English | 简体中文

Learning Edge

An agent skill that finds the exact edge of what you already know, then teaches only what lies between that edge and a concrete goal: a prerequisite map built from your own material, short probes instead of "do you know X?", a climb in waves, spaced reviews on real dates, and a goal test at the end.

Ask an AI agent to teach you something and it usually starts at a guessed level, explains the subject from memory instead of from your book or repo, and moves on as soon as the explanation is clear. Learning Edge gives the agent a disciplined method and a small tool. It turns your goal into a test you will be able to pass, maps the concepts that test depends on (with the page, section or file each one comes from), locates where your real knowledge stops with a handful of two-minute tasks, and plans the shortest honest climb from there.

It works for books, courses, papers, exam topics, new tools and unfamiliar codebases, in English and Chinese.

Features

• Goal as a test. Every plan ends with a concrete task that proves the goal, agreed up front.
• Map from your material. 8–25 concepts, each with a source locator in your book, notes, docs or code.
• Honest links. Every "A needs B" is tagged cited, inferred or assumed, and the plan lists the assumed ones for you to confirm.
• Probes, not self-ratings. Short produce-or-predict tasks graded solid, shaky or missing, with a note on exactly what went wrong.
• Adaptive probing. The tool picks the probe that settles the most of the map; a solid answer marks its prerequisites as assumed known, a missing one blocks everything above it. Six to ten probes usually locate the edge.
• Climb in waves. Only what lies between the edge and the goal, in dependency order, with shaky concepts repaired first.
• Session loop. Bridge from something known, a worked example from your material, a fresh variant to do, a three-sentence explain-back, then record.
• Spaced review calendar. New concepts packed into your daily minutes, reviews on day 1, 3, 7 and 21 after each one.
• Misconception log with a correction and a re-test question for every wrong belief.
• Mermaid map coloured by status: known, assumed known, to repair, to learn, goal.
• Zero dependencies. One Python 3.8+ script, standard library only.
• Bilingual. Plans in English or Chinese ("lang": "zh-CN").

Installation

Clone the repository into your agent's skills folder. The folder name should be learning-edge.

Claude Code

# personal (all projects)
git clone https://github.com/fangdididi/learning-edge ~/.claude/skills/learning-edge
# or for one project
git clone https://github.com/fangdididi/learning-edge .claude/skills/learning-edge

OpenAI Codex CLI

git clone https://github.com/fangdididi/learning-edge ~/.codex/skills/learning-edge

Cursor

git clone https://github.com/fangdididi/learning-edge .cursor/skills/learning-edge

Gemini CLI, OpenCode, Windsurf and other agents

Clone into the agent's skills directory if it has one (for example .agents/skills/). If your agent has no skill system, add a line to AGENTS.md, CLAUDE.md, GEMINI.md or your rules file:

When I ask to learn or study something, follow ./learning-edge/SKILL.md and keep the map with its script.

Requirement: Python 3.8 or newer for the helper script. The method itself works without it.

Usage

Just ask to learn something, ideally with material and a goal. The skill activates on phrases like "teach me", "help me learn", "where do I start", "get me up to speed", "教我", "从哪开始学", "帮我入门", "备考":

Get me up to speed on this repo so I can fix the retry bug in worker/retry.py. I have an hour a day.
我想读懂现金流量表，教材第 3–5 章和一份年报都在这个文件夹里，每天 45 分钟。
Teach me Bayes' rule well enough to solve medical-test problems. No material, just start.

The agent agrees a goal test with you, reads your material, drafts the map, gives you a few probes, and replies with your learning edge, the first two sessions and the path to the plan file. After each session it records the result and re-plans.

You can also run the tool directly:

python3 scripts/edge.py check examples/retry-bug.json                  # validate and warn
python3 scripts/edge.py probe examples/cash-flow.zh-CN.json --top 3    # what to probe next
python3 scripts/edge.py mark  examples/cash-flow.zh-CN.json depreciation solid --note "timing explained"
python3 scripts/edge.py plan  examples/retry-bug.json -o retry-bug.plan.md
python3 scripts/edge.py graph examples/retry-bug.json                  # Mermaid only

Examples

Choosing the next probe on a fresh map (Chinese finance example, before any probing beyond the basics):

6 concept(s) still undetermined. Probe next:
- indirect-method (间接法（从净利润推到经营现金流）)
    probe: 给你净利润、折旧、应收变动、存货变动、应付变动五个数，算出经营活动现金流净额……
    if solid: settles 1 prerequisite(s); if missing: blocks 3 dependent(s)
- fcf (自由现金流)
    probe: 用年报第 89–90 页的数字算出公司 2025 年的自由现金流……
    if solid: settles 3 prerequisite(s); if missing: blocks 1 dependent(s)

The start of a generated plan (English codebase example):

## Where you stand
- known: 1
- assumed known (not probed): 2
- to repair: 1
- to learn: 5

## Your learning edge (ready now)
- Idempotent jobs: Name two side effects in send_invoice() that would repeat on a retry…
- Exponential backoff with jitter: Without notes: base 2 s, cap 60 s, full jitter…
- Wall clock vs monotonic clock: Explain why time.time() is the wrong tool for a 30 s timeout…

Reading it: tracing a job through the worker loop came back solid, so the event loop and leases are assumed known and listed for one quick check. Backoff and the monotonic clock are missing, which blocks the retry state machine, so the climb starts at the edge and reaches the fix in three waves over four days.

Example	Language	Learner situation
examples/retry-bug.json → .plan.md	English	New to a codebase, needs to fix one bug
examples/cash-flow.zh-CN.json → .plan.md	Chinese	Self-study of financial statements from a textbook and an annual report

File structure

learning-edge/
├── SKILL.md                         # the skill (English), loaded by the agent
├── SKILL.zh-CN.md                   # the skill (Chinese)
├── README.md                        # this file
├── README.zh-CN.md                  # Chinese README
├── LICENSE                          # MIT
├── scripts/
│   └── edge.py                      # check / probe / mark / plan / graph
└── examples/
    ├── retry-bug.json               # English map: learning a job-queue codebase
    ├── retry-bug.plan.md            # its plan
    ├── cash-flow.zh-CN.json         # Chinese map: reading a cash flow statement
    └── cash-flow.zh-CN.plan.md      # its plan

Principles

1. Start from what the learner can do, not from what they say they know.
2. Teach from their material, in its words, with its page numbers.
3. Every link in the map says why it is there.
4. Explaining is half a session; doing and explaining back is the other half.
5. Done means the goal test passed, and the reviews stay on the calendar.

License

MIT © 2026 fangdididi