Module 01 · Lesson 03

The Five Universal Capabilities

Reading time: 20 minutes Track: Universal Foundations · Required for all learners Prerequisites: Module 01 · Lessons 01–02

What this lesson does

Lessons 01 and 02 gave you the strategic frame and the mental model. This lesson gets specific about what skills you're actually building.

There are five capabilities every biotech professional needs to develop, regardless of role. These are the skills that show up in every successful AI workflow in this industry — from a Bench R&D scientist accelerating literature review to a CFO drafting investor materials. The role-specific applications differ. The underlying capabilities do not.

By the end of this lesson, you'll be able to:

1. Name and describe the five universal capabilities
2. Self-assess your current level on each
3. Identify which capability is your weakest leverage point
4. Understand how the rest of the curriculum builds each one

These five capabilities are the spine of everything else you'll learn in HaiPhai. Each subsequent module deepens at least one of them, often more.

01 · The five capabilities, briefly

Before we go deep on each, here they are in one place:

If you mastered nothing else from this curriculum and just got competent at these five, you'd be in the top 5% of biotech AI users. Most learners are weak in 3-4 of them when they start, and that's normal. The goal isn't to be strong in all five immediately — it's to know which ones you're developing and why.

Let's go through each.

02 · Capability 1 — Specification writing

What it is

Specification writing is the skill of turning a vague task ("help me with this report") into a precise prompt that produces reliable, near-final-draft output ("Draft Section 5.2 of the ISS, covering Grade 3+ AEs by SOC, formatted as standard ISS narrative with subgroup analysis flagged separately, using FDA preferred terminology, tone aligned to our previously approved Phase 2b submission").

It's not a writing skill in the literary sense. It's an engineering skill. The output of a well-specified prompt is predictable, repeatable, and high-quality. The output of a vague prompt is a coin flip.

Why it matters most

Of the five capabilities, specification writing has the highest immediate ROI. Most people stuck at 2x productivity gain are bottlenecked here. The leap from 2x to 5x productivity gain is almost entirely about specification quality.

In biotech specifically, specification writing carries an additional weight: it forces you to think clearly about what you actually want, which is itself a valuable discipline. The prompts you write for AI tend to make you a better drafter of human requests too. Your colleagues notice this within weeks.

The shape of mastery

You know you've mastered specification writing when:

• Your first prompt produces 80% of the final output, not 30%
• You can think about a task and immediately recognize the 6-7 specifications it requires
• You catch yourself spotting under-specified requests from colleagues and either fixing them or asking the right clarifying questions
• You build prompt templates for tasks you do regularly, instead of rewriting them each time

Common failure pattern

The most common failure isn't writing bad specifications — it's writing no specifications. People type a question and hit enter, then get frustrated when the output is generic. The specification doesn't exist in the prompt, so the AI fills the gap with its averaged training-data default. The fix isn't to "try harder on the prompt" — it's to systematically write out role, context, goal, format, constraints, success criteria.

This is what Module 02 teaches in depth.

03 · Capability 2 — Verification habit

What it is

Verification habit is the automatic, non-negotiable practice of checking AI-generated content against authoritative sources before using it for anything that matters.

The word "habit" is intentional. Verification as a one-time effort is useless — you'll forget when you're tired or rushed. Verification as a habit, built into your default workflow, is what separates the safe user from the dangerous one.

Why it matters most in biotech

You learned in Lesson 02 that AI produces confident text regardless of accuracy. In most industries, this produces embarrassment when caught. In biotech, it produces regulatory exposure, patient safety risk, and career damage.

A medical writer who got 3 fabricated citations past initial review and into a senior reviewer's hands will not be promoted soon. A pharmacovigilance scientist who relied on AI for a drug-interaction profile and missed an inversion will be the subject of an internal investigation. A regulatory affairs specialist who used AI for an FDA Information Request response and didn't catch a misstated study endpoint will be in a difficult meeting with their VP.

These aren't hypotheticals. They are recurring incidents in this industry right now, and the rate is increasing.

The shape of mastery

You know you've mastered verification habit when:

• You don't think about whether to verify — you just do, automatically, before any AI output exits your workspace
• You have a default verification depth calibrated to output stakes (Lesson 02, Section 04)
• You routinely ask adversarial questions ("what's wrong with this?") rather than confirmatory ones ("is this right?")
• You document AI usage in version control or document metadata so reviewers know what to scrutinize
• You catch yourself feeling annoyed when colleagues skip verification, because you recognize the risk

Common failure pattern

Verification fatigue. After dozens of cases where the AI was right, you start to skip the checks because "it's usually accurate." This is the trap. The accuracy of AI output is not uniformly distributed — it's high on common topics and craters on rare ones. The cases where verification matters most are the cases that look just like the cases where verification didn't matter. You can't tell which is which without checking.

Module 03 builds verification into the foundation of how you work, then every role module reinforces it in the context of that role's actual outputs.

04 · Capability 3 — Workflow design

What it is

Workflow design is the skill of going beyond one-off prompts to build repeatable systems that handle entire categories of work.

A one-off prompt is "draft a protocol deviation memo for this missed visit." A workflow is "here is the template, the prompt sequence, the QC checks, the formatting rules, and the audit trail logic that produces compliant PD memos for any deviation scenario in our trials, ready for QC review."

The difference is the difference between using AI as a typewriter and using it as infrastructure.

Why it matters more than people think

Most AI users plateau because they treat every task as a fresh prompt. They get good at prompting (Capability 1), they verify well (Capability 2), but they don't build systems. So every task is full-cost, every time.

The 5x-to-20x productivity gain — the gain that actually transforms a function — comes from workflow design. The first PD memo takes you 40 minutes with AI assistance versus 90 minutes without. Substantial gain. But once you've built a workflow, the 100th PD memo takes 8 minutes, the AI does most of the drafting, and your time is spent on the 5% of judgment calls that actually require a human. That's the curve people aren't on, and it's where the real value lives.

The shape of mastery

You know you've mastered workflow design when:

• You can identify which of your recurring tasks are workflow candidates (high frequency, structured output, stable requirements)
• You build first-version workflows even when imperfect, because you know iteration is the path to good ones
• You document your workflows so others on your team can use them
• You think about handoffs between people and AI as deliberate gates (where does the human decide, where does the AI draft, where does QC happen)
• You spot opportunities to compose workflows together — outputs of one becoming inputs of another

Common failure pattern

Building too sophisticated too early. Some learners try to design a perfect, comprehensive workflow on their first attempt — and either give up because it's too hard or build something so complex no one else can use it. The right approach is iterative: build a workflow that handles 60% of cases, use it, refine. After 3-4 iterations, you have something genuinely good.

Modules 06 (Connectors & Tools), 07 (Agent Design), and 08 (Skills & Customization) cover workflow design in depth. Module 09 (Cross-Functional) covers how workflows hand off between functions.

05 · Capability 4 — Tool and model selection

What it is

Tool and model selection is knowing which AI tool, which model, and which configuration fits which task — and not defaulting to one tool for everything.

This sounds technical and small. It's actually large. The same task — say, drafting a clinical study report section — gets meaningfully different output from:

• Claude Opus vs. Claude Sonnet (depth vs. speed tradeoff)
• Claude with file uploads vs. Claude in the API (context vs. iteration tradeoff)
• A general-purpose tool vs. a regulated-environment tool with BAA (capability vs. compliance tradeoff)
• A tool with web search vs. one without (recency vs. determinism tradeoff)

A user who defaults to one tool for everything is leaving 30-50% of value on the table. A user who matches tool to task is operating at a different level.

Why it matters more over time

In 2026, the AI tool landscape is fragmenting rapidly. There's a frontier model from each major lab (Claude, GPT, Gemini), specialized models for code, specialized tools for retrieval, specialized agents for specific functions. Two years from now, the landscape will look completely different.

The skill you're developing isn't "knowing which 2026 tool to use" — it's the meta-skill of evaluating new tools against your tasks and picking the right one. That meta-skill is durable across the changes ahead.

The shape of mastery

You know you've mastered tool and model selection when:

• You have a default routing rule for common task types (e.g., "Claude Opus for regulatory drafts, Claude Sonnet for routine summaries, an open-source local model for anything involving sensitive patient data")
• You can evaluate a new AI tool's fit for your work within 30 minutes of trying it
• You don't get attached to one vendor — you're outcome-oriented
• You know the data handling, training, and retention policies of the tools you use, well enough to make compliant choices automatically
• You stay current enough to know when something materially better has shipped

Common failure pattern

Vendor capture. Some learners pick one tool early (often whatever their company deployed first) and never reassess. They miss substantial gains because better-fitting tools exist for specific tasks. The fix is curiosity and structured re-evaluation every quarter or two.

Module 05 (Connectors & Tools) covers this in operational detail. Each role module also notes where tool selection matters for that role's specific outputs.

06 · Capability 5 — Governance fluency

What it is

Governance fluency is the ability to make compliant choices about AI use automatically, without consulting a flowchart for every decision.

It's the same skill that lets an experienced clinical operations lead glance at a protocol deviation and immediately know it's Minor vs. Major without rereading the SOP. It's pattern recognition built on top of internalized principles.

For AI use in biotech, governance fluency means:

• Instinctively knowing what data not to put into a tool
• Recognizing which tools have what data handling
• Documenting AI usage appropriately without it feeling burdensome
• Catching colleagues' near-miss governance mistakes before they become incidents
• Making judgment calls in gray zones using internalized principles

Why this is non-negotiable

Of the five capabilities, governance fluency is the only one where being weak is dangerous in addition to being suboptimal. You can be a bad specification writer and just be slow. You can have a weak verification habit and just produce some errors. But if you're weak on governance, one bad choice — one paste of PHI into a public tool, one shared confidential trial document, one MNPI leak — can end your career, expose your company to regulatory action, or harm patients.

This is why Module 03 (Governance) is required before role specialization. It is the only required module in the curriculum besides the foundations and capstone. You cannot complete certification without it, and that's intentional.

The shape of mastery

You know you've mastered governance fluency when:

• You don't need to think about what's safe to share — you just know, automatically
• You can teach the principles to a new colleague clearly and concisely
• You catch governance issues in your colleagues' AI workflows during review
• You can defend your AI usage to an FDA inspector, an internal auditor, or your company's general counsel
• You see governance as enabling AI use, not restricting it — the constraints make confident use possible

Common failure pattern

Treating governance as a one-time training rather than an ongoing practice. New scenarios emerge constantly (new tools, new use cases, new data types). The fluency comes from continually applying principles to new situations, not from memorizing a list of rules. Module 03 emphasizes the principles, not the rule list, for exactly this reason.

07 · Self-assessment — where are you now?

Spend 60 seconds rating yourself on each capability. Use this scale:

• 1 — Novice: I don't really do this, or I do it inconsistently and produce mixed results
• 2 — Developing: I do this sometimes, often well, but it's not yet automatic
• 3 — Competent: I do this regularly and reliably, and the results show
• 4 — Strong: I do this automatically, can teach others, and continuously improve
• 5 — Expert: I design how others should do this, and I'm helping define best practices

Write your ratings down somewhere you can revisit them. We'll come back to this self-assessment at the end of Module 10 to see how much you've moved.

What your ratings tell you

If most are 1-2: You're at the start of your AI fluency journey, which is exactly where this curriculum expects you to be. Don't rush. Each capability needs deliberate practice.

If most are 2-3: You're a typical mid-development biotech AI user. The curriculum will push your strongest capability to 4 and your weakest to 3, which is a significant performance gain.

If most are 3-4: You're already operating well. The curriculum will tighten weak spots and give you the framework to teach others. Capability 3 (workflow design) is most likely where you have the most upside.

If most are 4-5: You're an outlier; this curriculum is mostly reinforcement and framework for you. Pay particular attention to advanced modules (06-09) and consider whether you should be involved in defining your organization's AI practices.

Your weakest capability is your highest leverage

Don't try to push your strongest skill from 4 to 5. Push your weakest skill from 2 to 3. The gain is bigger and the impact on your overall capability is larger. This is the same principle as exercise — train your weakest muscle group, not your strongest.

08 · How the curriculum builds these capabilities

Here's a transparent map of where each capability gets developed across the rest of the curriculum:

By the time you reach the Capstone, you should have moved at least one point on every capability and two or three points on at least one. If you complete the curriculum and your ratings haven't changed, something went wrong — either with your engagement or with our content. Either way, we want to know.

09 · Knowledge check

Three questions to lock in this lesson.

Q1. Of the five universal capabilities, which has the highest immediate ROI for someone new to structured AI use?

a) Workflow design — it produces the biggest long-term gains b) Specification writing — it most directly improves the quality of every individual output c) Tool and model selection — picking the right tool is half the battle d) Governance fluency — without it, nothing else matters

Q2. Why is verification habit emphasized as a habit rather than a task?

a) Because verifying is harder than other AI skills b) Because verification as a one-time effort is unreliable — fatigue and confidence cause skips; verification as an automatic habit, built into default workflow, is what makes it actually happen consistently c) Because regulators require it to be habitual d) Because habits are easier to teach than skills

Q3. Why is governance fluency the only capability where weakness is dangerous in addition to being suboptimal?

a) Because non-compliant AI use is the slowest of the failure modes b) Because compliance is an industry norm c) Because a single non-compliant AI choice — sharing PHI, sponsor data, MNPI — can produce career-ending or company-damaging consequences, unlike other capabilities where weakness produces lower-quality work rather than risk d) Because governance fluency is the hardest of the five capabilities

Answers: Q1: b · Q2: b · Q3: c

10 · What's next

You now have the strategic frame (Lesson 01), the mental model (Lesson 02), and the skill map (Lesson 03). One more lesson to close out Module 01:

Lesson 04 · Your 90-Day Learning Path — How to plan your journey through the rest of the curriculum, given your role, your stage, and your goals.

After Lesson 04, you'll move into Module 02 (Prompt Mastery), which begins building Capability 1 in depth.

End of Lesson 03.