The Promise and the Problem
Every other week, a new app launches claiming to use AI to plan your life. The pitch is always the same: tell the AI your goal, and it will create a plan for you. The reality is almost always disappointing. You get a vague list of steps that could apply to anyone, with no real structure, no timeline awareness, and no connection to the rest of your life.
The problem is not that AI cannot do this. It is that most products use AI in the laziest possible way. They send a single prompt to a language model and return whatever comes back. That is not a planning system. That is a chatbot with a save button.
This article explains how ArcusVision's AI roadmap generation pipeline actually works, why structured output matters more than raw model capability, and what separates useful AI planning from generic list generation.
Step 1: Context Collection, Not Just a Text Box
The first thing most AI planners get wrong is the input. They give you a text box, you type "I want to learn machine learning," and the AI generates a plan. The problem is that a good plan for a computer science graduate is completely different from a good plan for a marketing manager. Without context, the AI has to guess, and it guesses by producing the most generic possible output.
ArcusVision takes a different approach. When you create a goal, you assign it to one of six life pillars: Ambition, Foundation, Intellect, Vitality, Wealth, or Social. Each pillar has its own set of context questions that the system asks before generating anything.
For an Ambition goal like "transition to a senior engineering role," the system asks about your current role, years of experience, target timeline, specific areas of interest, and known gaps. For a Vitality goal like "run a half marathon," it asks about your current fitness level, any injuries, available training time, and whether you have run before.
This context collection is not optional decoration. It is the foundation that makes the output useful instead of generic.
Step 2: Pillar-Specific Prompt Engineering
Here is where most AI planning tools completely fall apart. They use a single, generic prompt template for every type of goal. Something like: "Create a step-by-step plan for the following goal: {goal}. Include milestones and a timeline."
That prompt will produce output. It will even look reasonable at first glance. But it will lack the domain-specific structure that makes a plan actually actionable.
ArcusVision uses pillar-specific prompt chains. The prompt template for an Intellect goal (like learning a new programming language) is fundamentally different from the prompt template for a Wealth goal (like building an emergency fund) or a Vitality goal (like establishing a daily exercise routine).
Each prompt template includes:
- Phase structure requirements: The AI must organize the plan into sequential phases with clear entry and exit criteria.
- Task granularity rules: Tasks must be specific enough to complete in a single work session (typically 30 to 90 minutes), not vague multi-day activities.
- Dependency mapping: The prompt requires the AI to identify which tasks depend on others and cannot be started until prerequisites are complete.
- Resource linking: Where applicable, the AI includes specific resources like courses, documentation, tools, or communities rather than generic advice like "find a good tutorial."
- Hour estimation: Each task includes an estimated time commitment, which enables deadline validation.
Step 3: Structured Output Parsing
Sending a good prompt is only half the battle. The other half is getting structured, parseable output back from the model.
ArcusVision uses purpose-built AI models optimized for planning. The models support structured output modes, but the key is not just requesting JSON. It is defining a strict schema that the model must conform to.
The output schema includes:
- An array of phases, each with a name, description, and estimated duration.
- An array of tasks within each phase, each with a title, description, estimated hours, priority level, and optional resource URLs.
- A dependency graph that maps task relationships.
- Milestone markers that indicate significant progress points.
If validation fails, the system retries with the fallback model before presenting an error. In practice, the retry rate is low because the prompt engineering and schema constraints are tight enough to produce valid output on the first attempt in the vast majority of cases.
Step 4: Deadline Validation
This is something almost no other AI planner does. Once the tasks and hour estimates are generated, ArcusVision checks whether the plan is actually completable within the user's target timeline.
The system takes the total estimated hours across all tasks and phases, factors in the user's available time per week (derived from their time block schedule if they have set one up, or from a default assumption), and calculates whether the math works.
If a user wants to "learn data science in 3 months" but the generated plan requires 400 hours of work and the user has 10 hours per week available, that is 40 weeks of work compressed into 12. The system flags this and presents the user with options: extend the timeline, reduce the scope, or increase available hours.
This sounds simple, but it is the difference between a plan that exists on paper and a plan that is actually achievable. Most AI planners will happily generate a 3-month plan that would take a year to complete, because they never do the math.
Why Most AI Planners Fail
Having built this system, we have a clear picture of why most AI planning tools produce disappointing results:
Single-prompt architecture: They send one prompt and return one response. No context collection, no pillar-specific templates, no validation layer. The output quality is entirely dependent on whatever the user typed into the text box. No structured output: They return free-form text that looks like a plan but cannot be parsed into actionable tasks with dates, priorities, and dependencies. You get a nice-looking list that you still have to manually break down. No timeline awareness: They generate plans with no concept of how long things take or whether the user has enough time. A plan without time validation is just a wish list. No integration with execution: Even when the plan is decent, it lives in a separate space from where work actually happens. You get a plan in one tool and then have to manually copy tasks into your actual task manager. Generic prompts: They use the same prompt structure for "learn Spanish" and "prepare for a triathlon" and "build a SaaS business." These are fundamentally different types of goals that require fundamentally different planning structures.The Model Matters Less Than You Think
One of the most counterintuitive lessons from building this system is that the choice of language model matters less than the prompt engineering and output validation pipeline around it.
The AI models ArcusVision uses are not the most powerful models available. They are fast and cost-effective. But with the right prompt structure, context, and output constraints, they produce better plans than a more expensive model with a generic prompt.
This is because planning quality is bounded by input quality and output structure, not by raw model intelligence. A smarter model with bad input and no structure will produce more eloquent garbage. A capable-enough model with great input and strict structure will produce actionable plans.
What This Means for Users
If you are evaluating AI planning tools, here are the questions to ask:
- Does it ask you context questions before generating a plan? If it just has a text box and a "Generate" button, the output will be generic.
- Are the generated tasks specific and time-bounded? If the tasks are things like "learn the basics" or "practice regularly," the AI is not doing the hard work of breaking things down.
- Does it validate the timeline? If you can set any deadline and the AI happily generates a plan without checking feasibility, the plan is decorative.
- Does the plan connect to execution? If the generated plan lives in a separate view from where you actually track and complete tasks, you will lose the thread within a week.
- Does it handle different types of goals differently? A fitness goal and a career goal require completely different planning structures. One-size-fits-all prompts produce one-size-fits-none plans.