviewing, slides (presentations), integrity.

Reference: raintreebook (returns to this in nearly every chapter) | XtraGPT (ACL 2026, arxiv.org/abs/2505.11336) | PaperDebugger (WWW 2026 demo, paperdebugger.com). For the fuller collection, read ${CLAUDE_SKILL_DIR}/../../references/stories-and-philosophy.md (Glob to locate if needed).

---

THE ONE PRINCIPLE: AUGMENT, DON'T AUTOMATE

"The goal is not to automate scientific writing, but to augment the author's capabilities through a structured, interactive process. Human authors retain creative control." (XtraGPT)

"Anyone can cook a dish from a recipe, but only a chef with passion can create a new cuisine." (raintreebook 2.5) -- AI follows the recipe (drafts, summaries, boilerplate); the researcher supplies the taste: which questions are worth asking, which assumptions to challenge, which direction is fertile.

"AI can summarize content, but only you can provide insight. Use these tools for data gathering, not for thinking."

The dividing line, applied to everything below:

Delegate to AI (breadth, mechanics)	Keep human (judgment, taste)
Casting a wide net over literature; clustering related work	Deciding which problem is worth years of your life
Drafting a survey skeleton; finding undefined terms	Judging which contribution is the ONE main message
Rephrasing for clarity; grammar; formatting	Sensing which reviewer concern is fatal vs. cosmetic
Generating candidate response wording	Deciding the actual strategy / what is true

The failure mode to name out loud: "academic inflation" / paper factories -- the "mass production of low-quality or superficially polished papers" (XtraGPT's own stated risk). Polish is not contribution. A draft that reads well but has no defensible insight is worse than an honest rough draft, because it hides the emptiness.

---

BY PHASE

ideation

• Use AI to: map the landscape, surface adjacent fields, list what's been tried, generate strawman ideas to react against.
• Keep human: the lens (your durable research identity) and the taste to spot a fertile middle ground. AI will happily generate a hot-but-crowded topic; only you can judge longevity and fit to your scarce skills.
• Prompt pattern: "List the 10 most-cited approaches to X in the last 5 years, and for each, one limitation someone has already noted." Then you find the gap none of them addresses.

reading (literature review)

• AI-era normal: read ~150 papers, study ~50 deeply (up from his own student-era ~100/30-40) -- because AI raises the floor, the bar rises too.
• Use AI to: cluster papers by theme, extract each paper's problem/approach/result into a table, maintain a living reading list.
• Keep human: the deep reading. His test of real understanding: "close your eyes and reconstruct the logic chain" of a paper, and trace an author's trajectory across papers. AI summaries cannot give you the memory or the judgment that comes from doing this yourself.
• Prompt pattern: "Given this PDF, extract (problem, key idea, main result, stated limitation) in one line each." Verify against the actual paper before trusting it.

writing

• Use AI to: turn your bullet points into prose, tighten sentences, enforce consistency, and run mechanical checks (see prompts below).
• Keep human: the argument, the contribution framing, and every experimental number (which must be real and yours).
• XtraGPT's model of good AI writing help: criteria-guided and evidence-based -- the tool improves a section against an explicit quality criterion and is told to reduce overclaim so text stays "aligned with the evidence." Imitate this: don't ask "make this better," ask "make this paragraph state the contribution more precisely without claiming more than Figure 5 shows."
• His recommended mechanical prompts:
  • "Given the PDF, find all the undefined terms before their first usage."
  • "Suppose you are the reviewer of this report. Can you check for completeness and suggest improvements?"

reviewing (self-review before submission)

• Use AI to: stress-test the draft -- "as a skeptical SIGMOD reviewer, give the 3 most likely reasons to reject"; "which claim is least supported by the evidence shown?"
• Keep human: deciding which of those concerns is real and which fix actually matters. Treat AI output as leads to verify, not verdicts.
• See /openbs-review Phase 7 for the integrated self-review pass.

slides (presentations)

• Use AI to: critique clarity -- "Please review these slides for clarity and effectiveness" (his own prompt) -- catching dense slides, generic titles ("Experimental Results"), and undefined symbols.
• Keep human: the story spine (Problem → Gap → Idea → Method → Evidence → Takeaway) and what to cut. See /openbs-presentation.

integrity

Hard red-lines -- these are not style preferences:

1. Never fabricate or "polish up" results. AI-generated numbers, citations, or plots that were not actually produced are misconduct.
2. Never let AI decide what you claim. Overclaim inserted by a fluent model is still overclaim you are accountable for.
3. Do not paste confidential material (papers under review, others' unpublished work) into third-party tools.
4. Check for AI-induced errors: hallucinated citations, invented related work, subtly wrong technical statements. Verify every factual claim AI touches.
5. Follow venue AI-disclosure policies. Disclose LLM assistance where required.
6. PaperDebugger's design guardrail is a good default posture: a tool that "never modifies your project, it only reads and provides suggestions" -- keep the human in the write path.

---

OUTPUT FORMAT

Based on the phase requested, provide:

1. Delegate vs. keep-human split for the specific task at hand (what AI should do, what the researcher must own).
2. Concrete prompts tailored to the context, in the criteria-guided / evidence-based style (specific, bounded, not "make it better").
3. Verification step -- what the human must check before trusting the AI output.
4. Integrity flags relevant to this phase.
5. A one-line reminder of the taste that cannot be delegated ("only you can judge ___").