Writing GPT-5.6 Prompts

An agent-agnostic Agent Skill for writing, rewriting, reviewing, and handing off lean, outcome-led prompts for GPT-5.6 and GPT-5.5.

It helps coding agents remove legacy scaffolding and duplicated instructions while preserving the real contract: outcome, evidence, constraints, autonomy, output, validation, and stopping conditions.

Install

Install globally with the open skills CLI:

npx skills add CodingCossack/writing-gpt-5-6-prompts -g -y

The skill itself contains no harness-specific workflow or tool dependency. The CLI detects installed agents and supports Codex, Claude Code, Gemini CLI, OpenCode, Cursor, GitHub Copilot CLI, Windsurf, and many other Agent Skills-compatible tools.

Install for selected agents:

npx skills add CodingCossack/writing-gpt-5-6-prompts -g -y \
  -a codex -a claude-code -a gemini-cli -a opencode

Codex native installer

Ask Codex:

$skill-installer install https://github.com/CodingCossack/writing-gpt-5-6-prompts

Restart the agent after installation if it does not reload skills dynamically.

Manual installation

Clone the repository into the global skill directory used by your agent:

git clone https://github.com/CodingCossack/writing-gpt-5-6-prompts.git \
  ~/.codex/skills/writing-gpt-5-6-prompts

Common global roots include:

Agent	Global skill directory
Codex	~/.codex/skills/ or ~/.agents/skills/
Claude Code	~/.claude/skills/
Gemini CLI	~/.gemini/skills/
OpenCode	~/.config/opencode/skills/

For several agents on one machine, keep one clone and symlink it into the other roots. That gives every agent the same payload and one update point.

Update

If installed with skills:

npx skills update writing-gpt-5-6-prompts -g -y

If installed with Git:

git -C ~/.codex/skills/writing-gpt-5-6-prompts pull --ff-only

Use

The skill is designed to trigger automatically when an agent is asked to create, rewrite, optimise, review, or hand off prompts and agent instructions for GPT-5.6 or GPT-5.5.

Explicit invocation also works:

Use $writing-gpt-5-6-prompts to rewrite this coding-agent brief for GPT-5.6.

Use $writing-gpt-5-6-prompts to review this system prompt for duplicated rules,
unclear autonomy boundaries, missing validation, and procedural over-scaffolding.

Use $writing-gpt-5-6-prompts to turn these notes into a compact handoff prompt.

What it optimises

• Outcome-led task contracts instead of unnecessary procedures.
• Explicit success criteria, evidence rules, constraints, and finish lines.
• Clear autonomy and approval boundaries for action-taking agents.
• Content-preserving brevity rather than vague “be concise” instructions.
• Conditional guidance for research, coding, creative work, tools, subagents, long-running tasks, and handoffs.
• Removal of duplicated harness rules, generic metacognitive exhortations, and redundant tool instructions.

The skill does not choose API parameters or application architecture unless the prompt-writing task specifically concerns them.

Structure

.
├── SKILL.md
├── agents/openai.yaml
└── references/
    ├── source-map.md
    └── task-modules.md

SKILL.md contains the core method. The references are loaded only when the target prompt needs their detail, keeping normal context use small.

Source grounding

The guidance is derived from OpenAI's official model prompting documentation:

• GPT-5.6 model guidance
• GPT-5.5 model guidance

The included source map records how those sources translate into the skill and which API/runtime concerns were deliberately excluded.

Contributing

Issues and focused pull requests are welcome. See CONTRIBUTING.md. Report security issues using SECURITY.md.

Licence

MIT