Three Must-Have Skills Unlock GPT‑5.6’s Super‑Human Performance

Yesterday OpenAI released GPT‑5.6, marketed as “Frontier intelligence” while merging Codex into ChatGPT. The author examined the model report and summarized the headline claim: the model is smarter, cheaper, and more durable for long‑running tasks.

Official benchmarks note that GPT‑5.6 uses 1.6× fewer tokens for slide‑comparison tasks and reduces token consumption by over 20% in 30 real‑world website‑building dialogues, indicating better focus and tool usage.

Skill 1 – Anthropic frontend‑design

The skill’s core rule is “don’t produce one‑click templates.” It forces the model to define a concrete subject, then choose colors, fonts, and layout while taking an aesthetic risk. The author fed the prompt:

帮我给独立咖啡品牌『夜行』做落地页，深色底配单一高饱和强调色，hero区放一个会晕开的咖啡渍动画当signature，别给我奶油色那套。

GPT‑5.6‑Sol generated a landing page with a dark background, a high‑saturation accent color, and a slow radial coffee‑stain animation in the hero area, avoiding default skins and gradient corners. The output demonstrated genuine design judgment, as the model resisted the usual template‑like defaults.

Skill 2 – guizang‑ppt‑skill

This community‑maintained skill converts an article into a single‑file HTML horizontal‑paging PPT with two visual systems, a Swiss‑style narrative, and 22 locked layouts that prohibit arbitrary structure changes. The author supplied the prompt:

帮我把这篇文章做成瑞士风PPT，7页左右，要2-3张配图，主题色用克莱因蓝IKB。

Official documentation claims the skill produces a fully editable presentation with built‑in hierarchy and design sense. In the author’s test the claim held up, and the generated PPT respected the locked layout constraints.

Skill 3 – Deli_AutoResearch (DeepSeek)

Victor Chen’s Deli_AutoResearch framework addresses the failure modes of long‑duration agents—cognitive loops, dead‑states, and fragility—by enforcing zero interaction, persisting all state (task_spec, progress, findings, directions_tried) to files, and forcing a pivot when stuck. Sessions are limited to 15 rounds or 30 minutes. A three‑layer watchdog monitors the process: L0 stays in a shell, L1 checks hourly, and L2 reports within the business loop, allowing any layer to revive the others.

The author prompted the framework:

用Deli框架自主调研『Skill加持的Harness技术综述论文』，每次爹地啊里程碑写进findings.jsonl，卡住自动换方向，三层心跳全开，全程不许问我。

After execution, the findings.jsonl file contained a long list of milestones, and the iteration_log showed two wall‑hits and two automatic pivots, shifting the focus from pure technical routes to production capacity and geopolitical considerations. The author observed that GPT‑5.6 stayed focused, used tools effectively, and required almost no human intervention, producing clear reports and intuitive graphs.

Reflection

The three skills compress hard‑won know‑how—template‑design frustrations, PPT layout constraints, and autonomous‑agent safeguards—into reusable artifacts that empower GPT‑5.6. The author argues that while manual crafts are devalued, those who codify them into skills are gaining value.

Relevant resources:

https://victorchen96.github.io/auto_research/framework.html
https://github.com/op7418/guizang-ppt-skill
https://github.com/anthropics/skills/blob/main/skills/frontend-design/SKILL.md