Get straight to
the techie stuff.

Codex desktop & CLI

Use a skills-enabled Codex host. Download the ZIP and extract the skill so that SKILL.md sits directly inside the prompt-adviser folder, not a second nested folder.

1. Create the current personal skills directory.
2. Extract the ZIP’s skill contents into ~/.agents/skills/prompt-adviser/.
3. Start a new Codex session and invoke $prompt-adviser with your starting brief.

mkdir -p ~/.agents/skills/prompt-adviser
# Extract the skill contents here:
# ~/.agents/skills/prompt-adviser/SKILL.md

$prompt-adviser

The Codex ZIP does not automatically install into ordinary ChatGPT. Codex desktop and CLI are different surfaces.

Claude Code & supported accounts

For Claude Code, extract the skill contents into your personal skills folder and check that SKILL.md is directly inside prompt-adviser.

mkdir -p ~/.claude/skills/prompt-adviser
# Extract the skill contents here:
# ~/.claude/skills/prompt-adviser/SKILL.md

Start a new Claude Code session, then invoke:

/prompt-adviser

Claude account uploads

Upload the packaged ZIP only where your Claude account exposes skill controls. Availability depends on your account, plan and enabled features; an ordinary chat attachment is not equivalent to installing a skill. Native Claude behaviour has not yet been validated for this project.

Google Antigravity

Download the packaged edition and extract it into the project’s skills directory. Replace <project-root> with the actual project folder; do not paste the angle-bracket placeholder into a shell command.

<project-root>/.agents/skills/prompt-adviser/SKILL.md

Open the project in Antigravity and explicitly ask the assistant to use the Prompt Adviser skill. Check your current host guidance for skill discovery and reload behaviour.

Shared source. Platform editions.

The core workflow is maintained in a shared source template, with platform-specific files and packaging. When contributing, edit the source and rebuild the generated editions using the repository’s documented process rather than editing a generated ZIP.

The skill is a set of instructions interpreted by the host, not a separate AI service. Host capabilities, available models and permission systems still determine what can happen.

No automatic upgrades or telemetry. Download and replace your installed edition deliberately when you choose to update, checking the changelog first.

The first 80% checkpoint

Completeness describes coverage of goal, context, output, constraints and checks. It is not a quality rating or a guarantee of success. You may finalise at any score.

Questions and model guidance

Review rounds contain three useful numbered questions plus “Shall I ask more?” Answered context is carried forward into another completeness check. The adviser may recommend a suitable available model, but does not switch it. A model that fits the task saves tokens and speeds up your reply.

Sources, without exposing the draft

At the first 80% checkpoint, a separate generic-task prompt-directory lookup may identify at most one useful pattern. It must be attributed and checked for reuse rights. No suitable match is a valid result. If lookup is unavailable, an offline fallback must be labelled, not presented as a live source result.

Optional roles

Role choices include No role. A role is inserted only after selection; a recommendation is not consent. Roles provide framing, not credentials or additional model capability.

Optional Jev assessment

The adviser considers whether the task involves repeated structured decisions. Jev is optional and never called during ordinary prompt review. One creative brief generally does not need it.

Jev & OpenRouter

Jev is intended for structured, repeated routing, classification and scoring. It is not the tool for generating prose, images or video.

The OpenRouter Decisions API requires an API key and incurs billed usage. Using it requires specific authorisation for the data being shared, the service and the intended action. The helper is never called automatically; ordinary review does not make a Jev request.

Review is not execution.

A final prompt is copyable text. Producing it, choosing a role or accepting a suggestion does not authorise task execution. Execution starts only on a direct “Run this prompt”, subject to the host’s permission controls.

Source queries use generic task terms rather than private drafts. That does not replace your assistant host’s or provider’s policies: private context entered into those services remains subject to their data handling rules.

This informational website collects no visitor prompt content. The walkthrough is local and scripted: no AI calls, Jev calls, authentication or form submissions. If you opt in, Google Analytics counts page views and edition download clicks only; you can reject or withdraw at any time via Cookie settings. See Privacy & cookies.

Methods research.
Not a product benchmark.

Relevant details reduce ambiguity, and clarification can expose missing context. Longer prompts are not automatically better. These papers support techniques that informed the workflow, not measured gains for Prompt Adviser.

Validation & contribution

Codex exercises have tested parts of the workflow. Claude and Antigravity instructions have been simulated through Codex, not validated in their native hosts. There is no product-quality benchmark yet.

Help by reporting reproducible issues, checking native platform behaviour or contributing improvements. Use fictional or anonymised examples only; do not include private prompts, customer records or API keys.