The 90-day plan and the job operating system

Days 1–30: listen, map the current AI work, find the golden-set situation (usually: none), pick one thin slice with a named owner, publish a scoring rubric. 31–60: shadow the slice, first eval bar, data-handling note, no platform rewrite. 61–90: assist-mode users, a steering paper with kill/scale, a second slice or a documented no. You will not "transform the enterprise" in 90 days. Anyone who says they will is lying. You will install an operating system that can fund and kill.

Write each block with a visible output, because interviewers ask "what will you have in 30 days." 30: a map of in-flight AI work (a table, not a workshop series), a scoring rubric, one named owner, one named slice, a written hole list. 60: a shadow running, an eval table with a bar, a data-handling one-pager Security can comment on, a death date on any inherited POC. 90: assist-mode or a documented kill, a steering paper, a recommendation on a second slice or on extraction-to-platform. Outputs, not activities. "Meet people" is an activity. "A map with owners" is an output.

Days 1–30 are mostly listening, and listening still needs a calendar. Week 1: sit in the existing standups, read the last three steering packs, ask who owns evals (nobody), ask what the approved model path is, ask which POCs have death dates (none). Week 2: draft the scoring rubric (value, feasibility including data/evals/ACL, risk, change) and socialize it as a draft, not as a decree. Week 3: pick the slice with a sponsor who will name an owner. Week 4: publish the map and the pick. That is a month. It is enough. It is not a platform.

Days 31–60 you will be pressured to demo. Use pushback A. Shadow first. The eval bar is the thing you add that the company probably does not have. Protect it. If they inherited a vendor, use drill C: golden set, death date, no write tools. If legal is fuzzy, use drill D in week 5, not in week 12. Early path-or-kill beats a quarter of quiet noncompliance. You are new. New people can ask naive questions. Use that window. It closes.

Days 61–90 you must show a decision, not a status. Scale, assist, or kill. A second slice only if the first produced a pattern worth copying. A platform recommendation only if two slices share ingest/eval/gateway. If the first slice missed the bar, the 90-day paper is a kill with evidence. That paper is a success. A 90-day paper that says "we are still exploring" is a failure, even if everyone was nice. You were hired to reduce a museum of POCs, not to add an exhibit.

Seat variants. FDE: 30 is a customer map and a first retrieve on their corpus; 60 is an eval with the customer; 90 is expand or kill on that account — you are not also running the AI office. Solutions: 30 is a qualification rubric and two discoveries; 60 is a POC on a golden set; 90 is a yes or a no that still leads. AI PM: 30 is a one-pager with evals; 60 is a shadow behind a feature flag; 90 is a ship-or-don't with a refusal policy. Transformation: 30 is intake freeze and a scoring model; 60 is two slices and a kill list; 90 is an operating cadence the exec team actually uses. Same spine, different nouns. Write yours.

You know how to run a pipeline. Stages: researched target, conversation, application, loop, offer. WIP limits. A weekly standup with yourself: 5 targeted applications or 5 conversations, not 40 spray. Update the pack when a loop teaches you something — that is your retrospective. Conversations outperform applications. Use the 90-word paragraph. Ask for 20 minutes, not a job.

WIP limits, written: no more than 3 live loops; no more than 5 conversations in flight; no more than 5 targeted applications in a week; no take-home that breaks the 6-hour cap if you already have a live loop. When a stage is full, you do not start another. You finish, kill, or park. This is the same rule you used on programs that tried to start eight workstreams. Spray-applying is eight workstreams. It feels like progress. It is inventory without owners.

Weekly cadence you can keep, not a hero week. Example: 4 hours. Monday 60 minutes — targets and one pack tweak from last loop. Tuesday 45 — two conversation asks (90-word paragraph, 20 minutes, not a job). Wednesday 45 — one application to a 5-scoring post, primary paper only. Thursday 60 — artifact upgrade or the one next skill, not both. Friday 30 — standup with yourself: WIP counts, what to kill, what to park. If you have a full-time job, this is already a lot. Do not write a 12-hour plan you will abandon on day 8 of the search.

Conversations: three this week to people who have seen you deliver. The message is positioning v1, 90 words, a specific ask ("20 minutes to walk whether [primary seat] is the right name for what I do"). Not "let me pick your brain." Not a referral form in sentence one. If they offer a referral, take it. If they offer 20 minutes, take it. If they ignore you, one bump in seven days, then park. You already know how to chase a stakeholder without becoming a pest. Use that skill. Volume of ignored InMails is not a metric. Conversations held is.

Applications: only to posts you would have scored 4–5 on day 22, with the primary paper, with a first-line mapping if the title is noisy. A short note that names the slice you can walk. No spray. No "easy apply" on 30 roles at 11pm. Easy apply is how your primary paper goes into a pile with a title mismatch and you never hear why. Five targeted beats thirty sprayed, including for your morale. Morale is a delivery resource. Protect it.

Retrospective on the pack: after every loop, one change. They poked cost — tighten the band. They poked Python — add a sentence to the stance, or schedule the one next skill. They poked "why this seat" — your intro is still muddy. One change, not a rewrite. Version the pack (date in the cover). You are running a product. The product is you-in-this-seat. Treat it like one. Then go home. A search that eats every evening is a program without a WIP limit. You would not let a team do that. Do not let you do that.

Pick one. Not ten. If FDE is primary: make the wiki copilot actually retrieve on a folder of 20 PDFs — a running slice, an eval table that is measured, not sketched. If delivery is primary: sit with your security team (or a public equivalent policy) on the real approved path and write the data-handling note against it. If solutions is primary: run one discovery using the day-20 playbook on a volunteer (a colleague's team, a nonprofit, a friend who runs ops). If AI PM: write a one-pager for a feature in a product you actually use, with evals and a refusal policy. If transformation: score a real intake (even a fake-but-specific list of 12 ideas from your company) and produce a kill list.

Four-week mini-plan, two hours twice a week — not another 30-day crush unless you want it. Week 1: set the environment or the meeting or the volunteer. Week 2: first pass (retrieve runs, or the security conversation happens, or discovery is booked). Week 3: eval or write-up. Week 4: put the output in the appendix of the pack and add one bullet to the resume only if it is still honest. That is depth. A new cert stack is a detour. You were warned on day 1. The warning still holds.

Do not pick a skill because Twitter said agents. Pick the skill that closes a fail-gap for your primary, or that makes a labeled hole measured. If your fail-gap is "cannot talk through a thin slice," the running RAG is the skill even if delivery is primary. If your fail-gap is "no specific delivery anecdote," the skill is not technical — it is writing the messy-program story until it is timed. Be honest about which gap is actually open. Ego will pick a glamorous skill. The loop will pick the fail-gap.

Time the skill so it cannot eat the job OS. Two hours, twice a week, on the calendar, titled. If a loop is booked, the skill yields. Loops are the workstream. The skill is a backlog item. You would not let a tech-debt epic eat a go-live. Do not let a side tutorial eat an interview. This is obvious on paper and hard on a nervous Sunday. The calendar is the control.

What you will not do this quarter: a master's, a full ML specialization, LeetCode 150, a public Twitter brand, five certs, a startup on the side to "prove" you can build. Those are allowed later. They are not this quarter. Write them on the will-not-chase list from day 22 and leave them there. The people who stay in the course-pile for a year are the people who keep adding a stack instead of running the pipeline.

If you already have an offer, the one skill is still useful, and it becomes week-4-of-the-job optional homework, not a search blocker. If you are employed and not leaving yet, the skill can be internal: the data-handling note with your real Security, a discovery on a real team. Internal proof beats another independent design. Use your access while you have it. Do not wait for the new title to start the muscle. The title follows the muscle more often than the other way around.

You can explain models without magic. You can design RAG and say no to an agent. You can cost a workload. You can write evals and HITL. You can charter a slice. You can walk it in 12 minutes. That is a legitimate pivot story for the seats this academy named — if you keep the honesty pass. The market is real. It rewards people who reduce the risk of AI work, not people who recite model names. You already knew how to reduce delivery risk. Now you can point it at this object.

Inventory the six artifacts with one sentence of proof each, written tonight, past tense, no hype. Example: "1. Prompt library — I can version a contract for status and RAID. 2. Wiki design — I can walk a thin slice with a refuse path. 3. Operating pack — I can put owners, bars, cost, and a kill on one table. 4. Resume — the primary seat is in line 1 and the independent work is labeled. 5. Walkthrough — I can do 12 minutes and the likely eight tired. 6. 90-day — I can say what I would output in 30/60/90, and I have a search cadence with WIP limits." That list is the close. Read it out loud.

Also list what you cannot prove. You cannot claim a production deployment you do not have. You cannot claim you are a staff engineer. You cannot claim a cert you did not earn. You cannot claim a percentage you did not measure. Writing the cannot-prove list is part of the honesty pass, last time. Keep it next to the can-prove list. Interviews will ask both. You will not have to invent in the chair.

If a thing is almost true ("I ran a notebook once"), put it on the can-prove list in its small form, not its large form. Small form survives. Large form dies in the engineer round. You have spent a week on this distinction. Do not get generous tonight because you want a cinematic ending. Cinematic endings are for films. You are closing a workstream.

Read positioning v1 out loud once. If it still names five seats, cut it now. If it still lacks a number, add one now. Then leave it. You will tweak after loops. You will not tweak because today is the last day and you are sentimental. Sentiment makes mush. Mush does not get offers.

The prove-tonight list is also what you send if a hiring manager says "what have you actually done." You may paste the six sentences. You may not paste this academy's curriculum. The proof is your objects, your numbers, your no. That is enough to start the pipeline tomorrow morning. It is not enough to retire on. It is enough to start. Starting is the point of day 30.

Write it as if your future manager asked in the last round. One page. Three blocks. Each block: output, owner (you plus who you need), risk. Not a manifesto. Not a 12-workstream Gantt. If you cannot fit it on one page, you do not own it yet — the same rule as the operating pack. Interviewers will read a page. They will not read your hopes.

Stealable shape, delivery primary. "30: map in-flight AI work; publish a scoring rubric; name one owner and one slice; confirm lightest approved path with Security. Output: a one-table map plus a pick. 60: shadow the slice; golden set of 40; data-handling note; death-date any inherited POC. Output: eval table and a security-commentable page. 90: assist-mode or a documented kill; steering paper; yes/no on a second slice. Output: a decision in writing. I will not staff a platform in this window. I will keep a list of what repeats." Rewrite with your seat's nouns.

Add a line on what you need from them: access, a sponsor who can name an owner, the banned-tools list, the last three steering packs. That line is an ask. 90-day plans without an ask are essays. You have spent two days on asks. Put one here. If they cannot give you a sponsor in 30 days, that is data about the job. You want that data while you can still leave, not in month five.

If you are applying internally, the 90-day plan should name real objects (their Copilot rollout, their vendor, their factory) and still be polite about holes. You are not performing a roast. You are offering a sequence. If you are applying externally, keep Atlas or a clearly labeled independent slice as the example, and say you would re-noun it in week 1 against their map. Do not pretend you already know their politics. Pretending is how external hires die in 90 days. Asking is how they live.

Put the page in the pack as artifact 6. Date it. Tick the last capstone box. Export the sendable PDF again so it includes this page. You now have a pack you could send after an interview when they say "what would you do in 90 days." That sentence happens. Having the page is how you look like a person who has done this. Not having it is how you improvise a transformation speech. You did not come this far to improvise a transformation speech.

If you already have the job, send the page to your manager in week one as a draft, not as a demand. "This is how I would use the first quarter; I want your red pen." That is a working relationship. If they tear it up, you learned the job. If they sign it, you have a contract. Either way you did not wait to be told. Delivery leads do not wait to be told the first 90 days. They propose.

A go-live has a list: what shipped, what is deferred, who is on call, what the first-week review is. Close this month the same way. Shipped: the six artifacts, a primary and a backup, a voice you can use tired. Deferred: the one next skill, the conversations not yet sent, the measured eval if you did not run code. On call: the job OS, WIP limits, take-home policy. First-week review: a standup with yourself next Friday, 30 minutes, counts only.

Send three 90-word paragraphs this week to people who have seen you deliver. That is the further work, and it is not optional if you want the pipeline to exist. Ask for 20 minutes, not a referral form. If you send zero, you have a notebook. If you send three, you have a workstream. The month does not get you a job. The workstream after the month gets you a job. You knew that on day 1. It is still true.

Do not announce on LinkedIn that you "completed an AI transformation." You may update the About you already wrote. You may not post a carousel of lessons. The seats you want are filled in rooms, not in carousels. If you want to write, write one specific paragraph about evals or a kill, in your voice, without a stock photo. Or write nothing. Silence plus three conversations beats a carousel plus zero.

Protect the pack from yourself. Make a dated export. Tomorrow you will want to redo it. Redo only after a loop teaches you. Otherwise the pack is frozen until next Friday's standup. Frozen artifacts are how you show up consistent. Living documents that change every night are how you send the wrong PDF. You have a version. Use it.

If this month felt too thin, that is allowed. You can run the four-week skill. You cannot erase the month because it was not a degree. It was never a degree. It was 60 hours pointed at five seats. 60 hours is a lot if you produced the objects. It is nothing if you only read. Check the inventory. If the objects exist, you have enough to start the pipeline. If they do not, do not apply yet. Close the fail-gaps this week, then apply. That is judgment. Judgment is the job.

You are done when: the 90-day page exists, the job OS has numbers, the one skill is named, the six sentences of proof are written, the last box is ticked, positioning v1 has been read out loud, and three names are on a list to message. Then stop. Go do something that is not this. A workstream that never closes is a museum. Close it. Tomorrow the pipeline is open, with limits, like a professional.