AI Project Management Simulator
Breaking the Illusion of a Safe Schedule
The Business Challenge
Unclear task interdependencies are a systemic driver of project failure. According to PMI’s 2026 Pulse of the Profession report, 31% of complex projects fail to achieve their originally intended benefits—more than double the rate from just two years prior.
Within that broader pattern, industry practitioner literature points to a specific, recurring failure mode: schedule errors that get visually corrected on the surface while the underlying structural delays cascade invisibly, pushing campaigns past their deadlines. This simulation is a self-directed spec project built to target exactly that failure mode—the client and campaign are illustrative.
Industry Sources:
PMI, Pulse of the Profession 2026: Driving Success in Complex Projects
PMI Press Release: Why Complex Projects Fail
The Digital Project Manager: Task Dependencies & Schedule Slippage
Target Audience & Needs Assessment
This simulation targets mid-level marketing project managers and program coordinators. This persona was synthesized from PMI industry data to model a widespread capability gap: professionals who own cross-functional Gantt charts (in tools like Asana or Monday) but lack formal PMP certification or dependency-modelling training.
Instructional Logic: The Trap and The Collision
In this simulation, a campaign schedule relies on a rigid sequence with a strict deadline of Day 8.
The Trap: The interface opens on a deceptive state. The "Legal" phase is set to a Start-to-Start dependency with zero lag. Visually, the chart looks neutral and mathematically safe (landing on Day 5). Functionally, it is impossible: Legal cannot review design work that hasn't started yet.
The Collision: If the user simply switches the dependency to Finish-to-Start without properly calculating a trailing buffer (lag), they trigger a collision. The schedule updates, Launch pushes to Day 10, and the campaign misses its deadline.
The Solution: Custom Socratic AI Application
A traditional authoring tool requires pre-scripted multiple-choice options (e.g., "Select a lag: 0, -1, -2"). That removes the cognitive load of actually calculating the schedule and diagnosing the error.
To evaluate a learner's free-text rationale at scale, I built a custom, single-screen web application featuring a stateless Socratic AI coach (Gemini 3.5 Flash-Lite). Instead of clicking through a quiz, learners must manipulate a live Gantt chart to find the mathematically tightest feasible schedule, and then defend their logic against the real-time math—without the AI ever handing them the answer.
Technical Architecture & Guardrails
To ensure security, cost-efficiency, and instructional integrity, I designed a custom, lightweight web application:
Frontend: A single-screen UI built with static HTML/CSS/JS, hosted on GitHub Pages.
Backend & Security: A Cloudflare Worker acts as a proxy, safely storing the API key as a secret and enforcing a strict rate limit (10 requests/hour/IP) to prevent abuse.
The AI Model: Gemini 3.5 Flash-Lite was selected specifically for its low latency and high cost-efficiency on classification and short-response tasks.
Stateless Evaluation: The AI evaluates each submission statelessly, meaning there is no conversational memory between calls. This simplifies the architecture and effectively prevents users from trying to prompt-inject or manipulate the AI into giving away the answer.
Before generating the code, I mapped the interaction sequence in Affinity to ensure the learning flow—moving the learner from a deceptive trap to an earned success—was mathematically and visually sound.
The Learning Sequence: From Trap to Success
Image 1 (The Trap): The simulation opens on the illusion of safety. The Legal phase is set to Start-to-Start with zero lag. Visually, it looks mathematically safe (Day 5); functionally, Legal cannot review work that hasn't started.
Image 2 (The Collision): Fixing the surface error breaks the deadline. When the learner changes the dependency to Finish-to-Start without a trailing buffer, Launch pushes to Day 10. The campaign fails.
Image 3 (The Socratic Coach): The AI coach evaluates the learner's free-text reasoning against the live math. A strict negative prompt bans the AI from using terms like "lag," forcing it to guide the learner rather than hand over the answer.
Image 4 (The Success State): The learner earns the success state by setting a Finish-to-Start dependency with a -2 lag. The schedule hits the strict Day 8 deadline, and the AI validates their mathematical logic.
Quality Assurance: Managing AI-Assisted Development
This custom web application was implemented using an AI coding assistant (Claude Code) under my direct supervision. As the instructional architect, I drove the pedagogical logic, engineered the system prompts, and conducted rigorous adversarial testing to protect the validity of the assessment.
Key edge cases resolved during development include:
Closing the Reasoning-Validation Gap: Initially, the system prompt only checked the mathematical outcome, allowing learners to pass if they submitted the correct lag value but typed gibberish. I resolved this by engineering a two-step evaluation prompt: Step 1 validates that the text is substantive, and Step 2 evaluates the schedule math.
Preventing Data Confusion (Constants vs. Variables): During testing, the AI hallucinated that the user's computed launch day was "Day 8" when it was actually Day 5. The LLM was confusing the fixed deadline constant (Day 8) with the calculated variable. I corrected this by refining the prompt instructions to rigidly distinguish between the two.
Enforcing Socratic Guardrails: To prevent the AI from giving away the answer, I initially banned terms like "lag" and "fast-tracking" in the evaluation step. However, if the AI declined a submission, it might reply, "I cannot give you the lag value," inadvertently leaking the term. I fixed this by applying the negative word-constraint globally across all AI outputs.
Structured JSON for Attempt Tracking: The frontend allows exactly three genuine attempts before locking. Initially, AI "declines" (asking the user to rephrase a non-answer) were incorrectly burning through these attempts. I redesigned the system prompt to output a strict declined: true/false boolean in JSON format, providing the frontend with an authoritative data point to control the session lock.
Catching a Hidden Data-Integrity Bug: I found a logic error where applying a negative lag to a Start-to-Start dependency caused the visual chart to stop at Day 1. It looked correct on-screen, but the hidden math behind it was corrupted. I directed the technical fix to correct the underlying data, ensuring the AI coach graded the learner against accurate schedule math, not just a screen that "looked right."
Measurement & Expected Business Impact
To validate the business impact of this tool in a live corporate deployment, success would be measured against Kirkpatrick Levels 3 and 4:
Level 3 (On-the-Job Application): Conduct spot-audits on active marketing campaigns in Asana or Monday at 30 and 60 days post-training. The key metric is tracking the reduction of zero-lag Start-to-Start dependencies on critical handoffs (e.g., Design to Legal), measuring whether project managers are actively building mathematically sound trailing buffers into their schedules.
Level 4 (Business Results): Correlate the corrected scheduling behaviors directly with the percentage of marketing campaigns that hit their locked launch dates without requiring emergency schedule compression.