Goal: Generate diverse, creative ideas
AI Role: Creative brainstorming partner
Key Tactic: Ask for multiple, varied, “weird” ideas
Risk if Skipped: Right answers to uninspired questions
Goal: Achieve precise, correct result
AI Role: Meticulous step-by-step calculator
Key Tactic: Demand step-by-step work, definitions, unit checks
Risk if Skipped: Creative ideas without rigorous support
Goal: Identify flaws and hidden assumptions
AI Role: Skeptical peer reviewer
Key Tactic: Ask “What’s wrong?” and “What if…?”
Risk if Skipped: Critical errors go unnoticed
AI tools like ChatGPT solve textbook problems well but struggle with creative research. However, if you give them worked examples they can become more reliable and predictable. If you do not traine them, it is likely the stryggle with creative endeavours and become sloppy.
In general, I have found that we can AI differently for different research phases. One phase is exploring broadly, as if you had a canvas for brainstorming. The other phase it is like having a scalpel for precise work, that you can use for sanity checks based on your expertise and sources of knowledge that are reliable.
Canvas mode, adopting now this terminology, can increase idea variety, making it better for discovery. However, it might be risky when you need precision. The scalpel mode reduces idea variability through step-by-step reasoning, giving you reliability but possibly locking you into familiar approaches, and knowledge. Finally, sanity checks help you probe assumptions that both other modes miss. This is where you exercise your critical thinking and skepticism.
Generate 8 different explanations for why heritability estimates vary between populations.
For each one:
- State the core mechanism in one sentence
- Give two testable predictions
- Describe the single most decisive observation that would falsify it
- Rate how likely it is (low/medium/high) and why
Include at least 3 weird but biologically plausible ideas.How would a control engineer think about gene regulation? An economist about
population dynamics? A physicist about protein folding? For each perspective,
what tools or principles could we steal for biology?Everyone assumes linkage disequilibrium decays exponentially. What if we
dropped that assumption? What alternatives are there, and what would each
predict about patterns in real populations?Act as [famous scientist] with their theoretical perspective, but with access
to modern knowledge. From that viewpoint, generate 5 unconventional hypotheses
for [your problem]. Focus on concepts that [scientist] would emphasise.
Examples:
- "Act as Sewall Wright in the 1940s, but with modern genomics knowledge..."
- "Take the perspective of Lynn Margulis and her focus on symbiosis..."
- "Channel Mary-Claire King's approach to genetics and human disease..."Quantitative Genetics Canvas Prompt
List 10 reasons why a selection experiment might give weird results that don't
match the breeder's equation. Include some exotic possibilities like epigenetic
effects, maternal environments, or developmental noise. For each one, what would
it predict about the pattern of response over time?Network Biology Canvas Prompt
Why might feed-forward loops evolve in gene regulatory networks? Think beyond
the standard 'noise filtering' explanation. Consider development, evolvability,
robustness to mutation, metabolic costs, etc. For each idea, give me a prediction
and a way to test it.Now you can pick one promising idea from your brainstorming and work through it carefully. This is where you want AI to be systematic and show all its work.
Bounds to Always Check:
Scalpel Template: Step-by-Step
Solve this step-by-step:
1) Define all variables and state your assumptions clearly
2) Write down the relevant equations or logical steps
3) State the key limitations or conditions under which this model fails
4) Plug in numbers and work through the calculation
5) Check your units and whether the result makes biological sense
6) Give me your final answer in one clear sentenceQ: A plant population has mean height 100 cm, total variance 25 cm², and genetic variance 10 cm². After selecting the tallest 20%, the mean is 105 cm. What’s the predicted response?
Step-by-step solution:
New population mean = \(100 + 2 = 102\) cm
Always demand:
If you believe you got a reasonable and clean answer. it is time to break it. This catches errors and reveals hidden assumptions that th eother two modes might have missed.
Reality Check Template: Attack Questions
I just calculated [your result]. What's the biggest flaw in this reasoning?
What assumptions am I making that might be wrong? In what situations would
this completely fail?Reality Check Template: Constraint Checking
Check if my answer violates any biological constraints. For instance:
- Is heritability between 0 and 1?
- Are population sizes positive?
- Do probabilities sum to 1?
- Are growth rates reasonable?Reality Check Template: Alternative Explanations
What's the strongest alternative explanation for this result? How would I design
an experiment to distinguish between my explanation and the alternative?Here are ready-to-use templates you can customise for your research:
Generate [number] distinct explanations for [your phenomenon].
For each one:
- State the core mechanism in one sentence
- Give two testable predictions
- Describe what evidence would prove it wrong
- Rate how likely it is (low/medium/high) and why
Include at least 3 weird but biologically plausible ideas.How would researchers from these 5 fields approach [your problem]: control theory,
economics, physics, computer science, and ecology? For each perspective, what tools
or principles could we borrow? What new predictions would that generate?Everyone in my field assumes [standard assumption]. What if we dropped that?
What alternatives exist, and what would each predict about [your system]?Solve this step-by-step:
1) Define all variables and state your assumptions clearly
2) Write down the relevant equations or logical steps
3) Plug in numbers and work through the calculation
4) Check your units and whether the result makes biological sense
5) Give me your final answer in one clear sentenceI'm using [method] to analyse [type of data]. Walk me through:
1) What assumptions this method makes
2) How to check if my data meets these assumptions
3) What the results mean biologically
4) What could go wrong and how I'd knowI just calculated [your result]. What's the biggest flaw in this reasoning?
What assumptions am I making that might be wrong? In what situations would
this completely fail?My data shows [pattern]. I think it's due to [your explanation]. What's the
strongest alternative explanation? How would I design an experiment to
distinguish between these possibilities?Canvas Phase: “Why might selection response be weaker than predicted? Generate 6 mechanisms including some unusual ones.”
Results might include: linkage drag, maternal effects, developmental constraints, inbreeding depression, environmental correlation, epigenetic inheritance.
Scalpel Phase: Pick “linkage drag” and model how linked deleterious alleles reduce response.
Reality Check: “What would prove linkage drag wrong? How could I test this experimentally?”
Falsifier: Response increases after breaking up linkage through recombination.
Canvas Phase:
“Propose 8 evolutionary explanations for why incoherent feed-forward
loops are common in regulatory networks.”
Possible mechanisms: - Gene duplication-divergence:
One copy retains activation, other evolves repression -
Falsifier: No paralogue age differences between
activating and repressing arms - Cis-regulatory
co-option: Recruitment from stress response circuits
- Falsifier: No shared regulatory motifs with stress
response genes - Post-transcriptional delay:
MicroRNA-mediated repression creates temporal lag -
Falsifier: Loss of transient dynamics when miRNA
biogenesis is disrupted
Scalpel Phase: Model population genetics of how gene duplication could create this pattern.
Reality Check: “What phylogenetic patterns would I expect if duplication-divergence is correct? What would falsify it?”
Canvas Phase: “I’m studying why genetic diversity varies so much between island populations. Generate 8 different mechanisms.”
Results: founder effects, ongoing migration, population bottlenecks, selection differences, mating system variation, mutation rate differences, genetic drift, population structure.
Scalpel Phase: Calculate expected \(F_{ST}\) values under the island model: \(F_{ST} = \frac{1}{1 + 4N_em}\)
With \(N_e = 1000\), \(m = 0.01\): \(F_{ST} = \frac{1}{1 + 40} = 0.024\)
Units check: Dimensionless âś“; Bounds check: \(0 \leq F_{ST} \leq 1\) âś“
Reality Check: “What would prove migration wrong? Design controls for drift effects.”
Falsifier: If \(F_{ST}\) values don’t correlate with geographic distance, ongoing migration model fails.
Good lab practice means knowing when to clean your workspace. AI conversations degrade over time: Answers get vaguer, errors compound, and the AI gets stuck repeating itself.
Copy key results into a notes file before starting fresh. Build on previous work without carrying conversational baggage.
AI can be powerful for research, but you need strategy. Don’t expect it to replace your scientific thinking. Use it to enhance your natural research process.
Canvas phase helps you think more broadly than you would alone.
Scalpel phase helps you work through complex problems without arithmetic errors.
Reality check helps you catch mistakes and strengthen arguments.
The key insight: different research phases need different AI interactions. Match your prompting strategy to your research needs.
Most importantly: you’re still the scientist. AI is a sophisticated tool, like a microscope or statistical software. It amplifies your capabilities but doesn’t replace your judgement, creativity, or domain expertise.