About CELPIP Writing Coach
An independent study site for the writing component of the CELPIP-General test: original samples with level breakdowns, practice prompts in the official format, and a free AI scorer built to be conservative.
Updated October 20, 2026
What the site is
CELPIP Writing Coach helps people preparing for CELPIP writing see what each level looks like and get feedback on their own writing without paying for a tutor. Everything on the site is original: the prompts, the sample answers, the level breakdowns and the guides were written for this site against the official test format and the official performance standards. No test questions, no texts from other sites.
How the samples were made
For each prompt we wrote two answers: one at a high level (9 to 11) and one at a mid level (5 to 7), with the typical errors and limitations of that level, then rated both on the four official categories and explained the ratings. The levels are our judgement, informed by the official descriptors; they are not official scores. The same answers were used as the first calibration set for the scorer.
How the scorer works
The scorer uses gpt-oss-120b (OpenAI), Cloudflare Workers AI. The model rates the four categories, lists errors with the exact passage and a rule, and quotes the evidence for level-9 features. Code then verifies every quote against the text, caps the level on objective measures (error density, missing parts of the task, length) and only allows level 10 and above with verified evidence. The method is described on the scoring page.
Calibration log
We publish every calibration run here. "Within 1" means the scorer's level was within one level of the level we assigned to the answer; "bias" is the average difference (negative means the scorer is stricter than we are).
| Date | Answers | Exact | Within 1 | Bias | Notes |
|---|---|---|---|---|---|
| 2026-10-20 | 28 | 7 (25%) | 25 (89%) | −0.07 | Launch configuration (two readings per text, lower one kept, no promotion by code). The three answers outside ±1 are two level-11 answers and one level-10 answer, all estimated at 8–9: the scorer does not reach the top of the scale. Mean absolute error 0.86 levels. |
| 2026-10-20 | 28 | 13 (46%) | 25 (89%) | −0.04 | First run, single reading per text, before the rules were finalised. Same three high answers outside ±1. |
What the site is not
- Not affiliated with, endorsed by or connected to Paragon Testing Enterprises, Prometric, or Immigration, Refugees and Citizenship Canada. CELPIP is a trademark of its owner, used here only to name the test.
- Not an official rater. Estimates can differ from official results, especially at the top of the scale, where the scorer is conservative by design.
- Not a source of test questions. The only official practice material is the free sample test on celpip.ca.
Contact
Corrections, questions and official-score reports (which we use to recalibrate and never publish) are welcome through the contact page.