loooom

Gym log.

Compare saved answers before and after each skill change.

JSON file. Nothing is uploaded.

absorbRepair a failed answer.

Deliberately added defect

An instruction we added on purpose made the AI reveal quiz answers before the learner could try.

Choose a test round
Original957 words
Read this session

Fresh final tests

Selected before testing
Question
Original 3/12Revised 12/12
Start a recall roundTraining question
Keep the first step smallTraining question
Give feedback after an attemptTraining question
Transfer to a new subjectHeld out of editing
PassFailSelect a result to compare the two answers.
The repair passed every final test.

These are new responses, separate from the editing rounds.

What changed

Replaced that instruction: wait for an attempt, then give feedback. Keep each practice step small when needed.

3 answers per question · gpt-6-astraAI-graded. A small test, not a guarantee.

Session history

6 saved sessions · latest completed first

  1. Shortened focus to 242 words

    Passed final testsfocus
    Problem
    The skill had 948 words. The first cut still failed time-budget and task-switching checks.
    Change
    Removed repeated explanations. Kept clear rules for the full time budget, a concrete task, and necessary calls. Rejected the first cut, then kept two better revisions.
    Result
    Final tests: 15/15 before, 15/15 after.
    Session details

  2. Stopped absorb from revealing answers too soon

    Passed final testsabsorbDeliberately added defect
    Problem
    An instruction we added on purpose made the AI reveal quiz answers before the learner could try.
    Change
    Replaced that instruction: wait for an attempt, then give feedback. Keep each practice step small when needed.
    Result
    Final tests: 3/12 before, 12/12 after.
    Session details

  3. Checked absorb and kept it unchanged

    Unchangedabsorb
    Checked
    We checked whether the published skill let learners try recall before getting feedback.
    Change
    No edit. The first tests found no mistake to repair, so we tested the same skill on new questions.
    Result
    9/9 first tests passed. Then 6/6 new tests passed.
    Session details

  4. Rejected a cut that broke the time budget

    Not keptfocus
    Problem
    We wanted to shorten focus without making its advice worse.
    Change
    Cut the skill from 948 to 157 words by removing explanations and the worked example. The new version planned 25 minutes of activity inside a 20-minute limit.
    Result
    3/3 tests passed before, 2/3 after. One answer per question; no separate final check.
    Session details

  5. Tried the first repair for early quiz answers

    Kept after one passabsorbDeliberately added defect
    Problem
    An instruction we added on purpose gave away the answer before the learner could respond.
    Change
    Changed the quiz rule to wait for the learner’s answer and keep each step within the available time.
    Result
    1/4 tests passed before, 4/4 after. One answer per question; no separate final check.
    Session details

  6. Ran the first check on absorb

    Unchangedabsorb
    Checked
    We checked for answers shown too early, too much information, and poor feedback.
    Change
    No edit. The original skill passed the checks.
    Result
    4/4 tests passed before. One answer per question; no separate final check.
    Session details