Human judgment, AI execution.
A plain-language tour of two AI-assisted workflows you can actually use in graduate study and clinical training: vibe researching (directing AI agents through a literature review or analysis) and vibe coding (describing a tool in words and letting an agent build it). The throughline: AI can carry the mechanical load, but you keep the judgment, the verification, the ethics, and the accountability.
Simulated data Every demo below uses invented teaching data. Nothing here is real client, student, or patient data, and no demo is a validated clinical tool.
You can delegate the work.
You cannot delegate the responsibility.
That one sentence is the whole lesson. AI agents can search, draft, compute, and build, but if your name is on the assignment, the poster, the paper, or the therapy material, you answer for every claim in it.
Orientation
Why this matters for SLP
You do not need to become a programmer to benefit from these tools, and you should not hand your clinical reasoning to a machine. The useful middle path is learning to direct AI well and check it rigorously. Here is where that shows up in real SLP work.
Evidence you can trust
Finding intervention literature for aphasia, articulation, fluency, voice, language, literacy, or AAC, then verifying every source instead of trusting a summary.
Interactive activities
Prototyping a practice activity, a visualized progress-monitoring chart, or a de-identified practice dataset to use in a class or lab, tested before any learner sees it.
Therapy-material drafts
Sketching a therapy-material prototype to react to and refine, clearly labeled as a draft, never mistaken for a validated tool.
The core idea
Vibe researching and vibe coding are not about turning off your brain. They shift some mechanical work to AI agents while keeping human judgment, clinical reasoning, verification, and ethical responsibility firmly in the loop. The skill you are building is direction and verification, not coding.
What AI can help with, and what stays human
Throughout this lesson you will see this split again and again. A quick first look:
AI agents can assist with
Searching and organizing literature, drafting tables and summaries, writing and running code, cleaning and graphing de-identified data, and producing first drafts of text, quickly and tirelessly.
Humans must still own
Choosing the question, judging relevance, interpreting clinical meaning, verifying sources and numbers, protecting privacy, deciding what is good enough, and being accountable for the final product.
A caution before we start
These workflows are useful for learning, prototyping, and research support. They are not shortcuts around clinical judgment, evidence-based practice, or data-privacy requirements. Keep that in mind as the tools get more impressive.
Vibe Researching: human vision, AI execution
Adapted for SLP students from a 2026 paper that maps this practice carefully. We will define it, place it among other ways of doing research, walk the human–agent loop, and take an honest look at the risks, because the risks are exactly why your judgment matters more, not less.
Part 1 · Vibe Researching
2.1 What vibe researching actually is
A beginner-friendly definition, then the five ideas that separate it from “just using ChatGPT.”
Working definition
Vibe researching is a way of doing research where a human sets the question, direction, standards, and final interpretation, while AI agents help with labor-intensive tasks: searching literature, organizing sources, writing code, cleaning data, generating figures, or drafting text. The human is still responsible for deciding what matters, checking whether outputs are correct, and standing behind every claim.
A 30-second preview of the vocabulary
Tap any card to flip it; tap again to flip back. You will meet each term properly in context, nothing to memorize now.
It is more than “using ChatGPT”
A single question and a single answer is not vibe researching. Vibe researching is a workflow: repeated cycles of asking, reviewing, redirecting, and verifying. Compare:
One-shot prompting
“Tell me about aphasia treatment.” You get a generic paragraph with no way to check it.
Vibe researching
Ask the agent to search recent intervention studies, organize them by population and outcome measure, flag missing evidence, and build a table, which you then check against the original papers.
The five ideas that define it
1 · Human as creative director
You choose the questions, judge what findings matter, and make the strategic calls. An agent can summarize childhood-apraxia studies, but you decide whether the populations and outcome measures are truly comparable.
2 · Natural language is the interface
You start with structured requests, not code: “Find recent telepractice-for-aphasia studies; organize by participants, intervention, outcomes, and limitations; do not invent citations.”
3 · Delegation with oversight
Agents can draft, organize, compute, and build. You must check. This is where research parts ways with throwaway code: you cannot publish a paper you have not actually read.
4 · Iterative refinement
The work improves in loops. Too broad → narrow to adolescents who stutter → keep only empirical intervention studies → separate group designs from single-case. Each cycle tightens the focus.
5 · Human accountability
At the end of the day, your name goes on the work and you answer for every claim. AI agents are instruments, not co-authors. You must be able to defend every conclusion.
SLP example · refinement in action
Prompt 1: “Find papers on fluency intervention.” → too broad. Prompt 2: “Focus on adolescents who stutter and school-age outcomes.” → some are not intervention studies. Prompt 3: “Keep only empirical intervention studies; separate group designs from single-case.” → now the table is worth inspecting against the original papers.
Remember
You can delegate the searching, the drafting, and the formatting. You cannot delegate the reading, the judging, or the responsibility. You can delegate work; you cannot delegate responsibility.
Part 1 · Vibe Researching
2.2 Where it sits: the collaboration spectrum
Research with AI is not all-or-nothing. It runs along a spectrum from fully human to fully automatic. Vibe researching sits in a specific spot, and knowing where helps you use it well.
The key distinction
AI for Science uses an AI model for one computation (for example, predicting a protein structure) but leaves the research process unchanged. Vibe researching changes the process itself: agents help across literature, implementation, analysis, and drafting. Auto research tries to remove the human entirely. For student learning and responsible training, the human-in-charge middle is usually the right place to be.
Part 1 · Vibe Researching
2.3 Who does what: the division of labor
Across these paradigms, the same research tasks get assigned differently. Two rows never move in vibe researching, no matter how much you delegate: quality control and accountability.
| Research task | Traditional | Tool-assisted | AI for Science | Vibe researching | Auto research |
|---|---|---|---|---|---|
| Idea generation | Human | Human | Human | Human | AI |
| Literature review | Human | Semi-auto | Human | AI (human-guided) | AI |
| Experiment design | Human | Human | Human | Human + AI | AI |
| Domain computation | Human | Human + tools | AI model | AI agent + tools | AI |
| Implementation | Human | Human + tools | Human | AI (human-reviewed) | AI |
| Data analysis | Human | Human + tools | Human + AI | AI (human-directed) | AI |
| Writing | Human | Human | Human | AI (human-edited) | AI |
| Quality control | Human | Human | Human | Human | AI (or none) |
| Accountability | Human | Human | Human | Human | Unclear |
Notice the last two rows: even when agents do the literature review, implementation, analysis, and drafting, a human still owns quality control and accountability. In auto research, those become “AI or none” and “unclear”, which is exactly the problem.
Part 1 · Vibe Researching
2.4 Vibe researching vs. auto research
Both use the same underlying AI techniques. The difference is not technological: it is who orchestrates the process.
The one-liner
In vibe researching, the human is the meta-agent. You observe each step, decide whether to proceed or back up, and adjust direction in real time. Auto research hands that role to a pipeline or a controller agent.
| Aspect | Vibe researching | Auto research |
|---|---|---|
| Orchestration | Human decides what to do next | Meta-agent or fixed pipeline controls the flow |
| Goal specification | Open-ended; evolves through conversation | Defined upfront as a structured input |
| Error recovery | Agent self-checks + human redirects | Agent self-corrects or fails silently |
| Quality control | Human checkpoints, aided by agent verification | Automated evaluation (e.g., simulated review) |
| Scope of autonomy | Per-task; human grants and revokes | End-to-end; runs to completion |
| Adaptability | High; you can pivot at any moment | Low; the pipeline structure is fixed |
| Reproducibility | Requires logging the interaction | Deterministic pipeline is inherently reproducible |
Why the slower path is often the safer one
Keeping the human in the loop is less efficient in wall-clock time: you are the bottleneck. But it is far more robust to the errors that compound silently in automated pipelines: a flawed setup nobody noticed, a misframed question, a subtle statistical mistake a “simulated reviewer” misses. For SLP training, that reliability is worth the slowdown.
Part 1 · Vibe Researching
2.5 The human–agent loop
At its core, vibe researching is a five-step loop you repeat, sometimes dozens of times in an hour. You think; the agent acts; you decide what happens next.
Zooming out: the five-phase workflow
Those quick loops organize into a larger arc. Notice which phases are human-led and which are AI-executed with human oversight: the entry and exit of every phase stays with you.
You pick the research question.
Agents survey the literature and map the landscape.
Agents implement and run analyses under your direction.
Agents compile findings into draft text, tables, and figures.
You take ownership of the final output, verifying every claim.
Part 1 · Vibe Researching
2.6 What makes it work: enabling techniques
You do not need to build any of this, but knowing the moving parts helps you ask for the right thing and spot when something is off. Each technique, in plain language, with an SLP example.
| Technique | What it means | SLP example |
|---|---|---|
| Multi-agent systems | Different AI agents take different roles instead of one agent doing everything. | One agent searches aphasia literature while another builds the summary table. |
| Memory mechanisms | Project files and logs help the agent remember decisions across days or weeks. | The agent remembers this lesson should reuse the same navy classroom style. |
| Tool use & skills | The agent can search, run code, process files, and reuse procedures, not just talk. | The agent builds a progress-monitoring graph from a simulated dataset. |
| Planning & decomposition | The agent breaks one big request into an ordered list of smaller steps. | Plan the full lesson outline before writing any code. |
| RAG (source retrieval) | The agent looks up real sources before answering instead of guessing from memory. | It retrieves and lists papers so you can verify each one actually exists. |
| Self-reflection | The agent checks its own work and flags where it is unsure. | It notes “this citation needs verification” rather than stating it as fact. |
Optional detail · what a multi-agent SLP setup might look like
A natural division of labor: a literature agent searches databases and produces structured summaries; a coding agent builds an interactive articulation activity; an analysis agent graphs de-identified progress data; a writing agent drafts a plain-language summary. In every case, the human verifies all sources, data handling, and interpretation. The point of multiple agents is focus, not autonomy: you still decide when to call each one and whether its output is acceptable.
One principle ties these together
A vibe-researching agent should never be trusted to just “know” something: it should be equipped to look it up, run it, and check it. That shifts the failure mode from confident fabrication toward retrieval and reasoning errors, which are easier (though not easy) to catch.
Part 1 · Vibe Researching
2.7 Risks, limitations, and responsible use
The paper names seven technical limitations. For each, here is the concrete SLP danger and a one-line rule to protect yourself. These are not reasons to avoid AI: they are reasons your judgment matters.
Hallucination & lack of rigor
Agents can invent citations, misstate findings, and build plausible but wrong arguments. A fabricated reference can slip into a literature review unnoticed.
Rule: check original sources before trusting any summary.
Context loss
An agent’s working memory is finite. On long projects it forgets earlier constraints and can contradict its own earlier outputs.
Rule: keep project instructions and decisions in files.
Infrastructure not built for agents
Paywalled articles, missing APIs, broken links, and awkward file access mean an agent often cannot do what you could do in a browser in minutes.
Rule: expect friction; document what worked.
Limited multimodal understanding
Agents can accept images or audio in some tools, but that is not the same as clinical understanding of speech, language, voice, or swallowing.
Rule: don’t rely on AI alone for client-specific interpretation.
Verification asymmetry
The tasks most worth delegating are the hardest to check. Auditing a script or a stats analysis can take as much expertise as writing it.
Rule: if you cannot verify it, do not present it as fact.
Brittleness on novel tasks
Agents are strongest on familiar patterns and weakest on genuinely new problems: they default to safe, conventional answers.
Rule: treat unusual clinical/research questions with extra caution.
Data privacy & intellectual property
Cloud-based agents may process your input outside your device. Unpublished ideas, preliminary results, and especially identifiable client data can be exposed.
Rule: do not upload identifiable clinical or student data unless explicitly permitted by policy, consent, and privacy rules.
Part 1 · Vibe Researching
2.8 Impacts, trade-offs, and the road ahead
Vibe researching changes how science gets done, and the changes cut both ways. An honest view holds the benefits and the costs in the same hand.
Positive impacts
- Doing more with less: smaller labs and individuals can survey, prototype, and draft at a pace that used to need a team.
- Faster iteration: literature surveys and data wrangling compress from weeks to days.
- Expanding research coverage: side hypotheses worth “a look” can actually get looked at.
- Crossing disciplinary boundaries: agents act as translators into unfamiliar fields.
- Reducing cognitive load: less grind on formatting and reimplementation, more time to think.
- Surfacing hidden connections: scanning vast literatures can flag links a single-domain reader would miss.
Negative impacts
- Convergent thinking: similar models can push everyone toward the same framings and well-known methods.
- Credit & disclosure: norms for how much AI use to declare are unsettled.
- Flooding the literature: cheaper papers can mean more technically-correct-but-unexciting work.
- Polished mediocrity: fluent formatting can decouple how good work looks from how good it is.
- Erosion of public trust: if readers doubt the humans understand their own papers, trust frays.
- Devaluation of expertise: deep skill can become invisible when output looks similar.
- Erosion of training: skipping the formative struggle weakens the judgment good delegation depends on.
Student takeaway
AI can make weak work look polished. That is exactly why source-checking, clinical reasoning, and human accountability matter more, not less.
From risk to responsibility: what to do now
For every limitation, researchers are pursuing a future fix, but you do not have to wait. Each row below pairs the problem with the direction the field is heading and a concrete habit you can adopt today.
| Problem | Future direction (the field) | What you should do now |
|---|---|---|
| Hallucination | More reliable generation (retrieval + fact-checking, calibrated uncertainty) | Check every citation and number against the source yourself. |
| Context-window limits | Persistent project memory across sessions | Keep decisions in instruction files; re-state context each session. |
| Infrastructure gap | Agent-native infrastructure (open APIs, standard tool protocols) | Expect friction; record what worked and what didn’t. |
| Multimodal limits | Domain-specific vision and embodied/lab-automation agents | Don’t trust AI for clinical audio/visual judgment. |
| Verification asymmetry | Verification tooling (citation checks, statistical sanity checks) | Budget real time to verify; never skim a result you’ll stand behind. |
| Brittleness on novelty | Novelty-aware agents that escalate when uncertain | Supervise unusual questions closely; expect to do more of the work. |
| Data privacy | Privacy-preserving architectures (on-device, stronger open models) | Never upload identifiable data without explicit permission. |
| Convergence, flooding, polish | Community norms & standardized AI-use disclosure | Disclose your AI use; judge substance over polish. |
| Erosion of training | Research-education reform (fundamentals first, then delegation) | Learn the fundamentals the hard way before you delegate them. |
Reference for this section
Feng, Y., & Liu, Y. (2026). A visionary look at vibe researching. arXiv. https://arxiv.org/abs/2604.00945
Vibe Coding with Claude Cowork: from idea to working prototype
The same logic as vibe researching, pointed at software: you describe what you want in plain language, an agent builds it, and you test and refine. You will not write every line: you will learn to direct, review, and verify a small, useful tool.
Part 2 · Vibe Coding
3.1 What vibe coding is (and is not)
Working definition
Vibe coding is a workflow where you describe what you want the software to do in plain language, the coding agent writes or edits the code, and you repeatedly test, review, and refine the result. The “vibe” is not guesswork; it is a guided loop: give context, define the goal, review the result, and keep improving until the outcome matches your success criteria.
The most important caution
Vibe coding is not “accept everything.” The original idea included an “Accept All” habit meant for throwaway prototypes. For anything you will use with learners or clients, it is a cycle of specifying, reviewing, testing, and revising: the same delegation-with-oversight you saw in vibe researching.
Why Claude Cowork for this lesson
Built for non-coders
You can begin with natural-language directions: no terminal, Git, or JavaScript knowledge required to get started. The barrier to entry is a sentence.
Works across the project
It can work with web pages, JavaScript, Python, data files, Markdown instructions, and whole project folders, like the one that produced this page.
Still needs you
It needs direction, constraints, and review. It does not guarantee correctness: you provide the goals and the verification.
Honest about limits
A working prototype is a draft, not validated clinical software. The same risks from Part 1 (hallucination, privacy, verification) apply here too.
Part 2 · Vibe Coding
3.2 The four principles
These four rules come from widely shared observations about where coding agents go wrong. They are not just engineering taste: they protect non-coders from confusing, bloated, or unreliable output.
Think before coding
Before changing code, the agent should pause, state assumptions, name what is ambiguous, and ask questions, instead of guessing silently and running with it.
Fixes: wrong assumptions, hidden confusion, missing trade-offs.
SLP example
Instead of “make a language activity,” the agent asks: “Child language, adult aphasia, or AAC examples?” before building anything.
Simplicity first
The first version should be the smallest useful one: no speculative features, no abstractions for single-use code. If 200 lines could be 50, rewrite it.
Fixes: overcomplication, bloated abstractions.
SLP example
Skip accounts, logins, databases, and dashboards. Start with a single-page interactive demo that runs in the browser.
Surgical changes
When editing an existing project, change only what the task requires. Don’t refactor unrelated parts, rename things, or “improve” nearby code. Match the existing style.
Fixes: stray edits, touching code you shouldn’t.
SLP example
If the task is “fix the quiz feedback wording,” the agent should not redesign the page or rename classes.
Goal-driven execution
A strong request describes what success looks like and how to check it. Clear success criteria let the agent loop on its own; vague ones (“make it work”) force constant clarification.
Fixes: aimless effort, endless back-and-forth.
SLP example
“Make a flip-card activity with 8 cards; students reveal definitions by click or Enter; verify it works on mobile and keyboard.”
Part 2 · Vibe Coding
3.3 Good vibe coding starts before code
The single biggest upgrade to your results is not a better prompt: it is giving the agent the right project memory and constraints up front, in plain Markdown files.
| File | What it tells the agent |
|---|---|
| README.md | What the project is and who it is for. |
| SPEC.md | What content and features to build, and what not to build. |
| STYLE.md | How the page should look and feel. |
| CONTENT_SOURCES.md | Which sources to read and how to use them. |
| INTERACTIONS.md | What kinds of interactive activities to include. |
| DEMOS.md | How to propose and implement SLP demo tools. |
| CLAUDE.md | How the agent should behave while working: its rules, limits, and quality gates. |
A workflow you can copy
Tap each step for a sample prompt and the common mistake it avoids.
Part 2 · Vibe Coding
3.4 From vague to verifiable
A vague request gets a generic result. A well-framed one gets useful work on the first try. The fix is rarely “more words”; it is the right words: who, what, with what data, and how you’ll know it worked.
What not to do
Don’t upload identifiable client videos, audio, reports, or transcripts. Don’t accept every change without testing. Don’t assume a clinical tool is validated. Don’t let the agent rewrite unrelated parts. Don’t use AI-generated citations without checking them. Don’t confuse a working prototype with evidence-based clinical software.
What to do
Start with a clear purpose. Tell the agent the audience. Give examples. Define success criteria. Ask questions before coding. Keep the first version small. Test it yourself. Save instructions in project files. Document what the agent changed.
Reference for this section
Multica AI. (n.d.). andrej-karpathy-skills [Computer software]. GitHub. Retrieved June 16, 2026, from https://github.com/multica-ai/andrej-karpathy-skills
Three teaching prototypes, vibe-coded
Each demo below is a small, working tool built with the workflow you just learned. Try them, then read how each one was requested, what had to be checked, and where its limits are.
Part 3 · SLP Demos
4 Try the demos
These run entirely in your browser on simulated data: nothing is uploaded or saved. They illustrate the vibe-coding workflow in familiar SLP territory.
Demo 1 · Listen, identify & transcribe
Demo 2 · Language sample analysis teaching toy
Demo 3 · AAC core vs. fringe vocabulary sorter
What to notice about all three
Each is small, runs locally, uses only simulated data, and is clearly labeled as a prototype. That is not a limitation of the workflow. It is the workflow: build the minimum useful version, keep the human checking it, and never dress a draft up as a validated clinical tool.
Wrap-up
5 Responsible use & takeaways
Whether you are vibe researching a literature review or vibe coding a practice tool, the same seven habits keep you (and the people you serve) safe. Use this as your pre-flight check.
Bad prompt
Make me a therapy website.No audience, no task, no data policy, no success criteria. You will get something generic and spend more time fixing it than it saved.
Good prompt
Single-page /r/ minimal-pair practice for grad SLP students. Reuse the navy classroom style. 10 items, a progress counter, instructions, and a reset. Simulated data only. Propose the flow and success criteria before coding.Audience, task, style, scope, data policy, and a plan-first request, verifiable from the start.
The whole lesson in one breath
Vibe researching and vibe coding let you hand the mechanical work to AI agents while you keep the judgment. They reduce friction, which is exactly why they are risky: it becomes easy to produce polished work before you have done enough thinking. So check your sources, test your tools, protect your data, and stand behind what carries your name. The risk is not that AI replaces clinicians; it is that we use AI carelessly and call the result good work. Don’t.