Independent candidate proof of work
One challenge in the HBS AI Institute Instructional Designer role stood out to me: translating emerging faculty research into learning experiences executives can understand, practice, and apply. Rather than only describe how I would approach that work, I tested the process myself.
In a few focused hours, I researched the public role specification, reviewed HBS AI studies, selected a peer-reviewed research foundation, and built a complete 60-minute executive workshop — along with the participant, facilitator, measurement, accessibility, and AI-enabled production systems required to support it.
Independent candidate prototype created using publicly available HBS research. Not an official Harvard Business School product.
Research paper
Peer-reviewed field experiment
Executive decision
Where should AI act, and who verifies?
Learning experience
60-minute experiential working session
Workplace experiment
Seven days, one task, one decision rule
The goal was not to create the greatest possible quantity of material. The goal was to demonstrate an end-to-end research-to-learning production system.
01 — The challenge
The job description calls for more than conventional content production. It requires an instructional designer who can connect faculty, subject-matter experts, facilitators, and program teams; translate academic research into participant-centered experiences; produce assets across formats; support delivery; measure performance and impact; and build AI-enabled production systems that improve over time.
I treated those responsibilities as the project specification.
| Role responsibility | What it means in practice | Evidence in this project |
|---|---|---|
| Research translation | Preserve findings and limitations while converting them into executive decisions | Faculty research brief, claim ledger, framework rationale, workshop |
| Experiential learning | Make participants perform a meaningful task | Frontier Fit Map based on a real organizational workflow |
| End-to-end asset production | Produce a connected learning system rather than isolated slides | Deck, workbook, facilitator guide, takeaway card, assessment, digital specification |
| Facilitation readiness | Make another facilitator capable of delivering the experience | Minute-by-minute guide, speaker notes, backup demo, misconceptions, contingencies |
| Measurement | Demonstrate learning and transfer rather than satisfaction alone | Baseline, rubric, post-performance task, 7/30/90-day follow-up |
| AI-enabled production | Use AI to accelerate repeatable work while retaining human accountability | Production workflow, approval gates, SOP, prompt library, versioning |
| Continuous improvement | Combine learner data, facilitator observations, and faculty review | Feedback instrument, analysis process, revision example |
Opens the portfolio PDF. The map appears near the beginning and again at the end of the document.
02 — Research
I reviewed public HBS and Harvard AI research for three qualities: evidentiary strength, executive relevance, and the ability to support an experiential activity rather than a passive lecture.
Nine candidates reviewed. Scoring scale: 1 = weak fit, 10 = exceptional fit. Scores are prototype judgments, not Harvard ratings. Expand a candidate for researchers, finding, relevance, exercise concept, and the selection judgment.
RQ 10/10ER 10/10EL 10/10PU 10/10JD 10/10ID 10/10
RQ 10/10ER 10/10EL 9/10PU 9/10JD 10/10ID 9/10
RQ 9/10ER 10/10EL 10/10PU 10/10JD 10/10ID 10/10
RQ 9/10ER 10/10EL 9/10PU 10/10JD 10/10ID 9/10
RQ 9/10ER 9/10EL 10/10PU 9/10JD 10/10ID 10/10
RQ 8/10ER 10/10EL 10/10PU 9/10JD 9/10ID 10/10
RQ 8/10ER 10/10EL 9/10PU 9/10JD 9/10ID 9/10
RQ 9/10ER 9/10EL 9/10PU 9/10JD 9/10ID 9/10
RQ 9/10ER 9/10EL 9/10PU 9/10JD 9/10ID 9/10
Selected research
Navigating the Jagged Technological Frontier: a preregistered, randomized field experiment with 758 working consultants on tasks designed to resemble their work, published in Organization Science in 2026.
The same broad technology category did not create the same result across similar-looking knowledge-work tasks. The practical leadership problem is therefore not simply whether to adopt AI. It is how to test task fit and assign authority.
This study does not prove that every model, worker, industry, or workflow will produce the same results. The experiment used a specific 2023 version of GPT-4, a consultant sample, and designed work tasks. The location of the capability boundary changes.
The selected research is public HBS-affiliated research. The learning framework and executive application are my instructional translations and require faculty validation.
All figures above are drawn from the cited paper and the HBS AI Institute summary. Any executive implication beyond those figures is a prototype implication and requires faculty validation.
03 — The translation
It is: Where should AI act, where should a human verify, and what evidence justifies that decision?
Research finding
“AI capability is uneven across tasks.”
Executive problem
“Similar-looking workflows may require different operating modes.”
Learning capability
“A leader can test one task and assign AI an evidence-based role.”
Participant output
“A seven-day experiment with a baseline, quality metric, accountable human, escalation condition, and stop rule.”
04 — Backward design
By the end of the experience, a participant can decompose a real workflow, form a hypothesis about AI fit, design a representative test, assign an appropriate human–AI operating mode, and define a measurable seven-day experiment.
| Participants can… | Required knowledge | Learning activity | Evidence of learning |
|---|---|---|---|
| Decompose one business workflow into observable tasks and decision points | Workflow versus task; inputs, outputs, handoffs, exceptions | Workflow strip and task decomposition | At least 4 distinct tasks with owners, inputs, outputs, and failure consequences |
| Form an evidence-based hypothesis about where AI may help, harm, or remain uncertain | Jagged frontier; coherence versus correctness; moving boundary | Frontier prediction on each task | Rationale cites task characteristics rather than personal enthusiasm |
| Design a representative comparison test for one task | Baseline, representative case, acceptance criteria, failure mode | Test-plan canvas and demo audit | Test includes baseline, 3–5 cases, quality metric, and stop rule |
| Assign an appropriate AI–human operating mode and guardrail | AI-led, human–AI, human-led; verification and escalation | Mode decision matrix and peer challenge | Role assignment includes accountable human, verification source, and escalation condition |
| Commit to a seven-day experiment that can produce a go, revise, or stop decision | Transfer, iteration, monitoring | Action-plan card | Named owner, first case, metric, review date, and decision rule |
The deck is not the product. Improved decision capability is the product.
05 — The framework
A repeatable, five-step decision procedure a leader can run without a facilitator present. Select a step to see the guiding question, the participant action, the evidence produced, and the most common mistake.
Monitor returns to Decompose: the frontier moves when models, data, staff, or regulation change, so a passing test is a dated result rather than a permanent classification.
Step 1 of 5 · Decompose
Break the workflow into observable tasks. Name the inputs, outputs, decisions, owners, handoffs, and consequences.
The decision that matters most is not which tool to buy. It is who holds authority, on what evidence, and under what condition the work stops.
Strong repeated evidence; errors are detectable, reversible, and low-consequence.
Mixed evidence or meaningful context; independent human verification required.
Weak evidence or high-stakes, novel, regulated, difficult-to-verify, or irreversible consequences.
Every mode requires a named accountable human, an independent verification source, an escalation condition, and a stop rule.
Candidate-created instructional translation. Requires HBS/faculty validation.
06 — The workshop
Learn → see → do → reflect → apply → commit
Over half the session is participant work. The design sequence deliberately places prediction before evidence, and evidence before practice on the participant's own workflow.
0–5min
Learn / predict
“A faster answer can still be the wrong answer.” Individual vote and paired rationale make prior belief visible.
Participant actionVote on whether a polished, faster AI recommendation should be deployed.
Workshop deck
5–12min
Learn
Study design, inside-frontier gains, outside-frontier losses, and stated limitations.
Participant actionIdentify what makes a polished but incorrect output dangerous in their domain.
Workshop deckParticipant research brief
12–20min
See / reflect
The jagged frontier and the Frontier Fit Loop, applied to a short familiar example.
Participant actionClassify one task as promising, uncertain, or restricted and document why.
Workshop deckWorkbook
20–30min
See / do
Northstar versus Harbor. Structured metrics favor Northstar; a qualitative constraint reveals that the critical distribution agreement cannot transfer. The AI produces a polished Northstar recommendation from the visible metrics.
Participant actionChoose independently, inspect the AI output, rate confidence, uncover the missing constraint, and diagnose which Frontier Fit step failed.
Workshop deckFacilitator guidePrepared demo inputsBackup AI output
30–45min
Do
Participants apply the Frontier Fit Map to a real workflow: outcome, baseline, observable tasks, inputs, repeated decisions, human judgment, failure consequences, AI-fit hypothesis, representative cases, quality rubric, stop rule, operating mode, accountable human, escalation condition, success metric, and seven-day action.
Participant actionBuild a Frontier Fit Map and test plan for one consequential, repeatable task.
Participant workbookWorkshop deckFacilitator guide
45–53min
Reflect
Participants exchange maps and challenge assumptions: Where could a polished error survive? Is the test representative? Is the metric meaningful? Is the action reversible? Where should the human checkpoint move?
Participant actionRevise the decision after peer challenge.
Workbook peer-review pageWorkshop deck
53–58min
Apply
Participants convert the map into a small, reversible, decision-producing experiment.
Participant actionSet owner, first case, metric, review date, and decision rule.
WorkbookExecutive card
58–60min
Commit
“In the next seven days, I will test ______ before allowing AI to ______.”
Participant actionState the commitment and complete the post-assessment.
Executive cardWorkbookPost-performance assessment
Thirteen of the twenty-four slides, with the instructional reason each one is in the deck.
Surfaces prior belief before evidence arrives, so participants can later observe their own judgment changing.
Stage 1
A fictional company must choose a launch market. Structured metrics favour Northstar over Harbor. Participants decide independently, on the record, before any AI output appears.
Stage 2
Given only the visible metrics, the AI produces a fluent, well-structured, confident recommendation for Northstar. Participants rate their confidence in it.
Stage 3
A qualitative constraint shows the critical distribution agreement cannot transfer. The revised decision is a bounded feasibility test — and participants diagnose which Frontier Fit step failed.
The demonstration is not designed to embarrass the technology. It is designed so that overreliance is experienced rather than described — the AI is not wrong because it is weak, but because the evidence it was given was incomplete.
The company, market data, and constraint are fictional teaching content created for this prototype. A prepared backup output is included in case live AI access fails.
07 — Participant experience
Participants do not practise on a case study. They map a real workflow they own, design a test for it, assign authority, and leave with a seven-day experiment.
The completed example is a fictional escalation-triage workflow, filled in to the standard the rubric rewards.
Field 01
Name one consequential, repeatable workflow you own.
Field 02
What would improve, and for whom?
Field 03
List 4–7 tasks with inputs, outputs, and owners.
Field 04
Where does judgment actually happen?
Field 05
What happens if the task is done badly?
Field 06
Promising, uncertain, or restricted — and why?
Field 07
Choose 3–5 cases including at least one edge case.
Field 08
How will output be scored, and by whom?
Field 09
What would make you stop the pilot immediately?
Field 10
AI-led, human–AI, or human-led?
Field 11
What result would justify a wider pilot?
Field 12
Seven-day action, owner, review date, go/revise/stop rule.
Fictional teaching example. Assessment thresholds are prototype judgments requiring faculty validation.
The experience is a supported system rather than a single event: the facilitator, the participant, and the learning team each hold a defined set of assets at every stage.
Before the session
During the session
After the session
08 — Measurement
Performance is measured before, during, immediately after, and later — against decision quality rather than activity.
Five criteria, scored 0–3. A workbook that names a department rather than a task, or a success metric rather than a stop rule, cannot reach the top band.
| Criterion | 0 points | 1 point | 2 points | 3 points |
|---|---|---|---|---|
| Task decomposition | Job/department named | One broad activity | Multiple tasks but unclear boundaries | 4–7 observable tasks with inputs, outputs, decisions, owners |
| Frontier hypothesis | No rationale | Preference-based rationale | Some task cues | Evidence-based rationale with uncertainty explicitly stated |
| Test design | Demo only | Unrepresentative case or no baseline | Baseline plus typical cases | Same rubric, credible baseline, 3–5 representative cases including edge case |
| Authority and guardrail | “Human checks” | Human named but check unclear | Evidence source or escalation named | Accountable human, independent evidence, escalation and stop rule |
| Metrics and action | Vanity/activity metric | Time or volume only | Quality and time | Quality/correctness, time, failure severity, owner, review date, decision rule |
Prototype thresholds. Institutional assessment standards require faculty and program validation.
Satisfaction and self-reported confidence are worth collecting — but neither demonstrates that a decision improved.
Illustrative data created to demonstrate the analysis and revision process. No cohort has been delivered.
09 — AI-enabled production
The same discipline the workshop teaches executives governs how the workshop itself was produced: AI holds speed, humans hold judgment and authority.
Speed is worthless if the research is misrepresented. Eight controls sit in the workflow as hard requirements, not preferences.
The prompt library included in the package is safeguarded: it excludes confidential data and cannot be used to generate citations.
10 — Inclusive design
Executive audiences differ in technical background, authority, bandwidth, confidentiality constraints, and comfort speaking in a room. Each decision below removes a barrier to participating fully.
This website applies the same standards: semantic structure, keyboard-accessible controls, contrast that does not depend on color alone, reduced-motion support, and downloadable alternatives to every interactive element.
11 — The artifacts
Every asset exists because a specific person needs it at a specific moment: the facilitator before and during delivery, the participant during and after, and the learning team afterwards.
Independent candidate prototype created using publicly available HBS research. Not an official Harvard Business School product.
PPTX · 24 slides
Facilitator, projected to participants
The primary facilitation deck for the 60-minute session: provocation, study design, inside- and outside-frontier findings, limitations, the Frontier Fit Loop, live demonstration, participant exercise, peer challenge, debrief, and seven-day commitment. Includes facilitator notes and per-slide source blocks.
Facilitation assetsResearch fidelityExperiential sequencing
PDF · 19 pages
Facilitator, learning-operations team
Makes the experience deliverable and repeatable: minute-by-minute plan, room and virtual setup, SAY/ASK/DO/WATCH FOR/DEBRIEF language, transitions, expected responses, misconceptions, difficult questions, technology contingencies, accessibility considerations, assessments, follow-up system, production SOP, and safeguarded prompt library.
Delivery supportAI-enabled productionInclusive deliveryVersion control
PDF · 14 pages
Participants (executives and managers)
The active participant tool — not a printed copy of the slides. Participants select a real workflow, decompose it into observable tasks, predict AI fit, design a representative test, assign authority, define a stop rule, critique a peer's plan, and build a seven-day experiment. Includes a completed B2B example and the application rubric.
Participant-centered formatsExperiential learningAssessment design
PDF · 1 page
Participants after the session
The one-page takeaway: the Frontier Fit Loop, AI-led / human–AI / human-led operating modes, signals that should not justify scaling, and the participant's seven-day experiment.
Transfer designParticipant-centered formats
PPTX · 7 slides
Interview panel
A seven-slide, five-to-seven-minute overview of the challenge, selected research, learning objective, participant exercise, AI production system, measurement strategy, and faculty-validation plan. This is not the participant workshop.
Stakeholder communicationNarrative clarity
PDF · 39 pages
Hiring manager, faculty reviewer, instructional-design leader
The intellectual source of truth: role analysis, research landscape and scoring, faculty research brief, personas, objectives, learning architecture, workshop specification, assessment and feedback systems, AI-enabled production workflow, SOP and prompt library, accessibility review, source ledger, and quality audit.
Research translationBackward designQuality assuranceSource discipline
Independent prototype artifacts. Not Harvard Business School materials and not endorsed, reviewed, or approved by HBS or the HBS AI Institute. Faculty validation would be required before any institutional use.
12 — Honesty and validation
Naming the boundary is part of the discipline this project teaches. A polished document is not evidence, and an instructional translation is not a research finding.
Nothing in this project is a Harvard Business School product. It is an independent demonstration of a process — and the process includes knowing which claims are not yet mine to make.
Twelve capabilities named in the role specification, and where each one is demonstrated in this project.
| Capability | Evidence |
|---|---|
| Faculty and SME partnership | Faculty research brief, approval gates, validation questions |
| End-to-end learning development | Research intake through delivery, assessment, revision, and archive |
| Research translation | Study findings transformed into an executive decision problem and experiential framework |
| Experiential learning | Real-work Frontier Fit Map, live demonstration, peer challenge, seven-day experiment |
| Facilitation assets | Complete deck, speaker notes, facilitator guide, misconceptions, contingencies |
| Participant-centered formats | Workbook, executive card, digital companion specification |
| Digital and multimedia learning | Video script, infographic architecture, interactive companion specification |
| AI-enabled production | Workflow, human/AI responsibility matrix, SOP, safeguarded prompt library |
| Measurement | Baseline, formative checks, post-performance rubric, transfer measures |
| Continuous improvement | Feedback instrument, analysis pipeline, revision example, version history |
| Quality assurance | Claim ledger, source blocks, accessibility audit, slide-overflow test, release controls |
| In-person and virtual readiness | Room and virtual setup, low-tech fallback, backup demo, inclusive participation |
13 — About this project
I am Amir Daniel, an instructional designer working at the intersection of research translation, experiential learning, and AI-enabled production. I built this project independently, in a few focused hours, because the most honest way to show how I would approach the HBS AI Institute Instructional Designer role was to actually do the work — select the research, design the experience, produce every supporting asset, define the measurement, and mark the limits.
The judgment I want to demonstrate is not that I can generate documents quickly. It is that I can decide what a study does and does not support, design a capability worth building, and keep human authority over every claim an executive audience would carry back into their organisation.