FORWARDED THIS? Subscribe here to get the next one.

SPECIAL EDITION · CLAUDE + COMPENSATION

A NOTE FROM GIAC

I'm writing this from an apartment in Madrid I moved into on Monday, surrounded by boxes I haven't opened and furniture I haven't assembled yet, but with ultrafast wifi, writing this comp leader guide to Claude, because priorities are priorities.

Last Sunday, to be the contrarian in the room while everyone was writing that AI will take over jobs and replace McKinsey, BCG and the lot, I thought of faking a post myself by saying Fable 5 is here to say goodbye to comp processes, just to remind comp teams that on Monday their budget and salary planning will still be waiting for them.

(Said by someone who doesn't have to go through it this year, for the first time in six years of in-house comp)

It started a chain I did not expect. People started commenting ‘guide’ and then more people did because they saw the others doing it.

So I spent the past evenings writing it properly, and it turned out to be the most useful thing I've written in a while, because researching it changed how I use these tools.

Testament to the fact you never stop learning with AI.

Giac

THE BREAKDOWN

The joke zeroed in on an argument I have had some version of four or five times since the beginning of the year. Nobody serious in those conversations thinks comp gets automated away. What we disagree about is what is already happening while we argue, and my honest read is that most of us are describing last year’s trends to each other in real time.

Something did move this year. Anthropic's own documentation now says prompts and skills written for earlier models are often too prescriptive for Fable 5, and can make the output worse. If you spent 2026 building a comp prompt library, part of it is working against you, which is why item 01 in the guide is deleting most of what you wrote last year.

Most of the 21 are practices with a prompt to start you off. Three ship as installable skills, which is the difference between something you have to remember to use and something Claude reaches for on its own.

Here are 2 of the 21, the ones that work on whatever Claude you already have.

Take the comp prompt you're proudest of. Delete every numbered step that describes how to do the work. Keep the goal, the constraints, the output format and the things that must not happen.

It takes about 10 minutes to find out how much your library is costing you.

This is the answer to the question people keep asking about what Fable is even for. It plans better than it executes. Ask for the plan, read it, correct it, and only then let the work happen. Practitioners landed on this independently and Anthropic now runs webinars about it.

For comp it maps almost too neatly. You'd never let an analyst start rebuilding a structure without seeing their approach first.

Anthropic merged Claude's memory across chat and Cowork on Tuesday. What you tell it in conversation is now there when you put it to work, which makes number 2 in the guide, writing your comp context down once, worth more than it was last week.

One catch. Memory is shared when Cowork runs in the cloud, and sessions running locally on your machine don't use it. So the mode that keeps your pay data on your own hardware is the mode that forgets you.

WHY FABLE 5 WON'T REPLACE YOUR COMP PROCESSES

I made that claim up to take the mickey out of it. Three reasons I don't buy it.

Rules and thresholds can live in the code AI writes for you, and I'd encourage that. The decision itself can't move. There's nowhere you can say the model decided it and have that hold, whether the question comes from an employee or from a works council under EU pay transparency. A process that ends in a decision somebody has to own doesn't get replaced. It gets a faster middle and the same ending.

A comp analyst who gets a match wrong has a bad day and fixes it. Build something that does benchmarking or job matching, put it in front of managers, and they'll expect it to be right nearly every time and to explain itself the way an analyst would. I'd put that at 95% and I think I'm being generous. Automating a process you can't explain doesn't retire the process. It moves the explaining to a worse moment.

A Range member asked how AI could help with job levelling. Their comp team was building a levelling agent. Their HRBP team, who hadn't been in the room, was building something that writes a job description to a desired level. So the levelling call was already made in the HRBP tool before comp ever saw the job.

Two owners starting from opposite sides debating who should make a decision.

No AI model can fix that. What looks like an AI problem is usually a governance problem made fast, and running a broken sequence faster gets you to the wrong answer sooner.

So the processes stay. What’s changing is where your time goes, and how well you can explain the outcome to the employee reading it. I know in China they’ve released super bionic humanoids able to read and react to human emotions, but I don’t think the explaining part is close to being automated yet.

Something AI will never replace is my former manager's face when they have to tell me I'm getting 2% this year.

RECENT EDITIONS