Turn marketing judgment calls into Jev jobs

Read shared.md first. Save the job file with kind job.

Jev (by TypeSafe) is a decision model. It never writes. You give it text (the state) and closed questions. It returns a pick, a yes probability or a rubric score, each with a confidence. In a check on synthetic rows it took about a third of a second per call and about $0.03 per 1,000 short rows, so a judgment a team makes by hand on a sample can run on every row.

The script sits beside this file. Run it by its full path: python3 "<this skill's directory>/scripts/jev.py" <command> .... Resolve the directory from where you loaded this SKILL.md; do not rely on a variable for it. For API details the script does not cover, read https://docs.typesafe.ai/llms.txt.

The rule: Jev reads meaning, code does numbers

Split every workflow before building anything.

Never ask Jev to count, compute or compare dates; TypeSafe lists these as weak spots (https://docs.typesafe.ai/model-jaggedness/jev-1.13.md). Never ask it for a rationale.

The split is per row. "What share of these posts mention pricing" is a Jev question on each post (yes, no, unclear) plus a code count of the yes answers. Use assets/jobs/content-mentions.json.

Asked to write something? Say Jev cannot write. Have a writing model draft, then offer a Jev job that scores the drafts and keep the top ones.

Workflow

Copy this checklist and tick it off:

Jevify progress:
- [ ] 1. Pick the discipline and read its file (table below)
- [ ] 2. Map the workflow: mark each judgment Jev, code, writer or person
- [ ] 3. Pick one Jev judgment: many rows, options a person could apply the same way
- [ ] 4. Write the job file (copy a ready job from assets/jobs/; fill every REPLACE)
- [ ] 5. validate the job against the input file; fix every error and warning
- [ ] 6. Sample run on 20 rows; read every answer yourself
- [ ] 7. review, a person labels the sheet blind, calibrate, set the cutoff
- [ ] 8. Full run
- [ ] 9. Report the result in marketing terms

If a sample answer looks wrong, fix the question (usually the option descriptions) and repeat steps 5 and 6. Typed output guarantees the format, not the truth.

Save the job file you wrote with python3 "KIT" save TASK_ID job < JOB.json. Results land next to the input CSV in the project, not in .opascope-work.

Disciplines

Read only the file the task needs. Before writing questions, read references/writing-questions.md. Before setting a cutoff, read references/calibration.md.

Running it

python3 "SCRIPT" validate JOB --items rows.csv
python3 "SCRIPT" run JOB rows.csv --limit 20          # the sample
python3 "SCRIPT" review rows.JOBNAME.csv --question QID
python3 "SCRIPT" calibrate rows.JOBNAME.review-QID.csv rows.JOBNAME.csv --question QID
python3 "SCRIPT" run JOB rows.csv                     # the full run

SCRIPT is the full path to scripts/jev.py. JOB is a path, or the name of a ready job in assets/jobs/. validate works without --items; with it, it also checks the columns.

run prints a cost estimate first and refuses above --max-cost (default $1). Quote that estimate and the cost the run reports; never guess a cost. Jev is billed per input token and output is free (https://docs.typesafe.ai/models.md), so longer rows cost more. In a check on synthetic rows, 1,000 pages of 3,000 characters cost about $0.05.

Results are written as rows.JOBNAME.csv: the original columns plus, per question, the answer, its confidence and its route (auto or review). A choice answer is the option name. A score answer runs from 0 to the number of levels minus 1. A noul answer is the yes probability.

Keys: OPENROUTER_API_KEY (https://openrouter.ai/keys) is the default route, through OpenRouter to the latest Jev. TYPESAFE_API_KEY is used when it is the only key set. With neither, stop after writing and saving the job file, and tell the user which key to add.

Reporting

Lead with the marketing result, then the numbers behind it. A made-up example of the shape:

Sorted N search terms by buyer intent for $X. A look like negatives; B are unsure and need a person. Cutoff C, set from a 20-row blind check (K of 20 right).

Always include rows run, the share sent to review, the cost, and how the cutoff was set. Claim no accuracy beyond the check that was actually run. If no one has labeled the check yet, say the cutoff is unchecked.

Follow shared.md's Asking and Reporting sections.

Keep out

• Personal data: names, emails, phone numbers, anyone under 18. Strip them in code first. Zero data retention is an enterprise option (https://docs.typesafe.ai/legal.md), not the default.
• A model scoring its own fit for a use case is not evidence that it fits.