What the plan is for
The exposure map tells you what AI's impact is on each task in your week. The plan answers the next question: what to do about it.
Take a Product Manager at Mid-level IC. The map says that a task like Synthesize interview findings is already work AI can do most of, and that a task like Prioritize the backlog is still done mostly by the person. Those are two different situations, and getting better means something different in each. On the first, the skill worth building is directing AI well, so that the draft it produces reflects the decisions you made and not ones it made for you. On the second, the skill worth building is your own judgment, because AI cannot make that call for you, and as AI speeds up the rest of the task, that judgment becomes most of what the task is.
The plan looks at every task in your job at your level, picks the eight where the map's ratings say there is the most to gain from getting better, and writes each one up as a development opportunity. For Synthesize interview findings the development opportunity is “Prompt AI to synthesize interviews, then catch what it smoothed over”, it builds the skill of critiquing and redirecting, and it says in one sentence what your work becomes: “AI could draft the cross-interview synthesis and a first set of proposed implications. Your work becomes stress-testing that draft against what you actually heard, catching where AI smoothed over a contradiction, and building the case you can defend when someone pushes back.” Under that, the plan shows three ways to build the skill, so you can start this week. The aim is to take the guesswork out of where to start, and out of what to do next once you have started.
How the eight development opportunities are chosen
The eight are chosen by a fixed rule from the same ratings the map shows. Every task in your job at your level is considered and scored on how much there is to gain from getting better at it, and the tasks come in two kinds. Get better at the tasks that stay human marks work that is still done mostly by a person, for a reason that will not weaken as AI improves. That is a strength to build on. Get better at directing AI marks work that AI already does most of, and does well, so the skill to build is directing the AI. The rule takes the strongest of each kind in turn, so a plan is never all one kind. That matters most at senior levels, where nearly every task is still done mostly by a person: without the second kind, a director's plan would be eight ways to protect strengths and none to put AI to work. The eight chosen are the tasks that carry a chip on the map, and the chip is the kind.
Once the eight are picked, they are put in order by the date the map gives each task, which is when working on it with AI becomes common practice. The soonest goes first and the furthest off goes last. In the Product Manager plan at Mid-level IC, the first development opportunity is on a task the map dates “in 1–2 years” and the last on one dated “in 2–4 years”. So the order on the plan is a suggestion for what to start on, not a ranking of how much each one matters, and you can work through them in any order you like.
Each development opportunity is one of three kinds, and the rating decides which. Where the task is still done mostly by a person, the opportunity is to get better at the part only a person can do, and on the map that task carries Get better at the tasks that stay human. Where AI already does most of the task, and does it well, the opportunity is to learn to direct the AI, and the task carries Get better at directing AI. The third kind is also work AI does most of, so if the map gave it a chip, it is the same chip. It applies where the way the task is done is changing for the whole team: the opportunity is to work out the new way of doing the task and set it as the standard your team follows. In the Product Manager plan at Mid-level IC, “Hold your prioritization call under real pushback from engineering” (Prioritize the backlog) is the first kind and “Prompt AI to synthesize interviews, then catch what it smoothed over” (Synthesize interview findings) is the second. At Senior IC, “Design the reusable launch template for a brand-new launch type” (Draft the launch checklist) is the third kind, and the plan says what your work becomes: “AI could draft a checklist from the closest precedent, but there is none for a genuinely new launch type. Your work becomes designing the template itself: deciding what steps a launch like this actually needs, and turning that judgment into the standard the rest of the team reuses.”
How a development opportunity is written
An AI writes each development opportunity, and it works from a fixed set of facts rather than from its own general knowledge. For each of the eight tasks it is given the task and how someone at your level does it; the rating, which says how much of the task AI can do today, when working that way becomes common practice, and why a person is still needed; the sentence about what the work becomes; which of the three kinds this is; how much of your week the task takes; and the skills already named for the task. From those it writes the headline and the three ways to build the skill. The paragraph under the headline is written separately, from the three things the map already says about the task at your level: what the task is at your level, what your work becomes, and why it is worth investing in. It says those once, in its own words, and the page goes straight to the practical part.
Take “Prompt AI to synthesize interviews, then catch what it smoothed over” again. For Synthesize interview findings at Mid-level IC, the paragraph begins: “Reviewers will test the implications you take from your area's interviews, and you are the one who has to defend them.”
Every development opportunity also names the one skill it builds and the other skills you pick up along the way. For “Prompt AI to synthesize interviews, then catch what it smoothed over” the skill is critiquing and redirecting, and along the way synthesizing complex information, defending decisions to stakeholders. The skill names come from one shared list that every job on the site draws from.
Examples based on your experience with AI
Before it shows you how to build a skill, the plan asks one question: how much do you already use AI? The three answers are “New to AI”, “I use it most days” and “Advanced AI user”, and every development opportunity is written three times, once for each answer.
On “Prompt AI to synthesize interviews, then catch what it smoothed over”, the first thing the plan asks someone new to AI to try at work is this: “Paste your interview notes into an assistant and ask for a themed synthesis with proposed implications, then read it against your memory of the calls and mark anything that feels too clean.” It asks an advanced AI user something else entirely: “Build a synthesis prompt that ingests multiple transcripts and returns themes, proposed implications, and a separate list of contradictions it had to reconcile, so smoothing-over becomes visible by default.”
These are not the same advice with easier or harder words. For someone new to AI, the work is one prompt, one result, and one judgment about whether the result is any good. For someone who uses it most days, the work goes in rounds: ask for a draft, say what is wrong with it, ask again, and compare the two. For an advanced AI user, the work is to build something reusable, such as a template or a set of instructions, so that the next time the task comes round most of it is already done. Pick the answer that sounds like you. You can switch at any time, and the plan keeps a separate set of ticks for each answer.
Within each version there are three ways to work, and none depends on the others. Practice prompts are short exercises: prompts you paste into any AI assistant to practise on, away from your real work. Try this at work is two or three changes to how you already do the task, and any one of them is worth trying on its own the next time the task comes round. The project is one real piece of work, done in order over two or three hours, some of it at your desk because it uses your own material. Each prompt sits inside the part of the project that uses it, and the project says what you will have at the end and how you will know it worked.
The links, and what “checked” means
Under every development opportunity, “If you want to go deeper” lists up to four things to read or watch. Under “Prompt AI to synthesize interviews, then catch what it smoothed over” there are four, the first of them “Claude Code + 50 interview transcripts: three ways I query the data” from Else van der Berg, Substack. They were found by an AI searching the web, task by task, with a set number of searches to spend, and every link is checked the same way. It must be free to read. It should be about doing this task with AI, not about AI in general, and published within the last twelve months. It should come from someone who does the work, rather than from a company selling a product: at most one link per task may be a vendor's own training or documentation. Where fewer than three links meet every test, the plan adds the next best, such as a useful article with no date on it, or a broader article with one good section on the task, and the note under that link says which part to read. The date on each card is the day the link was opened and checked against those tests. Across the site today there are 532 links over 277 tasks, 92 of them from vendors, all checked between 2026-09-08 and 2026-09-27.
The links belong to the task, not the level. A director and a junior analyst working on the same task see the same reading list, because what is written about doing a task with AI rarely differs by seniority.
Who to follow
At the end of the plan there is a short list for the job as a whole: people, newsletters and podcasts about doing this job with AI, all free and all active in the last twelve months. An AI searches the web for them, job by job, and checks each one. The list is the same whichever development opportunity or level you are on. For Product Manager it is four people, two newsletters and two podcasts.
What the plan keeps about you
Your ticks, your “marked complete” and your answer to the AI-experience question are saved in your own browser and never sent anywhere. There is no account, and we hold nothing that identifies you.
The site does keep a count of what gets used, with nothing in it that points to you: that a plan was opened for a job and a level, which development opportunity was opened, which AI-experience answer was chosen, that a box was ticked or something marked complete, that a prompt was copied, that a link was clicked, and that the reader went back to the map. That is how we learn which parts of the plan get used and which never do, and it is the only reason those records exist.
Catalog version pilot-trackA-v2, published 2026-09-07. None of the plan is written by hand. When a line reads wrong, we fix the rule that produced it and write the plan again for every job, so the same rule applies everywhere. How the map is made.