Why You Can Ditch Most Superpowers Skills in the GPT‑5.6 Era

With GPT‑5.6 (or comparable models such as K3, Claude Fable 5) the coding agent can perform most basic development actions without explicit Skills. A bug‑fix request only needs a symptom description and expected outcome; the agent reads the project, follows the call chain, patches code, adds tests and runs verification automatically. Long tasks, approvals and multi‑Agent collaboration are also handled natively.

Progressive disclosure and context limits

The agent initially loads only each Skill’s name, description and path. Full Skill content is fetched only when a task matches – a mechanism called progressive disclosure. According to the Codex Skills documentation, the Skills list occupies at most 2 % of the model’s context window, with an absolute ceiling of 8 000 characters. If the list exceeds this budget the descriptions are truncated. With 100 installed Skills the initial view may already be incomplete, and overlapping descriptions can cause conflicts that the model must resolve on the fly.

Separating rules from Skills

Project‑wide conventions go into AGENTS.md or a dedicated rules file.

Task‑specific processes belong in a Skill.

Time‑consuming investigations are delegated to a Subagent.

Collections of Skills, Hooks, MCP and connectors are packaged as a Plugin.

Mechanical constraints that must never be missed belong to Hooks, CI, linters or tests, not to natural‑language reminders.

Mixing project rules with Skills often yields a bloated SKILL.md that is hard to maintain.

Security considerations

Because a Skill is an instruction set for the agent, third‑party Skills may contain dangerous commands or overly broad permissions. Review SKILL.md, the scripts/ directory and the references/ directory before installation.

Which Skills are worth keeping?

In the strong‑model era the author retains three categories of Skills:

Personal preferences or fixed artifacts that models cannot infer (e.g., article style, chart standards, internal templates, specific release processes).

Tasks that require professional judgment, scripts or reference material (e.g., security reviews, complex document handling, framework migrations, production checks).

Skills that prevent direction errors by asking clarifying questions before execution. An example is the grilling Skill, which repeatedly asks the user to clarify plans, decisions and dependencies, proceeding only after each answer.

Case study: SpringAI interview platform

The open‑source SpringAI interview platform (≈2.8 k Stars on GitHub) was used to test the grilling Skill. The goal was to connect mock interviews with a knowledge base. The agent first asked whether the first release should focus on fully personalized interviews or on standard Java/system‑design Skills. It then recommended limiting the knowledge base to a single knowledgeBaseId for simplicity.

The discussion highlighted that stronger models execute faster, but missteps become costlier, making direction‑checking Skills like grilling valuable.

Skill reduction checklist

Before keeping a Skill, ask:

Does the model already perform this action?

Do we repeatedly fail at this point without the Skill?

Does the Skill encapsulate scripts, templates, expertise or personal preferences?

Can we detect and remove the Skill when its rule expires?

Generic reminders such as “read project → write code → run tests” are omitted in favor of project‑wide rules in AGENTS.md. Only Skills that address frequent, high‑cost failures are retained, keeping the total number under twenty.

Further reading

References (plain URLs):

[1] Codex Skills documentation: https://learn.chatgpt.com/docs/build-skills

[2] AIGuide: AI application development, AI coding practice and interview guide – https://mp.weixin.qq.com/s/le3RzJsaAH22auUoB05y1Q

[3] Superpowers repository: https://github.com/obra/superpowers

[4] mattpocock/skills repository: https://github.com/mattpocock/skills

[5] grilling Skill definition: https://github.com/mattpocock/skills/blob/main/skills/productivity/grilling/SKILL.md

[6] Codex usage guide: https://javaguide.cn/ai-coding/practices/codex-best-practices.html

[7] AI coding practice guide: https://javaguide.cn/ai-coding/