How to AI-proof your job using Claude

What the data actually says

The number worth knowing is narrower than the headlines. Stanford's Digital Economy Lab, working from real payroll data, found employment of 22 to 25 year olds in the most AI-exposed occupations sits about 19% below where it would otherwise be. Experienced workers in those same occupations show no comparable gap, and the mechanism is slower hiring rather than layoffs.

The part that matters more for you is the sorting rule underneath it. Declines cluster where AI substitutes for the human task. Where AI complements the person, employment is flat or rising. So the useful question is not "is my job on a list", it is "which of my tasks can be handed in finished by a model, and which ones need me in the loop". These four builds answer that, then move your week toward the second kind.

The four builds, ranked

Ordered fastest first, so the quick one tells you where to spend the bigger effort.

1. 🧠 The task audit (about 30 minutes). Paste a real week of work into Claude and have it sort every task into delegate, do together, and human-only. Mechanism: you stop guessing which parts of your role are exposed and get a sorted list with hours attached, so the next three builds aim at something real.
2. 👩‍💻 Your SKILL.md (about 30 to 45 minutes). Package the thing you are genuinely best at into a skill file Claude loads automatically. Mechanism: your unwritten method becomes a portable asset with your name on it, instead of knowledge that only exists while you are in the room.
3. 📊 One visible tool (about 30 to 60 minutes). Turn a spreadsheet your team already argues over into a one-page dashboard and share it the same day. Mechanism: leverage that people can open and click is evidence. Quiet competence is not visible when decisions get made.
4. 🗂️ Own the workflow (about 45 to 60 minutes). Rebuild a recurring deliverable as a Claude Project so it runs in minutes instead of hours. Mechanism: the person who operates the automated version of a process cannot be swapped out by that process. They run it.

The task audit

This is the one to do today, even if you do nothing else. It takes a vague worry and turns it into a list with hours next to it. Open any Claude chat, paste last week's calendar and task list, and run this.

I want to audit my work week and see which parts of my job you can already
do, which parts I should be doing WITH you, and which parts are human-only.

My role: [job title, industry, team size, who I report to].
My week, pasted from my calendar and task list: [paste the real thing here].

Sort every single task into three buckets. Bucket 1, delegate now: you could
produce a usable first version from inputs I already have. Bucket 2, do
together: you draft or analyse, I direct, verify and decide. Bucket 3,
human-only: it runs on relationships, judgment, taste or accountability that
I hold. For every bucket 1 task, tell me exactly which inputs you would need
from me and estimate how long that task takes me today.

Then total the hours sitting in bucket 1 and give me a redeployment plan.
Pick the single task I should automate first, choosing the best ratio of
hours saved to effort to set up, and explain why you picked it. Say which
bucket 3 work I should move the recovered hours into so my week shifts
toward judgment instead of drafting. Finish with one honest sentence I could
say in a performance review about that shift.

Whatever comes out on top of that redeployment plan is what you build in the next three sections.

Your SKILL.md

A skill is a small folder with one markdown file inside it. It holds a name, a one-line description of when to use it, and your steps. Claude reads that description on its own and pulls in the whole skill the moment a request matches, so you explain your method once and it runs your way in every conversation after that. The mechanism here is ownership: a generic model has no access to how your team actually does the work until somebody writes it down, and that somebody should be you.

1. 🎯 Pick the task you are best at, not the one you hate most. The skill should encode judgment other people ask you for. The weekly report, the client email that lands, the way you turn messy notes into decisions.
2. 📄 Fill in the starter below. Save it as a file called SKILL.md inside a folder named after the skill. The folder plus that one file is the whole thing.
3. 📦 Zip the folder. Right click the folder, compress it. That zip is what you upload.
4. ➕ Add it to Claude. In the Claude app, open Customize, then Skills, then Add, then Upload a skill and pick your zip. In Claude Code, drop the folder into ~/.claude/skills/ instead. Either way it is now live.
5. 🔁 Run it on real work twice. Read what comes back, tell Claude what is off, and have it rewrite the SKILL.md. Two passes is usually the difference between "close" and "this is exactly how I do it".

Here is the starter. Everything in square brackets is yours to replace, and the rest is the structure that makes it work. Two formatting rules for the top block: the name must be lowercase and hyphenated, under 64 characters, with no "claude" or "anthropic" in it or the upload gets rejected, and the description has to fit in 200 characters.

---
name: your-skill-name-here
description: Replace this line with what the skill does AND when to use it, in one sentence.
---

# [The task, in plain words]

## When to use this
Use this whenever [the trigger, e.g. I paste raw notes and ask for the
Monday status report].

## What I will give you
- [input 1, e.g. my raw notes from the week]
- [input 2, e.g. last week's version, for continuity]
- [input 3, e.g. the metrics export]

## Steps
1. [the first thing you always do, e.g. pull the wins, blockers and next
   steps out of the raw notes]
2. [the checking step, e.g. verify every number against the metrics export
   and drop anything unsupported]
3. [the formatting step, e.g. write it in our structure: one-line summary,
   then Wins, Blockers, Next]
4. [the finishing step, e.g. list anything needing a decision from my
   manager separately at the bottom]

## Rules that matter
- [non-negotiable 1, e.g. never invent a number, ask me instead]
- [non-negotiable 2, e.g. keep it under 250 words]
- [non-negotiable 3, e.g. plain language, no buzzwords, no filler]

## Worked example
Input: [paste a short, real example of what I hand over]
Output: [paste the finished version I was happy with]

Four tasks that make excellent first skills, because they are repeated, opinionated, and currently living only in your head:

Ship one visible tool

The mechanism here is visibility, not sophistication. When headcount and scope decisions get made, a small tool your team opens every week is undeniable evidence you create leverage. The bar is genuinely "my team opened it twice this week", not production software. Take a spreadsheet your team already touches (campaign tracker, pipeline, budget burn, bug list), paste it into Claude, and run this.

I'm pasting in a spreadsheet my team uses every week: [say what it is, e.g.
our campaign tracker, our hiring pipeline, our monthly budget burn].

Build me a single-page dashboard as an artifact that my team could open and
actually use. Before you design anything, read the data and tell me the
three questions someone would open this to answer, then build around those
three questions instead of around every column I gave you.

I want a row of the key numbers across the top, two or three charts that
show movement over time rather than a snapshot, and one filter or toggle so
people can narrow it to [the dimension that matters, e.g. channel, owner,
month]. Keep the styling clean and warm, label everything in plain language
rather than column names, and make sure it is readable on a phone.

Flag anything in my data that looks wrong, duplicated or inconsistent
instead of quietly charting it. When it's finished, write me two sentences I
can post in the team channel that say what it answers and who it's for, not
how it was built.

Own the recurring workflow

This is the biggest payoff and the reason it comes last. Pick the recurring deliverable your team quietly dreads, the weekly status report, the campaign recap, the meeting-notes-to-actions pass, and rebuild it as a Claude Project. A Project is a workspace that keeps files and custom instructions attached to every chat inside it, so the context stops being something you re-paste and starts being something the workflow owns. And the workflow runs through you.

1. 🗂️ Create the Project. In Claude, open Projects and make a new one named after the deliverable. Projects are available on every plan including Free, where you can have up to five (Anthropic's Projects overview).
2. 📄 Add three finished examples. Three past versions you were genuinely happy with. This is the single highest-leverage step: three real examples teach format and voice better than any amount of describing it.
3. 📥 Add the raw inputs those examples came from. The messy notes, the exports, the threads. Now Claude can see the whole transformation, input to output, not only the polished end.
4. ⚙️ Write the project instructions. Use the template below. This is the part most people skip and it is what makes the output consistent instead of merely decent.
5. 🔁 Run it, then fix the instructions, not the output. When something comes back wrong, resist editing the text by hand. Add a line to the instructions so it never happens again. That is how two hours becomes ten minutes permanently rather than once.

You produce our [name of the recurring deliverable, e.g. Monday client
status report].

Context. I work as [role] at [type of company or team]. The audience is
[who reads this and what they do with it]. In this project you have [number]
past versions I was happy with, plus the raw inputs each one was built from.
Treat those past versions as the standard for format, structure and voice,
and match them closely rather than improving on them.

Every time I paste raw inputs, do this. Pull out only what the audience will
act on and leave the rest out. Check every number, name and date against the
raw inputs I gave you. Write it in the exact structure of the past versions.
Keep it to [length, e.g. under 300 words or one page]. If something the
audience needs is missing from my inputs, ask me one round of questions
before writing rather than filling the gap yourself.

Never invent a figure, a name or a status. If a claim is not supported by
the inputs, leave it out and list it at the bottom under "needs
confirmation" so I can chase it.

Tone: [how your team actually writes, e.g. plain, direct, no buzzwords, no
hedging].

Where this backfires

Which is the real argument for build one. The audit is not busywork before the fun part, it is how you find your own frontier before you bet a deliverable on it. Practically: run any new workflow twice on work you already know the right answer to, and only then trust it on work you do not. Test before you trust, every time.

What you need for each build

The proof behind all of this

• Stanford Digital Economy Lab, Canaries in the Coal Mine. The 19% figure, why it is specific to young workers in AI-exposed occupations, and the substitute versus complement finding underneath it. digitaleconomy.stanford.edu
• PwC 2026 Global AI Jobs Barometer. The 62% wage premium for workers with AI skills, up from 57%, and AI-skill postings growing 69% against 9% for the market overall. PwC 2026 barometer release
• HBS and BCG, Navigating the Jagged Technological Frontier. The 758-consultant experiment, the gains inside the frontier and the losses outside it. aiinstitute.hbs.edu
• Anthropic Economic Index, January 2026. Augmentation rose 5 points to 52% of conversations while automation fell 4 points to 45%, which is the working-with pattern showing up in usage data. anthropic.com/research
• Agent Skills, official docs. The SKILL.md format, how the description triggers a skill, and both install paths. platform.claude.com docs
• How Anthropic teams use Claude Code. Non-engineering teams building their own internal tools, with the specific examples quoted above. claude.com/blog