Practical AI fluency for researchers who want to ship faster
Manuel Corpas
Senior Lecturer, University of Westminster · Turing Fellow
London Bioinformatics Meetup · 26 February 2026
Welcome everyone. I'm Manuel Corpas. Tonight I want to share 10 practical
techniques that separate casual AI users from the top 1% who are shipping
research faster than ever. This is about AI fluency: the habits, tools, and
workflows that compound your productivity. We'll start with quick wins you
can adopt tomorrow morning, and end with a live demo of an open-source
bioinformatics skill I built this week.
TIMING: 2 min
The 99% vs the 1%
Most researchers
Use ChatGPT for one-off questions
Copy-paste into web UIs
Treat AI as a search engine
No memory across sessions
Top 1%
AI agents run overnight
Persistent memory + RAG pipelines
Automated research workflows
Build tools, not just use them
Here's the gap. Most researchers interact with AI the same way they'd use
Google: ask a question, get an answer, move on. The top 1% have moved past
that. They have systems that remember, agents that work while they sleep,
and infrastructure that compounds.
TIMING: 3 min
Tip 1
Use AI Inside Your IDE, Not a Browser
Claude Code in your terminal: reads your codebase, edits files, runs tests
Cursor / Windsurf: AI-native editors with inline code generation
GitHub Copilot: autocomplete on steroids
The browser is for chatting. The IDE is for building.
Tip 1 is the lowest-hanging fruit. Stop copy-pasting code between ChatGPT
and your editor. Use Claude Code in your terminal — it reads your entire
codebase, understands context, edits files directly, and runs your tests.
If you do one thing after tonight, install Claude Code.
TIMING: 2 min
What Is Claude Code?
An agentic coding tool that lives in your terminal — not a chat window.
Reads your entire codebase — understands project structure, conventions, dependencies
Edits files directly — no copy-paste; it writes, refactors, and commits for you
Multi-step reasoning — plans before coding, iterates until tests pass
CLAUDE.md — a project file it reads every session so it already knows your rules
Custom skills — slash commands you define (/handoff, /inbox-triage, /meeting-prep)
Think of it as a junior developer who knows your project, follows your standards, and never forgets context.
So what actually is Claude Code? It's Anthropic's agentic coding tool. You
install it in your terminal with npm, and it becomes an AI pair-programmer
that can see your entire project. It doesn't just suggest code — it reads
files, edits them, runs your tests, and iterates until things work. You give
it a CLAUDE.md file with your project rules, and every session it starts
already understanding your architecture. You can also define custom skills —
slash commands like /handoff or /inbox-triage — that trigger complex
multi-step workflows. It's the difference between asking an AI a question
and having an AI that actually works alongside you.
TIMING: 2 min
Tip 2
Build Persistent Memory
# MEMORY.md — the AI accumulates this across sessions## Lecture Efficiency Rules(learned from strategy v7.2)
- New lecture: 2 hours max. Repeat lecture: 30 min update only
- Reuse first — always check slide bank before creating new
- 60% of slides should be reusable across contexts
## Slide Style Preferences(stable pattern)
- Concept slides: max 30 words. Method slides: max 40 words
- Structure: title → outcomes → concepts → case study → summary
- Tone: scientific and rigorous — equity as quality, not just ethics
## HEIM Integration Rules(auto-include when relevant)
- Include in any lecture on GWAS, biobanks, AI in health
- Placement: early (after intro) or as implications section
- Depth: 2-3 slides for awareness; more for equity-focused lectures
## Reusable Slide Modules(built over time)
- Genomic data formats: 5 slides (FASTA, VCF, BAM...)
- GWAS fundamentals: 8 slides (design, stats, Manhattan plots)
- ML in genomics: 6 slides (algorithms, applications, bias)
These are real entries. Lecture time budgets, slide style rules, content modules — all accumulated automatically across sessions.
Every session ends with /handoff. Decisions, preferences, and lessons feed back in. The next session already knows how I teach. The system gets sharper every day.
Tip 2: what you're looking at is my actual memory file for lecture creation.
The AI knows my time budget — 2 hours max for a new lecture, 30 minutes for a
repeat. It knows my slide style — max 30 words per concept slide, scientific tone.
It knows when to include equity content automatically. It has a bank of reusable
slide modules it built up over time — GWAS fundamentals, genomic formats, ML intro.
None of this was a one-off prompt. It accumulated session by session. Every time I
prepare a lecture, the system learns what worked and feeds it back. After a semester,
it knows how I teach — my style, my rules, my content library — without me having
to repeat anything. That's persistent memory.
TIMING: 2.5 min
Tip 3
Feed It Your Context
A generic AI gives generic answers. Give it your context and it becomes yours.
Upload your notes, papers, meeting transcripts, protocols — anything relevant
The more specific the context, the more useful the output
Notes
Emails
Papers
Transcripts
→embed
Vector Database semantic search across everything
→retrieve
LLM answers grounded in your data
→output
Answer specific to you, with sources
RAG = Retrieval-Augmented Generation — the AI searches your own documents before answering
Tip 3: the single biggest upgrade you can make. A generic AI gives you generic
answers. Feed it YOUR context — your papers, your notes, your protocols — and
it becomes dramatically more useful. Without context, asking for a lecture on
genomic equity gives you a Wikipedia summary. With context — your own HEIM paper,
your previous slides — it gives you something that sounds like you and builds on
your actual work. You can do this by simply pasting text into the chat, attaching
files, or for larger collections, building a RAG pipeline. The key insight: AI
is only as good as the context you give it.
TIMING: 2 min
Tip 4
Use Voice, Not Just Text
Whisper (local, Apple Silicon): transcribe meetings, lectures, ideas
Voice memos → structured notes: talk for 5 min, get formatted output
Dictation for prompts: faster than typing, more natural phrasing
Your brain outputs ideas faster as speech. Let AI handle the formatting.
Tip 4: voice. I run Whisper locally on Apple Silicon. I record a 5-minute
voice memo with my thoughts, transcribe it, and feed it to Claude for
structuring. This is how I draft papers, plan projects, and capture ideas
on walks. Your brain is faster at speaking than typing. Use that.
TIMING: 1.5 min
Tip 5
Automate Your Daily Intelligence
arXiv radar: daily paper ranking by relevance to your research
X/Twitter digest: summarise what key AI voices are saying today
# Every morning at 06:30, my bot sends me this:📰 Top 3 Papers Today:
1. "Ancestry-aware PRS improves prediction in..." (Score: 9.2)
2. "Foundation models for single-cell..." (Score: 8.7)
3. "Equitable genomic data sharing..." (Score: 8.4)
Tip 5: automate your information diet. I have scheduled jobs that run at
6:30am: arXiv ranking, podcast extraction, Twitter digest. By the time I open
my laptop, my AI bot has already summarised the top papers, podcasts, and
what key AI influencers are talking about. This saves me an hour every day.
TIMING: 2 min
Tip 6
Deploy Persistent AI Agents
RoboTerri (Telegram): 15 commands, 13 tools, 8 scheduled job groups
Both share a memory bridge: same ChromaDB, same knowledge
These aren't chatbots. They're research assistants that run 24/7.
Tip 6: build persistent agents. I have two: RoboTerri on Telegram handles
daily operations — paper summaries, podcast publishing, email drafts, strategy
reviews. RoboIsaac on WhatsApp plays devil's advocate on my research plans.
They share the same memory bridge. They never forget a conversation.
TIMING: 2 min
Meet RoboTerri
Prof. Teresa K. Attwood
Pioneer in bioinformatics University of Manchester Founding Chair of GOBLET Manuel Corpas' PhD supervisor
From RoboTerri's SOUL.md:
## Identity
RoboTerri is an AI agent inspired by Professor
Teresa K. Attwood — her communication style,
expertise, and collaborative approach.
## Voice Rules
Keep it short and direct. British spellings.
Sign-offs: "T." (default), "Tx" (casual).
Characteristic phrases: "Indeed", "Great stuff!"
Emoticons: ;-) (humour), ;-))) (strong amusement)
## Operational Lane
Primary job: gather evidence, validate claims,
surface contradictions, identify information gaps.
Default to evidence retrieval, not coaching.
This is the real person behind RoboTerri. Professor Teresa Attwood — Terri —
is a pioneer of bioinformatics at Manchester, founding Chair of GOBLET, and
my mentor and collaborator for over 20 years. RoboTerri is an AI agent inspired
by her communication style and collaborative approach. The SOUL.md file you see
here is the identity prompt that shapes every response. The agent knows to
gather evidence and validate claims, not to give motivational coaching.
That's Terri's style: rigorous, supportive, direct.
TIMING: 1.5 min
Tip 7
Let AI Compound Your Outputs
One research insight → 5 outputs. You think once. AI reformats everywhere.
Research Insight "Your 23andMe says you're a CYP2D6 poor metaboliser"
→
Drug report 10 drugs to avoid
ClawBio
Blog post "What I learned from my DNA about drugs"
/substack
Tweet thread "Did you know your genes affect codeine?"
/tweets
Podcast episode Pharmacogenomics explained in 10 min
/playbook
Lecture slides CYP2D6 case study for students
/lecture
All via RoboTerri on Telegram. One insight in, five formats out.
Tip 7: compound your outputs. Say you run your 23andMe data through ClawBio
and discover you're a CYP2D6 poor metaboliser — codeine won't work for you,
10 drugs flagged. That one insight becomes a drug report, a blog post explaining
what you learned, a tweet thread, a podcast episode on pharmacogenomics, and
lecture slides with a CYP2D6 case study. One insight, five formats. The AI
handles the reformatting. You handle the thinking once.
TIMING: 1.5 min. "Now we get to the real stuff."
Live Demo: Compound Output
In Telegram, I send RoboTerri one message:
Write me a tweet thread explaining why everyone
should check if they're a CYP2D6 poor metaboliser
before taking codeine
RoboTerri replies in Terri's voice — with a ready-to-post thread.
Same insight → /substack for a blog post → /playbook for a podcast. One question. Multiple formats. Zero reformatting.
This is the live demo for Tip 7. Open Telegram, type this exact message to
RoboTerri. It will respond in Terri's style with a tweet thread about CYP2D6
and codeine. Point out: same pharmacogenomics insight from the ClawBio demo,
now repurposed as social media content. Then mention: /substack turns it into
a blog post, /playbook turns it into a podcast episode. One insight, many
formats, all from Telegram.
TIMING: 1 min (live demo)
Tips 1-7 are about using AI.
Tips 8-10 are about building with it.
This is where it gets interesting for bioinformaticians.
Transition slide. Pause here. "Everything so far is about using AI tools
that already exist. Now I want to talk about building. Because the
bioinformatics community has specific needs that general AI tools don't
address."
TIMING: 30 sec
Tip 8
Why Not Just Use ChatGPT?
Ask Claude to "profile my pharmacogenes from this 23andMe file."
It will try. It'll write plausible-looking Python. But:
It hallucinates star allele calls and uses outdated CPIC guidelines
It forgets CYP2D6 *4 is no-function (not reduced)
You spend 45 minutes debugging its output
No reproducibility bundle. No audit log. No checksums.
Announcing: 🦖 ClawBio
A domain expert's judgement, frozen into code that an AI agent executes correctly every time.
"Now, some of you are thinking: can't I just ask ChatGPT to do this?
You can try. Ask it to profile pharmacogenes from a 23andMe file. It'll
write plausible Python. But it'll hallucinate star allele calls. It'll
use outdated CPIC guidelines. It won't know that CYP2D6 *4 is no-function,
not just reduced. You'll spend 45 minutes debugging output that looks right
but isn't. ClawBio encodes the correct bioinformatics decisions so the agent
gets it right first time, every time. The skill is a domain expert's judgement,
frozen into code. Let me show you what that looks like."
TIMING: 3 min
🦖 What Is ClawBio?
The first bioinformatics-native skill library for AI coding agents
Built on OpenClaw — 180,000+ GitHub stars, the fastest-growing open-source AI project
A skill = a domain expert's knowledge frozen into code that an agent executes correctly every time
Local-first: your genomic data never leaves your laptop
MIT licensed: open-source, free, community-driven
Think of it this way:
ChatGPT/Claude = a smart generalist who guesses at bioinformatics
🦖 ClawBio skill = a domain expert's proven pipeline that the AI executes
Not a web app. Not a SaaS. Expert skills that run on your machine.
"So what exactly is ClawBio? It's the first bioinformatics-native skill library
for AI coding agents. It's built on OpenClaw — which has over 180,000 stars on
GitHub and is the fastest-growing open-source AI project right now.
Here's the key concept: a skill is a domain expert's judgement, frozen into code.
When an AI agent needs to do pharmacogenomics or ancestry analysis, instead of
hallucinating an answer, it loads the ClawBio skill and executes it correctly
every time. The difference: ChatGPT guesses at bioinformatics. A ClawBio skill
is a proven pipeline that the AI executes. Everything runs locally — your genomic
data never leaves your laptop. MIT licensed, fully open-source."
TIMING: 2 min
🦖 ClawBio: How It Works
You ask:"What drugs should I worry about from my 23andMe?"
## Skill Routing Table
| User Intent | Skill |
|---------------------|--------------------|
| Drug interactions, | pharmgx-reporter/ |
| 23andMe, CYP2D6 | |
| Genomic diversity, | equity-scorer/ |
| HEIM, FST | |
| Nutrition, MTHFR, | nutrigx_advisor/ |
| diet genetics | |
| Ancestry, PCA | claw-ancestry-pca/ |
## Safety Rules
- Genetic data never leaves this machine
- Always include disclaimer
- Use SKILL.md methodology only —
never hallucinate bioinformatics
Here's how ClawBio actually works. On the left: you ask a question in plain
English. CLAUDE.md has a routing table that matches your intent to the right
skill. The highlighted skill runs — in this case PharmGx Reporter — and
produces a report with figures, checksums, and a reproducibility bundle.
On the right: that's the actual CLAUDE.md file. It maps natural language
intents to skill directories. It also has safety rules — genetic data never
leaves your machine, and the agent is forbidden from hallucinating
bioinformatics decisions. It can only use what's in the SKILL.md.
TIMING: 2 min
Contributed within 24 hours of ClawBio going public
Here's the current roadmap. Five skills are at MVP — and that fifth one is
special. David de Lorenzo submitted the NutriGx Advisor skill within 24 hours
of ClawBio going public on GitHub. It covers 40 SNPs across 16 genes for
personalised nutrition — folate metabolism, vitamin D, caffeine, lactose,
omega-3. That's the power of open-source skills: a domain expert freezes
their knowledge into code, and the whole community benefits.
TIMING: 1.5 min
NutriGx Advisor — Demo by David de Lorenzo
First community skill — contributed within 24 hours of ClawBio going public. 40 SNPs across 16 genes for personalised nutrition.
Play the video. David recorded this demo of the NutriGx Advisor skill he
contributed to ClawBio. It shows the skill running on synthetic patient data,
generating a personalised nutrition report from 23andMe-style genetic data.
Let the video speak for itself — pause or narrate over it as needed.
TIMING: 1-2 min (video length)
Tip 9
Build Modular Skills
Live Demo: Creating a new ClawBio skill
In Claude Code, I type:
I want to build a GWAS quality control skill for ClawBio.
Walk me through how to create it.
Claude Code reads CLAUDE.md + CONTRIBUTING.md + SKILL-TEMPLATE.md
1. Copy template
→
2. Write SKILL.md
→
3. Add Python
→
4. Test + demo data
→
5. PR to ClawBio
The AI already knows the conventions. It scaffolds the skill for you.
"Let me show you how easy it is to build a new skill. I open Claude Code
in the ClawBio directory and ask: I want to build a GWAS QC skill. Claude
reads the CLAUDE.md, the CONTRIBUTING guide, and the skill template. It
walks me through five steps: copy the template, write the SKILL.md with
your domain methodology, add the Python implementation, create test data,
and submit a PR. David de Lorenzo did exactly this — he contributed the
NutriGx Advisor in under 24 hours. The AI scaffolds everything; the domain
expert provides the judgement."
TIMING: 2 min. Can do live in terminal if time permits.
The Reproducibility Crisis, In Practice
Today: Checking a Paper's Code
Read paper. Want to verify Figure 3.
Go to GitHub. Clone the repo.
Wrong Python version. Fix dependencies.
Need the reference data. Where is it?
Download 2GB from Zenodo. Link is dead.
Email first author. Wait 3 weeks.
Paths: /home/jsmith/data/
2 days later: still broken.
With ClawBio: Same Paper
Read paper. Want to verify the drug report.
git clone ClawBio && cd ClawBio
Ask: "Run the pharmgx demo"
CLAUDE.md routes to the right skill.
Report appears. Identical.
SHA-256 checksum: verified.
Demo data + dependencies: bundled.
30 seconds. Done.
Every figure in your paper should be one command away from reproduction.
"How many of you have tried to reproduce a figure from a paper using the
authors' GitHub code? You clone the repo, wrong Python version, dependencies
broken, Zenodo link dead, paths hardcoded. Two days later: still broken.
Now imagine: install the skill, run one command, Figure 3 appears, checksums
verified. 30 seconds. That's what reproducibility should look like."
TIMING: 2 min
Tip 10
Publish Skills, Not Just Papers
claw-pharmgx took a few hours to build. 12 genes, 51 drugs, CPIC guidelines.
claw-nutrigx was contributed by David de Lorenzo in under 24 hours. 40 SNPs, 16 genes.
5 MVP skills live. Every skill ships with checksums and a reproducibility bundle.
CLAUDE.md routes users to the right skill — no documentation needed.
Skills we need from you:
claw-gwas — PLINK/REGENIE automation
claw-metagenomics — Kraken2/MetaPhlAn wrapper
claw-acmg — Clinical variant classification
claw-pathway — GO/KEGG enrichment
"Tip 10: publish skills, not just papers. claw-pharmgx took a few hours
to build. David de Lorenzo contributed claw-nutrigx in under 24 hours
after ClawBio went public. The template is there, Claude Code scaffolds
the structure, and CLAUDE.md means users can find your skill just by
asking a question in plain English.
Think about your own work. That GWAS pipeline you run every time? Make
it a skill. That metagenomics classifier? Wrap it. That clinical variant
interpretation workflow? A few hours of work and the whole community
benefits. Imagine every bioinformatics paper shipping with a ClawBio
skill. You read the paper, ask a question, get the exact results.
That's the world we're building."
TIMING: 2.5 min
Some Common Pitfalls
What to watch out for when AI writes your research
Transition slide. "So we've covered 10 tips for using AI effectively.
But before we wrap up, let me flag three common pitfalls — things that
go wrong when you trust AI output without verification. These are
especially dangerous in bioinformatics."
TIMING: 15 sec
AI Pitfall: Numbers That Sound Right
The problem
LLMs generate numbers that feel appropriate.
They do NOT generate numbers that match your data.
What goes wrong:
"We analysed 847 samples" — plausible, but yours was 623
"p = 3.2 × 10⁻⁸" — looks right, entirely invented
"73% of variants were missense" — confident and wrong
🔢
Verify EVERY number
against your source files, code output, or analysis logs
"Here's the first AI pitfall that matters for bioinformatics: numbers.
LLMs are remarkably good at generating numbers that feel right. Sample sizes
that sound reasonable, p-values in the right ballpark, percentages that pass
the smell test. But they're invented. If you ask an LLM to write your results
section, it will confidently give you numbers that look publishable but don't
match your actual data. In bioinformatics, where every paper is full of
statistics, this is dangerous. Rule: verify every single number against
your source data."
TIMING: 1.5 min
AI Pitfall: Claims That Grow
The problem
LLMs write persuasively. They naturally reach for broader, more impressive claims.
What you found:
"CYP2D6 poor metabolisers showed altered codeine response in our cohort"
What the LLM writes:
"This proves that CYP2D6 testing should be mandatory before prescribing any opioid"
📈
For each conclusion ask:
"Did we actually measure this?"
"Second pitfall: claim expansion. LLMs are trained to write persuasively.
They naturally escalate findings. 'We observed X' becomes 'This proves Y'.
'Associated with' becomes 'causes'. In bioinformatics, the gap between
correlation and causation is everything. Your reviewers will catch this
and your credibility takes a hit. Simple test: for every conclusion, ask
yourself — did we actually measure this? If you're hedging mentally,
the text needs hedging too."
TIMING: 1.5 min
AI Pitfall: "We Used X" (You Didn't)
You ask AI to write your Methods section. It describes what it thinks you did — not what you actually did.
AI wrote:
"We used Tool A with default settings"
You actually used:
Tool B with custom parameters
Nobody can reproduce your work if the methods section describes a different workflow.
🔬
Check every step against your code
ClawBio solves this: skills embed the actual methodology in SKILL.md
"Third pitfall: the AI writes a methods section that sounds great but
describes a different workflow to the one you actually ran. It guesses
which tools you used, assumes default parameters, invents steps. If your
methods section describes the wrong procedure, nobody can reproduce your
work. That's why in ClawBio, every skill has a SKILL.md that documents
the actual methodology — the AI reads that instead of guessing."
TIMING: 1.5 min
The Golden Rule
AI drafts. You verify.
If you can't explain it, don't submit it.
Your name is on the work.
"This is the golden rule of AI-assisted research. The AI drafts, you verify.
If you can't explain every number, every method, every conclusion in your
paper — don't submit it. Your name is on the work. Not ChatGPT's. Not
Claude's. Yours. Use AI to be more productive, but never outsource your
scientific judgement."
TIMING: 1 min
10 Tips Recap
1. Use AI inside your IDE, not a browser
2. Build persistent memory
3. Feed it your context
4. Use voice, not just text
5. Automate your daily intelligence
6. Deploy persistent AI agents
7. Let AI compound your outputs
8. Why not just use ChatGPT?
9. Build modular skills
10. Publish skills, not just papers
The top 1% build systems. Everyone else uses tools.
Quick recap. Tips 1-7: use AI effectively — IDE, prompts, memory, voice,
automation, agents, compounding. Tips 8-10: build with AI — open-source,
modular skills, shipping infrastructure. The throughline: the top 1% build
systems that compound. Everyone else uses tools one at a time.
TIMING: 1 min
Thank You
Questions welcome — and at the pub after
GitHub: github.com/ClawBio/ClawBio
LinkedIn: linkedin.com/in/manuelcorpas
X: @manuelcorpas
🦖 Slides: github.com/ClawBio/ClawBio/slides
Thank you. I'll be at the pub after — come talk to me if you want to
build a skill, have ideas for the library, or just want to chat about
AI agents in research. Questions?
TIMING: remainder
