Goodbye Superpowers: Which Skills Are Worth Keeping After GPT‑5.6?

Basic steps no longer need explicit Skills

Earlier Skills listed detailed checklists (read project, locate call chain, run tests, submit PR) because model reliability was low. With GPT‑5.6, Codex can fix a bug by receiving only the symptom and expected outcome; it 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 by Codex itself. External capabilities can still be added via Skills, Plugins, MCP, and Hooks, but the earlier problem of agents “running wild when steps aren’t fully written” has largely disappeared.

Too many Skills can conflict

Codex does not read the entire SKILL.md at conversation start. It first loads each Skill’s name, description, and path, then loads the full content only when the task matches (progressive disclosure). This consumes up to 2 % of the model’s context window, capped at 8 000 characters. When many Skills are installed, descriptions are truncated and some Skills may never be loaded.

With 100 Skills, the agent may not see all full descriptions. Broad descriptions cause multiple Skills to trigger for a simple change, and conflicting rules force Codex to decide which to follow, making debugging harder.

When the context window grows, old dialogues, tool explanations, and Skill descriptions compete for space, leaving only the most critical three‑to‑five sentences of project constraints reliably visible.

Separate rules from Skills

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

Task‑specific processes belong in a Skill.

Time‑consuming investigations belong to a Subagent.

Combine Skills, Hooks, MCP, and connectors into a Plugin for team distribution.

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

Why Superpowers is used less

Large Skill suites such as Superpowers add unnecessary complexity. In group discussions participants noted that Superpowers makes processes overly heavy, while the lightweight grilling Skill, though verbose, clarifies requirements and reduces rework.

The Superpowers workflow (brainstorming → writing‑plans → test‑driven development → git worktree isolation → Subagent execution → code review → final verification) remains valuable for complex, unfamiliar, or high‑risk changes, but for trivial edits the overhead outweighs the benefit.

Token consumption is a major concern; the full process consumes many tokens, whereas a strong model can adjust steps on the fly and request missing information only when needed.

Security risk: a third‑party Skill may contain dangerous commands or overly permissive requirements. Review SKILL.md, scripts/, and references/ before installation.

Skills that still matter in the strong‑model era

Three categories of Skills are retained:

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

Tasks requiring professional judgment, scripts, or reference material (security reviews, complex documentation, framework migrations, production checks).

Skills that prevent directional errors , such as the lightweight grilling Skill that asks clarifying questions before execution.

The mattpocock/skills repository (https://github.com/mattpocock/skills) exemplifies small, composable Skills that do not take over the entire development process.

Practical example with the “grilling” Skill

In the open‑source SpringAI Interview Platform (≈2.8k Stars, https://github.com/mattpocock/skills/blob/main/skills/productivity/grilling/SKILL.md) the grilling Skill was used to clarify interview scope before coding.

First it asks whether the interview should be fully user‑profile‑driven or use predefined Java/system‑design Skills, then recommends the former because existing question banks, difficulty grading, and scoring rules already align with that path.

Next it determines whether a single knowledge base or multiple should be used (the request only contains knowledgeBaseId), recommending a single‑base approach for the initial phase.

Finally it confirms that the interview entry points—separate pages for knowledge‑base interviews and regular mock interviews—share the same configuration components, avoiding duplicate UI logic.

Although code changes have not started, product goals, data model, and entry reuse are already defined, demonstrating that the Skill forces the model to ask ambiguous questions before proceeding, a step that strong models cannot replace.

Skill pruning checklist

Does the model already perform this task?

Without the Skill, does the same failure recur?

Does the Skill encapsulate scripts, templates, expert material, or personal preferences?

Can the Skill be detected and removed once its rule becomes obsolete?

Skills that merely repeat “read project → write code → run tests” are omitted because Codex handles them. Only tasks that cause costly errors merit dedicated Skills, and they should contain just enough logic to solve the problem without expanding into a full workflow.

Conclusion

As models become stronger, the cost of executing the wrong direction grows. Concise, purpose‑specific Skills—especially those that capture personal conventions, embed expert scripts, or enforce critical checks—still provide real value. Apply a disciplined pruning process: evaluate each Skill against the checklist, keep only those that prevent high‑cost failures or encode irreplaceable knowledge, and store project‑wide conventions in AGENTS.md instead of bloated Skills.

References

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

[2] AIGuide: AI Application Development, AI Coding Practice & 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 file: 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/