I have been using LLMs to assist me with work: Gemini, Glean, Rovo, Gemini CLI and Claude Code. As an example, I am close to being able to generate a useful first draft of an assessment outline for a client engagement. To get there I have built up a layered context made up of memory files describing my role, the structure of our SOWs and proposals, my writing style, and two years of past work parsed and indexed. The goal: drop in a transcript or two, point the LLM at the right structure, and I get back something that sounds like me, organised the way I would organise it.
The obvious next thought is: should I share this?
The instinctive answer is yes; sharing scales effort. If the skill is useful for me, it could be useful for someone else, and the marginal cost of distribution is close to zero. The prevailing industry framing on agent skills runs in this direction: institutional knowledge as portable asset, implicit expertise made explicit, a shareable bundle that lifts the whole organisation. The idea is appealing. However, I don’t think it understands what these skills are.
Most of what makes my skill useful is not the skill, it’s me. The structure, the prompts, the context files: those are packageable. The judgement about when the assessment outline is on point, which phrases sound right, what should be flagged for human review, which parts of a transcript matter, comes from years of practice doing the work. Hand the skill to someone who has not done the work and you have given them the scaffolding without the scaffolder.
This is the part I keep coming back to. There is a tempting middle path that says: fine, share it as an agent and let feedback improve it over time. The reality is more complex. If twenty people feed back into a single agent, you do not get the wisdom of twenty experts. You get the noisy mean of twenty different opinions, each valid in their own context, none of them coherent together. It is the same problem you get when you ask twenty friends and family for advice on a hard decision; every answer is a fragment, and the average tells you nothing. The skill works for me because the feedback loop is one operator deep. Widen the loop and the signal collapses.
There is a more pragmatic version of the same vision: agents that autonomously create and refine their own skills as they work alongside one operator. A single feedback loop, kept tight, where the skill is the encoded output of one person doing the work. That is closer to how expertise actually develops in a career: you do not improve by averaging other people’s notes, you improve by doing the thing, getting feedback, and doing it again. The skill is the residue of the practice.
So the question is not should skills be shared. It is which skills are separable from the person who built them.
Some are. A SOW review skill that runs against an explicit checklist (payment milestones, contract term, IP clauses, MSA alignment) is a fair candidate. The rubric is largely explicit. The voice is doing less work. Two reviewers will disagree at the margin, but the core is rules-shaped, and rules can be lifted out of one person’s head and dropped into a shared file. That kind of skill is closer to a checklist with a model attached than to a personal craft.
The assessment-outline skill is the opposite. The rubric exists, but the parts that do the real work are tone, judgement, what to leave out, when to push back on a transcript and when to take it at face value. Strip those out and you have a templating engine. Useful. Not the thing.
That gives me a working test. (1) If I can write down what good looks like as rules, the skill is shareable; package it. (2) If what I mean is “it sounds like me”, the skill is not shareable; leave it alone, and accept that this one is mine the way my handwriting is mine. (3) If a skill drifts between the two, split it. Share the rubric-shaped half. Keep the voice-shaped half.
The career analogy holds. We do not expect to share our experience as a portable asset; we expect colleagues to develop their own. AI skills are starting to look the same. There will be shared agents that are checklist-shaped, rubric-shaped, well-bounded, and governed centrally. And there will be personal skills, the way every craftsperson develops their own tools. Both will exist. Neither will replace the other.
What I will not do is pretend the second category can be shared by uploading it. The skill is not the file. The skill is the pairing of the file and the operator. Change the operator and you have a different skill, usually a worse one.