v0.6.0 A skill for Claude Code and other AI agents
Don't read about it. Predict, try it, then check.
A tutoring skill for any subject. It doesn't explain topics to you. It makes you predict, try it yourself and check, and it keeps a record of what you actually know.
Works with Claude Code, Codex, Cursor, Gemini CLI, Copilot, OpenCode and plain chat.
Step 1 · PredictPlaying itself. Tap anything to take over.
You're writing to a colleague who's been ill. Which verb goes in the gap?
Espero que bien.
How sure are you?
Pick an answer and how sure you are.
Right · Fairly sure
Answer: B, estés. A wish about someone else isn't stated as fact, so the next verb leaves the present.
Logged with your confidence, so it comes back in review.
One module
Every module runs the same loop.
Each stage has its own reference file the agent re-reads before running it, and every check feeds the review log that brings ideas back as warm-ups.
Warm-up
What's due comes first: ideas from earlier modules, scheduled to return just before you'd forget them.
Spot
Each module is anchored in something you already do. The tutor starts from your own work.
Probe
It starts with high-level questions and walks down until it finds the edge of what you know.
Map
What you hold is marked and skipped. The steps left form a dependency graph to the module's goal.
Steps
Predict, try, check. At most one prediction and one check per step. A missed check gets repaired before anything builds on it.
Close
Break it, Plug back and Explain back, graded strictly, then the map redrawn from memory. A module isn't done without all three.
Warm-up
Spot
Probe
Map
Steps
Close
Due today
1.1Espero que ___ bien
2.3Doubling n halves the error?
3.2Article 231, in its own wordsin 9 days
Scheduled with FSRS from your review log.
From your work
To: Lucía
Hola Lucía, espero que estás bien. Te escribo por el informe…
Start here: the verb after espero que.
Probe
Does que join two clauses?
Does the present state a fact?
Does a wish state a fact?
Edge found
Map v1
Two held ideas skipped. Four steps left to the goal.
PredictYou try itCheck
A procedure you can't do yet comes as a faded worked example.
Module 1.1 close
Break itChange one thing and explain what happened
Plug backOne rule for your own work
Explain backThe goal in your words, then the map from memory
The module map
Your gaps, drawn as a map.
Two to four probing questions find the edge of what you know. What you hold is marked and skipped; the steps left lead to the module's goal, and each one closes only when its check passes.
Probe
Two answers come back solid. Those ideas are already yours, so they're marked as held.
Map v1
The steps left are drawn as a dependency graph, ending at the module's goal.
Walk it
Each step closes only when its check passes. The map shows where you are.
Redraw it
At the close, before seeing it again, you redraw the map from memory. Each step you miss or misplace is logged for review.
Known beforeCheck passedYou're hereStill to do
Confidence
Is "Certain" actually right?
Confidence is recorded with every answer, because a confident miss is the most useful thing a question can find. Those are the beliefs that quietly break your work. They're named, traced, repaired and retested later.
The recap page charts how often you're right at each confidence level, and how that's changing. Example numbers below.
Right at each level
Certain: said 95%, right 58%
Illustrative curve. FSRS picks the real dates from your review log.
Spaced review
Back just before you'd forget.
Every check is logged and scheduled with FSRS, so ideas come back just before you'd forget them. Ask "What's due?" or "Quiz me" and earlier modules return as warm-ups.
Grounding
Sourced, not remembered.
A model's memory is stale and unlabelled. Before planning, a separate agent finds the authoritative sources. Before each module, every claim it rests on is checked against them, with a quote and a date. What can't be supported isn't taught as fact, and your own result beats any source.
From a test topic on Python concurrency: claims checked before a module.
Python 3.13 removed the GIL; threads run CPU-bound code in parallel by default
What's New in Python 3.13“experimental support for running in a free-threaded mode”
Contradicted
datetime.utcnow() is the recommended way to get UTC time
datetime docs“Deprecated since version 3.12”
Contradicted
The thread pool's default worker count depends on the CPU count
That makes each claim supported, dated and correctable. It doesn't make it infallible.
Any subject
Same loop, any subject.
What changes is the try, the thing you do between predicting and checking. Questions, confidence, misconceptions, spaced review and the module close work the same everywhere.
Predict
Adding one tool to the end of the list keeps every message cached.
You try it
Run cell 2.3 with and without the tool; compare the reused token counts.
Check
The shared run ends inside the tools block. Prediction wrong: the order is one sequence.
Visual pages
Pages it builds, live.
Real outputs of the skill's scripts. Each page is generated from your markdown records and can be thrown away.
The tutor offers a single-purpose page with real measured data, and builds it only if you say yes. It asks the prediction first.
Honest status
Where it is today.
Claude Code is the primary target and the most tested. Elsewhere the install and scripts are tested; a full session there isn't yet, so reports are welcome.
5
Tested
Full sessions as a Claude Code plugin
Spaced review scheduled with FSRS
Your own sources: textbook, notes, past exams
Skills CLI install and scripts: Codex, Cursor, Gemini CLI and others
Install it, then ask for what you want to learn. Your records are plain markdown in a folder you own, so switching agents, devices or models loses nothing.
Claude Code
/plugin marketplace add ITZSHOAIB/mastery-tutor
/plugin install mastery-tutor@mastery-tutor
Other agents
npx skills add ITZSHOAIB/mastery-tutor
Anything else
Copy skills/mastery-tutor/ into your agent's skills folder, or paste SKILL.md into a chat.