Research and compare a prompt
Reviewed November 2, 2026. This is an independent research workflow, with a local record tool. It does not connect to an OD account or run generations.
Open Test records · Read the source register · See retained native results
Choose evidence that answers the question
| Question | Useful evidence | What it cannot establish |
|---|---|---|
| What does a feature's guide say? | Dated official documentation, with the exact section and access date | Your account's current controls, entitlement or output quality |
| What has a creator found useful? | Credited creator material, with its stated model, route and limits | A universal parser rule, backend specification or reliability rate |
| Is a control available to me? | A permitted, manually recorded observation of the actual account control | Availability for other accounts or plans |
| Did this prompt help? | Exact inputs and all outputs from a bounded comparison using predefined criteria | General reliability from one attractive result |
| Does this local tool preserve my record? | Software checks and a browser exercise | Native OD behavior |
A search result is a lead. Read the source body before adopting a claim, and retain any extraction limit. A missing page, unknown model or absent cost is unknown; it is not proof of removal, a default model or a free action. Preserve conflicting sources rather than silently picking the most convenient one.
Where the plugins help
These roles were reviewed from the connector capabilities available on November 2. Availability does not establish access to every provider or permission to collect from every website.
| Capability | Practical use | Boundary |
|---|---|---|
| Exa search and page reading | Discover independent explanations and primary research; read selected source pages | Indexed snippets and partial reads need explicit limits; search does not validate native functionality |
| Scholar paper search and retrieval | Find upstream work on prompt evaluation and instruction following | Bibliographic records are not full-paper evidence; other models' results are not OD results |
| Firecrawl permitted page retrieval and structured extraction | Capture dated source fields, reconcile permitted documents, and inspect Duhployed's own pages | Review source permissions and provider price before collection; structured output still needs comparison with the source |
| vidIQ video search, captions and comments | Find tutorial questions and dated creator demonstrations | Captions may omit controls; comments are reports, not implementation documentation; omit viewer identities |
| GitHub and Drive | Compare exact revisions, schemas and authorized creator documents | Verify the repository's authority and document access; a community export is not an official native import schema |
| Adobe and Magnific | Prepare explanatory illustrations and separate renderer experiments with settings and lineage | An output from another renderer cannot count as a native OD trial; inspect cost and rights before generation |
| Sites | Build, check and publish this existing library | Publishing is a separate result from source verification or native testing |
No OD-specific connector or official API permission was established by this review. No plugin is treated as a way around a website's access restrictions.
Policy and cost checkpoint
The November 1 policy review found restrictions on automated gathering in the Terms of Service and automated scraping in the Acceptable Use Policy. Their displayed modification dates were August 26 and August 21, 2026 respectively. The review did not establish permission for agent-operated account inspection and output collection. Automated OD collection and native testing remain on hold pending explicit platform clarification or permission. These pages were not fetched again for this update.
The worksheet below supports user-operated manual recording. Follow current platform rules, use original adult fictional material, and keep tests private unless deliberately publishing material you have the right to share. Reading a guide or using a different browser does not grant an automation exception.
Before a manual action, inspect its current charge, route, model and batch size. Set an existing-credit ceiling for the whole comparison outside this recorder. Stop when price or access is unclear, or the action exceeds that ceiling. Do not purchase credits or change a plan as an incidental test step. Record the displayed cost separately from an actually confirmed charge or refund. An account balance change with concurrent activity is not an isolated charge measurement.
Plan one interpretable comparison
- Write the question and criterion before seeing outputs. For example: does the reply answer the question while leaving the visitor's actions to the user?
- Give the comparison an ID. Label the baseline and revised prompt variants, and keep the exact wording. Change one intended element; record any other difference you cannot hold steady.
- Record the actual route, model label, batch size, settings and relevant starting context. Leave unknown values blank or explicitly unknown. A fresh chat and a continuing chat are different conditions.
- Assign a trial label to every attempt. Save all outputs and actual errors, including failures. Distinguish repeats of the same prompt from tests of different wording. Keep failed requests separate from completed behavioral trials.
- Grade each criterion from the evidence. Use uncertain when the observation cannot resolve it, and record confounds. Keep a source claim separate from a native observation.
- Report only the scope supported by those records. One run can reveal a defect or demonstrate that an operation completed; it cannot supply a reliability estimate. This workflow prescribes no universal minimum trial count.
For an original still-image exercise, use concise visible facts such as clothing, pose, setting, framing and one lighting setup. Judge identity, clothing, scene, framing and extra people separately. Keep stable appearance distinct from scene wording. Use a separate negative field only if the selected route actually offers it. Do not copy text into an unrelated field to work around a gate.
Record, retain and review
The Test records tool now accepts evidence type and reference, settings, comparison/variant/trial labels, displayed cost, actual charge, cost unit, charge evidence and limitations. It retains exact input, criterion, actual output and verdict. New fields are optional so older version 1 backups remain usable. A blank evidence type does not become a native observation automatically.
Records stay in this browser's local project storage until you choose to download a backup or test-record JSON. They are not uploaded by this tool. Those exports contain the text you entered; review a copy before sharing it. Remove names, email addresses, credentials, account links, unrelated private conversations and private character fields. Keep enough redacted input and context for a reviewer to understand the result. Record a redaction instead of inventing replacement evidence.
The recorder stores your reported observation. It cannot independently verify a pasted output, enforce an external spending ceiling, or guarantee that a prompt will be followed.
Research behind the method
Sclar and colleagues, Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design, study sensitivity to plausible prompt formats. The partial source body read for this review supports keeping exact formats, a defined metric and a bounded evaluation budget. Its numerical results are not transferred to OD.
Lior and colleagues, ReliableEval, describe prompt variation as an evaluation concern. Only the abstract was read here; this guide does not adopt the paper's statistical procedure or a sample-size formula.
The practical comparison steps above are editorial guidance informed by those sources. Neither paper evaluates OD. No new OD control, generation, prompt-efficacy, cost or entitlement result was collected for this update. Earlier native pilot outcomes retain their original dates, cohorts and limitations.