Editorial standards

Principles

• No invented evidence. We do not publish fabricated test results, scores, ratings, reviews, statistics or author credentials.
• Labels over claims. A prompt is only described as tested when a test record exists and is published on its page.
• Limitations are part of the product. Every prompt page lists what the prompt cannot do and what you need to verify.
• One page per task. We do not publish near-duplicate pages for keyword variations of the same intent.

How a prompt is developed

1. Task definition. We start from a real task and write down what a practitioner would need to supply and what they would expect back.
2. Drafting. The prompt is drafted — with AI assistance — as a set of inputs, ordered steps, explicit constraints and an output format.
3. Structural review. Every variable must be used, every placeholder must resolve, references must point to real pages and real sources, and the page must state limitations.
4. Publication. The prompt is published with the testing label that matches the evidence we hold for it.

Testing status labels

Untested

Reviewed for structure and clarity by the editorial team. No recorded model test yet.

Tested

Run against at least one model with a recorded test: inputs, output and rubric notes are published below.

Independently reviewed

Tested, and the test record was checked by a second reviewer.

Model-specific

Tested and tuned for specific models only. See model compatibility.

Needs retesting

A model or the prompt has changed since the last test. Earlier results may no longer hold.

Current state of the library: the launch library has passed structural review and is labelled Untested. Labels change prompt by prompt as test records are added; nothing is upgraded in bulk.

Testing methodology

A test record contains: the task being tested, the model and version, the test date, the input conditions (the exact values used for each variable), the expected output, the actual output, observed strengths, observed weaknesses and limitations. Records are shown in full on the prompt page.

Each output is assessed against the same seven criteria. For each one the tester records met, partially met or not met, with a note.

1. Instruction clarity. Could the instructions be read two ways? Did the model do what was asked without guessing?
2. Task completion. Did the response complete every part of the task?
3. Context handling. Did it use the supplied context correctly and avoid inventing missing facts?
4. Constraint adherence. Were length, scope, tone and "do not" rules respected?
5. Output structure. Did the response follow the requested format?
6. Reliability. Does a second run with the same inputs produce a comparable result?
7. Practical usefulness. Could a practitioner use the output with reasonable editing?

We do not publish numerical scores. A single number hides which criterion failed, implies a precision a handful of runs cannot support, and invites comparison between prompts for unrelated tasks.

Editor’s Pick criteria

At most one prompt per category is marked Editor’s Pick. It is the prompt we would point a newcomer to first, chosen because it (1) covers the broadest version of the task in that category, (2) has the most complete structure — inputs, ordered steps, constraints and output format, and (3) surfaces its own assumptions. Editor’s Pick is an editorial judgment about scope and design. It is not a performance claim.

Model compatibility

Compatibility is recorded per prompt and per model as Expected to work, Tested or Model-specific version. See how compatibility is labelled.

Versions and maintenance

Prompts are versioned. Any change to the prompt text creates a new version with a change note, shown in the version history on the page. When a model vendor ships a significant update, tested prompts for that model are set to Needs retesting until they are run again. Reader feedback is reviewed and can trigger a review or a retest.

Publication workflow

Content moves through six states: draft, in review, approved, published, needs review and archived. Only published content is visible on the site or listed in the sitemap.

Sources

Guidance about how models behave is attributed to the vendors’ own documentation and linked. We do not cite sources we have not read, and we do not cite search volumes or statistics without evidence.

Corrections

If something here is wrong or out of date, tell us. Material corrections are noted in the version history of the page concerned.