Skip to content

feat: update prompting skill to GPT-5.5 (gpt-5-5-prompting)#366

Open
skyswordw wants to merge 1 commit into
openai:mainfrom
skyswordw:feat/gpt-5-5-prompting
Open

feat: update prompting skill to GPT-5.5 (gpt-5-5-prompting)#366
skyswordw wants to merge 1 commit into
openai:mainfrom
skyswordw:feat/gpt-5-5-prompting

Conversation

@skyswordw

Copy link
Copy Markdown

Summary

Renames the internal gpt-5-4-prompting skill to gpt-5-5-prompting and rewrites its guidance to follow OpenAI's published GPT-5.5 prompt guidance and migration guide (developers.openai.com).

Motivation

GPT-5.5 reasons more efficiently than GPT-5.4 and pays a real cost for noisy, process-heavy prompts. The current skill still teaches the GPT-5.4-era style (compact XML block stacks by default), which the GPT-5.5 docs explicitly recommend against — carrying a 5.4 prompt stack forward adds noise, narrows the model's search space, and pushes toward mechanical answers. This revives the direction of #301 (closed by its author as local-only) with content checked against the official docs.

Changes

  • Skill renamed: plugins/codex/skills/gpt-5-4-prompting/plugins/codex/skills/gpt-5-5-prompting/
  • Skill content rewritten for GPT-5.5:
    • outcome-first core rules (define destination and success criteria; let the model choose the path)
    • recommended prompt skeleton: Role / Personality / Collaboration style / Goal / Success criteria / Constraints / Output / Stop rules
    • reasoning-effort guidance aligned with the CLI's --effort values (nonexhigh): start lower than on 5.4, escalate only when the prompt is already tight; higher effort can cause overthinking
    • what to delete when migrating a 5.4-era prompt (step-by-step process rules, inline schemas, the current date)
    • stopping conditions, retrieval budgets, validation defaults, and tool preambles
    • absolutes (must/never/only) reserved for invariants; decision rules for judgment calls
  • XML block library and recipes kept but demoted to opt-in modules — they remain useful for explicit contracts (downstream parsing, high-risk fixes, adversarial review) but are no longer the default stack
  • Anti-patterns reference gains a GPT-5.5-specific section (carrying 5.4 stacks forward verbatim, over-specified process instructions, overused absolutes)
  • References updated in agents/codex-rescue.md and skills/codex-cli-runtime/SKILL.md
  • Tests updated in tests/commands.test.mjs

Testing

  • node --test tests/commands.test.mjs: 8/8 pass
  • Full npm test: 83 pass / 3 fail — the same 3 failures reproduce on unmodified main in my environment (pre-existing runtime.test.mjs/state.test.mjs issues unrelated to this change)

🤖 Generated with Claude Code

Rename the internal gpt-5-4-prompting skill to gpt-5-5-prompting and
rewrite its guidance to match OpenAI's published GPT-5.5 prompt
guidance and migration guide:

- outcome-first prompts instead of process-heavy XML stacks
- recommended skeleton (Role / Personality / Collaboration style /
  Goal / Success criteria / Constraints / Output / Stop rules)
- re-evaluated reasoning effort defaults (5.5 needs fewer reasoning
  tokens at the same level; higher effort can cause overthinking)
- what to delete when migrating a 5.4-era prompt
- stopping conditions, retrieval budgets, validation defaults, and
  tool preambles
- XML block library and recipes kept but demoted to opt-in modules

Update references in agents/codex-rescue.md and
skills/codex-cli-runtime/SKILL.md, and adjust tests.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant