Choose one of three approaches. The steps for that option appear below.
Choose an option above. On a phone, swipe the tabs left or right to see all options →
Claude arranges your checked utterances into a script and blank run sheet. You and the reviewer keep every meaning.
Choose situations and utterances
Arrange the script draft
Approve wording and meanings
Lay out run sheet
Sign off the pack
Use: You
Have ready: your own sample utterances, the language pair and dialect, and the intended setting. Number the utterances U1 onward.
The reviewer checks each utterance is natural and records the intended meaning without relying on an AI translation.
Use: Claude
Ask Claude to group the utterances by test purpose into a draft script table, with gaps marked 'still to confirm'.
What to paste into the tool
Task: A draft test script for a later human-run speech translation check. Use only my supplied utterances and stated setting. Group them by test purpose such as ordinary requests, numbers, negation or ambiguity only where the supplied text supports that label. Output test ID, source utterance ID, exact source wording, speaker cue, meaning to preserve, and a blank reviewer note. Do not add target-language translations, new facts, test results or unsupported claims about a product. Mark gaps: still to confirm. Do not run, record or send anything. Brief and utterances: [paste the brief and numbered utterances].
Use: You
Check the draft against your original utterances and the reviewer's meaning notes; correct cues.
Compare every line with its source. Reject invented words or claims a service passed. The reviewer supplies the expected meanings.
Use: Claude
Ask Claude to lay out a numbered run sheet and blank score table from the checked script and your scale.
What to paste into the tool
Task: A run sheet and blank scoring table for a later speech translation test. Preserve every approved utterance and expected meaning exactly. Use my supplied scale and label unconfirmed criteria: still to confirm. Add columns for test ID, language direction, delivery cue, observed translation, meaning preserved, numbers and units preserved, omissions, and reviewer notes. Leave all observations and scores blank. Do not predict quality, invent timings, translate the answer key or run a tool. Checked script and scale: [paste the checked script, human meanings and chosen scale].
Use: You
Check the run sheet and blank score table against the checked script before signing off.
Check no utterance or meaning changed, IDs align, result cells blank. The reviewer clears the pack; no test has run.
ChatGPT audits your script for coverage gaps and builds a retest pack. The reviewer writes all additions.
Prepare your existing script
Log the coverage gaps
Choose gaps needing additions
Build the retest pack
Close the review
Use: You
Have ready: your own existing script, the setting, language pair and known error concerns.
The reviewer checks the source wording and notes which expected meanings are checked and which are still to confirm.
Use: ChatGPT
Ask ChatGPT to log coverage gaps tied to script IDs and suggest questions for the reviewer, with no new translations.
What to paste into the tool
Task: A coverage-gap log for my speech translation test script. Inspect only the text and concerns below. For each supported concern, name the existing script IDs that address it or write missing. Look for checks involving supplied numbers, negation, names represented by labels, turn changes and unclear wording; do not assert an absent feature is required in every setting. Suggest questions for my bilingual reviewer, not new translations. Do not score a service, claim accuracy or run a test. Script and concerns: [paste the script and concerns].
Use: You
Choose which gaps need new utterances; the reviewer writes them and their expected meanings.
Reject concerns that do not fit the setting. The reviewer writes any additions; no assistant guess becomes an answer key.
Use: ChatGPT
Ask ChatGPT to build a revision log and blank per-gap check sheet from the approved script, gap IDs and rules.
What to paste into the tool
Task: A change-focused preparation packet for checking speech translations later. Match my approved revised utterances to the gap IDs. Output old test ID if supplied, revised test ID, exact approved utterance, gap addressed, approved expected meaning, human acceptance rule and blank observed-result field. Preserve supplied wording and mark unmatched links as still to confirm. Do not make translations, decide pass or fail, run tests or imply performance improved. Revised script, gap IDs and rules: [paste the checked material].
Use: You
Close the review: keep unresolved gaps visible and outcomes blank.
The reviewer checks new meanings and gap links. Keep gaps visible, outcomes blank and approval separate from product decisions.
The script stays fixed. Gemini turns approved meaning units into scoring sheets for two reviewers.
Set the meaning checks
Draft the scoring checklist
Check list, choose rules
Prepare blank reviewer sheets
Approve for later review
Use: You
Have ready: your fixed script, the reviewer's approved meanings and the test's purpose.
The reviewer splits each meaning into units that matter. Never use AI as the language authority.
Use: Gemini
Ask Gemini to turn the approved meaning units and weights into a draft scoring checklist with blank scores.
What to paste into the tool
Task: A reviewer scoring checklist for a future speech translation check. Use only my approved script IDs, meaning units and scoring priorities. Create rows for each unit with the supplied weight, a plain description of what the reviewer must check and blank score fields. If a weight or pass rule is absent write: still to confirm, do not choose one. Do not translate, add factual answer keys or rate a translation service. Inputs: [paste script IDs, checked meaning units and priorities].
Use: You
Check the checklist against the original meaning units and choose the weights and thresholds yourself.
The reviewer checks no wrong number or negation hides behind a vague score. Choose weights yourself; never accept tool recommendations.
Use: Gemini
Ask Gemini to copy the checklist into two blank reviewer sheets, R1 and R2, plus a disagreement log.
What to paste into the tool
Task: Two blank reviewer sheets and a disagreement-log template. Copy my checked scoring checklist, test IDs and order exactly for reviewer labels R1 and R2. Include blank fields for observed wording, each meaning-unit score and notes; in the disagreement log include test ID, unit ID, each reviewer note and a blank human resolution. Do not fill results, choose a winner, change weights, translate or run a test. Scoring checklist and order: [paste the checked scoring checklist and approved order].
Use: You
Check both sheets and the disagreement log before approving the material for a later session.
Check sheets match and fields are empty; a fluent human resolves disagreements. This material is not evidence of translation quality.
Choose A to build a balanced pack from utterances, B to find weak spots in an existing script, or C to turn a settled script into scoring and reviewer material.
Best when: You have examples and want a first test pack
What it does: Claude arranges script and run sheet; reviewers own meanings
Strong at: Each test stays tied to an utterance you supplied
Limit: Needs a fluent person for the answer key
Free plan: Yes, limited (usage caps)
Best when: You have a script but worry it misses cases
What it does: ChatGPT logs gaps; the reviewer writes additions
Strong at: Shows which concerns have no corresponding test
Limit: A gap is a review question, not proof of failure
Free plan: Yes, limited
Best when: Your script is settled; consistent scoring matters
What it does: Gemini builds scoring sheets; reviewers score separately
Strong at: Uses human-selected meaning units, not an AI key
Limit: Adds no missing speaking situations to the script
Free plan: Yes, limited (prompt caps)
Before you paste into any AI tool, remove any confidential information. Put it back at the end.
Deciding between ChatGPT and Claude for this? See the comparison.
Review every assistant reply before it enters the next step.
Leave results blank; no live translation check has been run.
Check the result against your original inputs. Approve it before sharing or acting.