- The four inputs you need before any prompt
- Stage 1: Diagnose before you design
- Stage 2: Build the path against the diagnosis
- Stage 3: Match the path to the platform you already pay for
- The counterintuitive thing about personalization at scale
- The failure mode: paths that never close
- What changes for the L&D team itself
LinkedIn’s 2025 Workplace Learning Report shows around 47% of L&D teams now have generative AI fully embedded in their L&D strategy, with personalised learning pathways and content authoring leading the use cases. The honest reading of that number is that nearly half of teams have crossed from experimentation into operations. The other half are still building the inputs.
Personalization at scale only works when the data inputs are real. The most common failure is to ask an AI tool to “personalise training for our team” and expect anything useful back. The model has nothing to personalise to. This post walks through the four data inputs that make the chain work, the prompt sequence that turns those inputs into a defensible per-learner pathway, and the tools that actually run it in production.
The four inputs you need before any prompt
A personalised learning path produced from thin inputs is just a generic learning path with a name attached. The four inputs that change that:
The role context. Not the title, the actual scope: what the person is responsible for delivering, who they answer to, what success looks like in 90 days. One paragraph per role.
The current skill level, in their words. A short self-assessment beats a manager rating, every time. Three to five sentences from the learner about what they currently do confidently and where they hit walls.
The development goal. One sentence the learner wrote: “I want to be running my first cross-functional project by Q4.” Goals manufactured by HR have lower follow-through than goals stated by the learner.
The constraints. Hours per week available for learning, preferred formats (video, written, interactive), and any blockers (timezone, accessibility, language).
These four inputs together are around 300 to 400 words per learner. Collected once, refreshed twice a year, they are the substrate that turns AI from a content generator into a personalisation engine.
Stage 1: Diagnose before you design
Drop the four inputs into Claude or ChatGPT and ask for a diagnosis, not a plan:
“You are an experienced L&D consultant. Below is a profile for one employee. Diagnose, do not prescribe yet. Identify the three skill gaps that, if closed, would have the biggest impact on this person’s stated goal in the next six months. For each gap, explain why it matters specifically for the role described, what the early signs of progress would look like, and what other capabilities it makes possible downstream. Do not yet recommend any courses or content. PROFILE: [PASTE THE FOUR INPUTS].”
The diagnostic step is what separates a personalised path from a curated playlist. Skipping it produces a list of courses that look thoughtful and address none of the things actually blocking the learner.
Stage 2: Build the path against the diagnosis
Once the diagnosis is named, produce the path:
“Using the diagnosis above and the constraints in the original profile, design a 12-week learning path for this employee. For each week, specify: the single capability being built, the two best content formats given their stated preferences and weekly time budget, one practice task they will apply in their actual job that week, and the visible signal a manager would see if the capability is taking hold. The path should escalate in difficulty across the 12 weeks. Output as a table with columns: week, capability, content format, application task, manager-visible signal. Do not include filler weeks.”
The “manager-visible signal” column is the part most L&D paths miss. It turns a learning path from something that lives in the LMS into something that shows up in the work. It also gives you the metric that lets you prove the path worked.
Stage 3: Match the path to the platform you already pay for
Now and only now, match the path to actual content. Sana and Cornerstone run AI-driven path recommendations against your existing course library. Articulate Rise authors and delivers the modules. Synthesia and HeyGen generate the video components without filming.
“For each week in the path above, recommend the format I should source: existing internal content (which type), a public course (suggest providers without inventing course names), a short AI-generated video (specify the script’s main argument and visual style), or a manager-led practice session (specify the conversation prompt). For any AI-generated video, write a 90-second script that opens with the application context, teaches one concept, and closes with the practice task for the week.”
Feed those video scripts into Synthesia or HeyGen, generate them with a presenter avatar in your brand voice, and you have weekly video content per learner without booking a studio.
The counterintuitive thing about personalization at scale
The teams getting employees to genuinely love their pathways are not personalising more, they are personalising the right two things. They personalise the application task (what the learner does in their actual job this week) and the manager-visible signal (what their boss sees them doing differently). They do not personalise the underlying content as much as you would expect. A 30-person cohort might share roughly the same six course modules, with personalised application tasks, personalised pacing, and personalised manager check-in prompts wrapped around them.
This is the opposite of how most L&D teams approach the problem. They invest heavily in custom content per learner and lightly in custom application. The learner experience is “I watched a video that nobody else watched, and then I went back to my normal job and forgot it.” Reverse the ratio. Shared content, personalised practice, personalised feedback. The path feels personal because the work feels personal, even when the underlying modules are shared.
The failure mode: paths that never close
The single most common reason employees stop loving their personalised path is silent path-rot. The pathway was set up beautifully in week one, the manager check-in happened in week two, the application task was tracked in week three, and then the system went quiet. By week eight, the learner has stopped logging in. The path is technically still active in the LMS. Nothing has actually happened in eight weeks.
The fix is mechanical. Build a weekly check-in prompt into the path itself, automated through your platform of choice:
“Generate a weekly two-question check-in for the employee in this path. Question one asks them to describe in one sentence the application task they completed this week and what surprised them. Question two asks them what blocked them, if anything. Output should be 80 words or less, conversational, and end with a single specific commitment for next week. Adapt the tone based on whether the previous week’s task was completed or not. Today’s week is [WEEK NUMBER] of 12.”
The check-in is short enough to answer in 90 seconds. The aggregate of those check-ins is the dataset you use to refine the next cohort’s path. Path-rot happens when the loop is open. Close the loop, and a 12-week path actually finishes.
What changes for the L&D team itself
The team that runs personalised pathways at scale spends less time authoring content and more time designing application tasks and reviewing check-in patterns. The role shifts from instructional designer to learning architect. The 47% of L&D teams already operating this way are not better at content design. They are better at deciding which parts of the path to personalise, which to share, and where in the chain AI is the right tool for the job. The teams employees actively look forward to learning from are the ones who got that split right.