Original video in Chinese.
Key Takeaways
- Stronger models can perform tasks that previously needed specialist presentation skills or heavily prescribed workflows such as Superpowers. Generic instructions, broad process guides, and role-playing skills are losing value.
- Five categories remain useful: tool operations, proprietary methods, high-risk workflows with strong constraints, deterministic production, and project knowledge or organizational protocols.
- A valuable skill preserves a judgment system or workflow you have already worked out. What matters is retaining the ability to define problems, standards, and final decisions.
As models improve, many skills cease to be necessary.
I reached that conclusion while testing presentation skills, especially those made by Chinese developers. Most were disappointing. Many developers seemed to treat visual attractiveness as the central objective, overlooking visual logic: consistency and directing the audience’s attention.
PPT Master was relatively good, but I still had to point out what needed fixing.
Then I remembered the newly released Fable 5 model and wondered what would happen if I asked a top model to build the presentation directly.
Fable wrote code and chose a Python script to produce the deck. Ten minutes later, it had made ten slides. For the first time, I felt an AI presentation was approaching human quality. Compared with the supposedly specialist presentation skills, the difference was clear.
We have often said that model upgrades eliminate startups whose narrow capabilities become part of the model itself. Now the same is happening to skills.
When heavy workflows become unnecessary
Superpowers was another example. It imposed a senior-engineer workflow on AI to improve quality and reduce bugs.
That was effective with earlier models, which more often lost track or missed details. Forced process could substantially improve success on complex projects.
With models of the Fable 5 and GPT-5.6 generation, understanding and independent planning had improved. Clear requirements were often enough for useful decomposition and execution, with less need for heavy external constraints.
These models were also expensive. A long mandatory plan and many prescribed steps could make Superpowers slower, more token-intensive, and less economical.
I suggested that community members review their installed skills, including asking Codex to inspect them. I had Codex perform a cleanup myself, and also disabled rarely used plugins. It saved a meaningful number of tokens.
Five kinds of skills worth retaining
1. Tool operations
The model may know what to do without knowing how to operate your specific local CLI, API, or MCP. These skills tend to have the longest useful life.
For example, my newtype Cut skill’s value lies in a reliable workflow connecting Whisper, timecodes, FCPXML, DaVinci Resolve import, and protection of original media.
2. Proprietary methods
A skill can encode a judgment system you have defined yourself, rather than generic investment or writing knowledge.
My newtype AI Stack, for example, defines a seven-layer industry chain and methods such as tracking the migration of scarcity, to guide research on AI industry-chain investments.
3. High-risk workflows with strong constraints
Some errors are costly. Investment research, for example, depends on data sources, freshness requirements, and rules about which provider’s financial data takes priority. Skills remain useful for defining those boundaries.
4. Deterministic production
When output must satisfy a machine-verifiable format, use a skill with scripts.
Better models improve judgment without eliminating randomness. Scripts replace repeatedly generating code with running a tool that has already been validated.
5. Project knowledge and organizational protocols
A model cannot know your team, project, and long-term agreements by default, including naming, directories, delivery, and archiving rules.
A skill can act as a router that determines which workflow applies and when.
Preserve your judgment
Skills once compensated for limited model intelligence. Stronger models allow many of those additions to be removed or made much lighter, especially generic capability descriptions, pure role-playing, and broad process instructions.
I think the best future skill resembles a lightweight operating-system entry point: short routing instructions, a few boundaries, reference material loaded as needed, executable scripts, and explicit verification.
The deeper issue is what you have delegated to AI. Delegating execution is sensible. Delegating judgment, taste, problem definition, and value priorities can gradually weaken your agency.
I now judge a skill by whether it helps preserve my decisions. It should encode what I have already thought through so that a model can execute it more consistently.
That is part of what I mean by newtype: a person who can define problems, workflows, and the relationship between themselves and AI. Models can absorb more tools, but they cannot take away judgment you have truly internalized—provided you have developed it in the first place.
That is all for this episode. If you want to understand AI, reclaim personal sovereignty, and meet people with similar interests, join the newtype community. See you next time.