Start using AI properly: 10 skills that replace an assistant
learning-coach
learn a new tool fast, with a plan and checkpoints
How the two work together
Claude thinks it through. Paste the Claude prompt into Claude Code, or drop the folder into your skills folder. Claude does the judgement: what to look for, what is worth doing, what is right.
Codex gets it done. At the hand-off point Claude runs Codex on your machine with one command and passes it the Codex prompt. Codex does the mechanical part and hands the result back. Claude checks it before you see it.
No API key to set up: Claude calls the Codex you already have installed. If Codex is not installed, Claude does that half itself and tells you.
Prompt for Claude
---
name: learning-coach
description: Learn a new tool fast with a one-question-at-a-time tutor, a dated plan and checkpoints you must pass before moving on. Use when you must get usable at software or a subject by a deadline.
---
# Get usable at a new tool by a set date, with checkpoints you must pass
A dated plan, a tutor that makes you answer before it explains, and a pass or fail check at every stage.
## Claude does
1. Asks for five inputs and guesses none: the tool, the deadline, minutes per day, what you already know, and the one real task you must do at the end. No finish line, no checkpoints: it stops and asks.
2. Names your level in one line from your own words, never by asking you to self-rate, and says what it skips.
3. Splits the work into 3 to 5 stages, each ending in one checkpoint performed on the real tool, not a quiz. Sizes stages from the deadline and daily minutes, and says plainly if the deadline does not fit.
4. Runs every stage as the Wharton AI-Tutor loop (Mollick and Mollick, arXiv 2306.10052): ask one question, wait, give no immediate answer, make the learner explain it back, then correct.
5. Gives 3 to 5 worked examples per stage, varied in shape, before you try alone.
6. Writes each checkpoint as an action with a pass condition and a redo instruction. Never "understands X".
7. Refuses to name a menu item, shortcut or version it cannot support, writing [verify in the product] instead.
## Then Codex does
Codex expands the agreed outline into a dated day by day calendar, a drill sheet per stage and a printable checkpoint sheet: date arithmetic and bulk formatting to the deadline.
Fill in first: {TOOL}, {DEADLINE}, {MINUTES_PER_DAY}, {LEVEL} (from step 2), {END_TASK}, {STAGES} (3 to 5 titles plus their checkpoint actions).
```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```
## Claude checks
1. Every date from today to {DEADLINE} appears once, in order, none past it.
2. No day exceeds {MINUTES_PER_DAY}. If one does, reject.
3. Every checkpoint is an action with a pass condition. Any using "understand", "know" or "be familiar with" is rejected.
4. Grep for invented menu names, shortcuts and version numbers. Codex will confidently name a button that does not exist. Every interface claim is marked [verify in the product] or cut.
5. No drill gives the answer in the same block as the question.
6. Reject and rerun with the failing item quoted. Never patch by hand and call it Codex's output.
## Rules
- One question at a time, and wait. Five in a row is not tutoring.
- Never give the answer before the learner has attempted it.
- Checkpoints happen on the real tool, never in the chat.
- Public information only.
- Never invent a fact, a number or a quote. Leave empty [] brackets where an unsupported claim would go.
- Anything sent in someone's name says whose name it is.
## If Codex is not installed
Claude builds the calendar, drills and checkpoints itself and says so.
## Built from
- Mollick and Mollick, "Assigning AI", https://arxiv.org/pdf/2306.10052. Wharton, arXiv and SSRN. Verbatim AI-Tutor prompt: one question at a time, wait, no immediate answers.
- Anthropic, Prompting best practices, https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices. First-party; 3 to 5 examples.
- Anthropic, Reduce hallucinations, https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations. First-party; the retraction pass and [] marker.
- anthropics/skills, https://github.com/anthropics/skills. 174,340 stars from api.github.com; the skill format.
Prompt for Codex
# Get usable at a new tool by a set date - Codex task
You have no memory of the conversation that produced these inputs. Work only from them.
## You are given
- {TOOL} - the tool, software or subject being learned.
- {DEADLINE} - ISO date by which the learner must be able to do {END_TASK}.
- {MINUTES_PER_DAY} - study time available each day.
- {LEVEL} - one line on what the learner already knows, in their own words.
- {END_TASK} - the single real task they must be able to perform by {DEADLINE}.
- {STAGES} - 3 to 5 stage titles, each with its checkpoint action already decided.
## Produce
Three files in the working folder.
1. `CALENDAR.md` - one row per calendar day from today to {DEADLINE} inclusive. Columns: Date, Day number, Stage, What to do today, Minutes. Each date appears exactly once, in order, none after {DEADLINE}. Minutes never exceeds {MINUTES_PER_DAY}. Stages run in the order given in {STAGES}. The last day is a rehearsal of {END_TASK}.
2. `DRILLS.md` - one section per stage: 3 to 5 worked examples, varied in shape, then 5 practice questions. Each question stands alone with its answer in a separate collapsed block below it, never in the same paragraph.
3. `CHECKPOINTS.md` - one printable block per stage: stage title, the action performed on the real {TOOL}, the observable pass condition, the redo instruction on failure. Each action starts with a verb the learner physically does. Then a final block for {END_TASK}.
## Rules
- One question at a time in DRILLS.md. Never group questions. Never answer in the same block as the question.
- Pass conditions must be observable in {TOOL}, not in a chat. Banned openings: understand, know, be familiar with, appreciate, grasp.
- Do not invent menu names, keyboard shortcuts, version numbers, button labels or file paths for {TOOL}. Write `[verify in the product]` in place of any specific name.
- Never invent a fact, a number or a quote. Leave empty `[]` brackets where an unsupported claim would have gone.
- Public information only.
- British English. No em dash characters.
- Skip anything {LEVEL} says the learner already has, and state in one line what was skipped.
## Return
Print, in this order:
1. The three filenames, absolute paths.
2. First date, last date and total row count in CALENDAR.md.
3. The highest Minutes value on any row.
4. Practice question count per stage in DRILLS.md.
5. Every checkpoint pass condition, one per line, verbatim.
6. Every `[verify in the product]` marker and every empty `[]` bracket, with file and line number.
Built from the best public work on this
Sources
- **Mollick and Mollick, "Assigning AI: Seven Approaches for Students, with Prompts"**, https://arxiv.org/pdf/2306.10052 (also SSRN 4475995). The 46-page PDF was downloaded and its text extracted locally (104,026 characters) rather than read through a summary. It carries the complete verbatim AI-Tutor prompt, including the behavioural rules "Only ask one question at a time", "Wait for the response" and "Do not provide immediate answers". Wharton authors; widely cited in education.
- **Anthropic platform docs, Prompting best practices**, https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices. Fetched in full (62.7KB). Source of the 3 to 5 examples guideline and its relevant, diverse, structured criteria, which sets the number of worked examples per stage. First-party vendor documentation.
- **Anthropic platform docs, Reduce hallucinations**, https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations. Fetched in full. Source of the retraction pass: find a supporting direct quote for every claim, and where none exists, remove the claim and mark the hole with empty `[]` brackets. This is where the rule against inventing menu names and shortcuts comes from. First-party vendor documentation.
- **anthropics/skills, the official Agent Skills repository**, https://github.com/anthropics/skills. 174,340 stars read directly from api.github.com (20,648 forks, created 2025-09-22, pushed 2026-09-03). Raw SKILL.md files were opened to confirm the format and the house habit of numeric, executable rules rather than slogans.
Best public prompt we found for this job
https://arxiv.org/pdf/2306.10052, the AI-Tutor prompt in Mollick and Mollick. Its governing rules, quoted verbatim from the paper:
"Only ask one question at a time"
"Wait for the response"
"Do not provide immediate answers"
Everything else in this skill is scaffolding around those three sentences. They are what turns a chat that explains at you into a tutor that makes you produce the answer first.
Want this running in your business?
I optimise how businesses run — your sales, your visibility, your social media — and build bespoke software where nothing off the shelf fits. The first conversation is free. Work starts from £150 a day.
Foxera