Ninety days in three blocks of thirty. Each block ends in a checkpoint: read the numbers, ask three questions, decide to continue, adjust or stop.
The first week is my recommended first move on either path: inventory, hats, numbers, one message. Then one change at a time.
Hand this plan to the person who runs it. You appear only on the message, the hats, the week 4 incentive check and the checkpoints.
Weeks 4, 8 and 11 are buffers: they take what slipped before anything new.
Who does what, in pictures: the cartoon of the Start path, a company of 25 people with no designer, from day 1 to day 90.
The plan is built from single practices that named companies used, with my judgment where they join. Both plans are pre filled in the 90 day plan database below, one row per day, week and checkpoint. Delete the rows of the path you are not on, then add a runner and a date to each.
Company: {{name}}. Size: {{number}} people. Path: {{Start or Fix}}. Way of working owner: {{name}}. Runs the plan day to day: {{name}}. Day 1: {{date}}. Day 90: {{date}}.
| Day | Do this | Part | Runner |
|---|---|---|---|
| 1 | Self check. Write down M6 from the bill before any tool is cut. Close any of flags 1 to 4 this week | 00 | You; at 150 people the tool owner takes the inventory, a day |
| 2 | Pick one team: the longest review queue, or the UI most customers touch. Not the most enthusiastic one. At 25 people it is everyone who ships. Score the first use case | 01 | {{name}} |
| 3 | Name the hats on people already doing the closest thing, with weekly hours and what each stops doing. Start: experience quality owner first. Fix: tool owner first | 02 | You |
| 4 | Write down M1, M2 and M5 and their value today | 06 | {{name}} |
| 5 | Tell everyone, in writing, where they already read: what changes first, what stays the same for everyone, and that AI use is allowed and judged by whether the work works | 02, 03 | You |
Why one team on the Fix path: a change you can see is a change you can judge. Five teams at once give five sets of excuses. Why numbers first: a pilot with no owner, no end date and no number written down before it begins rarely becomes practice. Both my judgment.
Start, about 25 people: sequenced by steps, not teams, since one team is everyone. Blocks 2 and 3 run together from week 6. Usual runner: the senior engineer, with the PM on the experience quality weeks.
Fix, about 150 people: hats first, then rules before the map, because the bill and the access list are already a live risk. One team in block 2, five in block 3. Usual runner: the engineering lead who holds the rules hat.
The week by week rows are in the 90 day plan database. Who did this: Linear, about 100 to 300 people, reviewers open the preview before the diff (C01). Musixmatch, an estimated 100 to 200 people, agrees the plan before code (C08). PostHog, about 150 to 200 people, never automates review of auth, secrets, billing or public APIs (C04). Weaviate, about 74 to 100 people, leaves most line level review to agents with humans on scope (C07); do that in week 5 only if a review tool cleared its bar. Cursor, about 60 to 150 people, said AI replies would be labeled after its bot invented a policy (C12).
Counted per team, in the Scorecard. Read at the monthly look and at checkpoints, never weekly. Below about ten events, write the list or the count, not a rate. M1, M2 and M5 on day 4. M6 from the bill on day 1. M4 starts with the severity scale, M3 with the sample.
Before each checkpoint, write down the one change its numbers judge, and hold the others steady. Continue: the next block starts as written. Adjust: one thing changes, written down. Stop: reverse the change for that team and restart at week 5 with the next scored use case, keeping block 1. Stop is a legitimate outcome.
| Checkpoint | The change it judges | The three questions |
|---|---|---|
| CP1, day 30 | Gates and rules. Not passed until the high severity list is written. On Start the decision is continue or adjust, since nothing has changed yet | Gates only: every gate named, each name decided something? |
| CP2, day 60 | Start: preview first and plan before code. Fix: the review step | Did lead time fall without the other numbers getting worse? Did the gates hold? Would the team keep it if you stopped asking? |
| CP3, day 90 | Start: the sample and patterns.md. Fix: the widening to five teams | The same three, per team and per lead |
Who did this: WeTravel, 100 to 250 people by estimate, tested an AI review tool, said no, and said no again on a second look (C09).