Learning Principles: Turn Effort into Verifiable Learning

Learning is easily mistaken for two things: spending a lot of time or feeling familiar with material. Finishing a video, highlighting a chapter, or understanding an answer can create “I know it” confidence. When the source closes, the cue disappears, and the task changes, another action must answer whether the ability remains.

I once confused visible effort with learning. At school I woke at 5:20, skipped midday rest, and read after lights-out. Time increased while performance fell. I interpreted failure as “not pushing hard enough” and added more. Later I could admit that, for me, chronic sleep loss, no error diagnosis, and mechanical hours made an increasingly expensive path that did not improve.

That experience does not prove early rising is ineffective or prescribe a schedule. It leaves a question worth retesting: when investment increases, do task performance, error type, retention, and physical capacity improve together? If not, add diagnosis before adding hours.

This chapter is the book's general protocol. It does not turn life into a laboratory. It helps you preserve enough evidence within finite energy: define a real delivery, save the unaided first version, use targeted input and retrieval, receive specific feedback, remove cues after a delay, change one condition, and keep the state outside any single conversation.

Chapter at a Glance

• Turn “I want to improve” into an audience, action, quality threshold, deadline, and evidence.
• Separate research findings, personal experience, and hypotheses waiting for a test.
• Preserve an unpolished first version instead of letting a final product erase the start.
• Build one learning loop from input, retrieval, output, feedback, and a state record.
• Classify errors by task impact rather than blaming laziness or vocabulary size.
• Make feedback change the next action instead of adding notes and advice.
• Let performance set spacing while protecting sleep, body, and a returnable minimum capacity.
• Keep goals, samples, errors, and next steps in a state file outside the session.
• Test transfer under a changed topic, audience, mode, or time limit.

1. Define One Real Delivery

“Learn English,” “learn programming,” and “communicate better” are too large to choose today's action. Name the conditions in which language or knowledge must work:

Element	Question	Example
Audience	Who will understand, decide, or act from it?	New colleague, client, teacher, self
Action	What should the person do after reading or hearing?	Retell, reply, run, choose, continue
Input	Which concrete source will you process?	Two-minute audio, API page, email
Quality threshold	What counts as done, and what risk is unacceptable?	State gist and owner; preserve limits
Conditions	What are time, device, background, and permitted tools?	20 minutes, no captions, dictionary allowed
Evidence	What work, test, or response will remain?	Recording, retelling, run result, reply

Write the goal like this:

In 12 weeks, I can deliver a five-minute project update covering context, progress, and risk. Listeners can restate the key points without switching language. I will keep recordings and feedback from weeks 1, 4, 8, and 12.

It guides action better than “move from B1 to B2”. A level describes performance under some conditions; it cannot replace a context, task, deadline, and evidence.

2. Separate Three Kinds of Claim

This guide keeps different evidence in different places:

• Research evidence reports observations under specified samples, conditions, and measures; it suggests what to try, not what everyone will achieve.
• Personal experience explains why I chose something; it is a narrative and a clue for your experiment, not a general rule.
• Testable hypothesis is a judgment without enough evidence; turn it into an action, counterexample, or stop condition.

When advice omits its population, conditions, outcome, and source, return it to the hypothesis column. Plausibility is not fit for your task.

3. One Complete Learning Loop

For 25-45 minutes, move through six steps:

1. Define the task, audience, conditions, and acceptance standard.
2. Attempt it independently with the minimum permitted help and save the raw sample.
3. Diagnose whether the failure is language, structure, background, strategy, attention, capacity, or device.
4. Target input only at information needed for the current error.
5. Produce actively by retelling, writing, speaking, running, or deciding after closing the source.
6. Record feedback and state: result, help received, unknowns, and one condition to change.

“I watched it” is not a seventh step. The loop ends when you can complete a related action after the tool or source leaves, or can state more accurately why you cannot yet.

4. A First Version Is a Measure, Not a Verdict

The first version need not be attractive, but it must be comparable:

Real task and audience:
Material, version, and source:
Start/end time:
Permitted dictionary, captions, search, AI, and human help:
My acceptance standard:
First-version location:
What I completed then:
What remained uncertain:

Do not replace the first version with a later polished one or call one failed attempt stable ability. It answers: what could I do independently under these conditions?

For four English skills, use the English Diagnostic for listening, reading, speaking, and writing. Use a skill card when a bottleneck needs finer resolution.

5. Diagnose the Error Instead of Judging the Person

A weak result can come from different layers:

Layer	Typical sign	Next question
Language form	Form, sound, grammar, or spelling was not recognised	Which form changed the meaning?
Structure and meaning	Words are known but participants, time, condition, or relation were not parsed	Which relation cannot I explain?
Background and concept	Language is readable but the object or premise is unknown	Which domain fact is missing?
Strategy and tool	The same place is looked up repeatedly or a cue cannot be removed	Which scaffold stays, and when is it removed?
Attention and capacity	Errors cluster later and the body drops sharply	What minimum task is sustainable today?
Environment and device	Sound, network, font, or interface changes the result	Does performance change with another setup?

“Careless,” “not talented,” and “too little vocabulary” can end diagnosis too early. Preserve the error, choose one changeable condition, and test it. Diagnosis finds the next lever; it does not write a character judgment.

6. Retrieval Is Stricter Than Familiarity

Repeated exposure can make material familiar without making it available when needed. Retrieval includes:

• writing a three-sentence summary after closing the source;
• stating gist, detail, responsibility, and next step after listening;
• explaining a concept without notes;
• placing a new chunk in two new situations;
• reproducing minimum run steps without example code;
• asking a reader or peer to retell what they actually received.

Retrieval need not be so difficult that it only measures frustration, background, or capacity. The useful difficulty exposes a gap while leaving a route to repair. Record completion, cue, time, error, and response rather than “I worked hard”.

7. Feedback Must Enter the Next Action

Feedback is not better in larger quantities. Start with the one to three items that change the task:

1. Did the audience understand, decide, or complete the intended action?
2. Are facts, structure, and responsibility accurate?
3. Which recurring error deserves repair in the next attempt?

Sentence-level feedback should distinguish necessary correction, ambiguity, register choice, acceptable variation, and style preference. Ask for a retelling before discussing whether something sounds natural. A reviewer may reveal a blind spot while also carrying different knowledge and aims; accept, partly accept, reject, or defer with a reason.

If feedback only beautifies an old draft, it is not transfer. Turn it into a condition for a parallel task: change topic, audience, sound, time, or constraint, and change one main variable.

8. Spacing, Interleaving, and Sleep

Retrieval and distributed practice have substantial support in learning research, but no fixed calendar fits every learner and material. Use same-day, day 1, days 3-7, day 14, and day 30 as reminders, then adjust:

• easy, accurate recall plus new-context use: lengthen the interval or add variation;
• effortful success: keep a similar interval;
• repeated failure: reduce material, shorten the interval, or repair prerequisites;
• recognition without use: add production and cross-mode tasks;
• sustained physical decline: downgrade, pause, or seek help instead of calling recovery laziness.

Interleave skills when useful, but do not change every variable at once or the cause of change becomes unknowable. Sleep, food, movement, and unassigned time are inputs to memory, attention, and return. Persistent sleep, mood, or physical problems need qualified professional care; this book does not diagnose.

9. Keep State Outside the Session

Chats, courses, and apps end. Learning state cannot live only inside them. Use the Learning State to keep:

Current main question:
What I know with evidence:
First and latest result:
Error categories and unresolved risks:
Tools used and verification boundary:
Smallest next task:
One condition to change next:
How I return within 24 hours after interruption:
Handover owner and review date:

At the next session, read the state and act. Do not ask a model to infer progress from an old chat. A state file is a fact page, not a performance diary.

10. AI Is a Scaffold, Not the Result

AI can question, explain, classify errors, generate counterexamples, simulate a reader, offer parallel tasks, and compare versions. Use an auditable order:

1. Save your task and first version.
2. Give the model only the minimum material needed for diagnosis.
3. Ask it to state uncertainty, sources, and possible misjudgments.
4. Return to primary material, dictionaries, run results, teachers, peers, or real audiences.
5. Close the tool and perform independently under a changed condition.
6. Write acceptance, rejection, and the next step into the state file.

The model cannot prove facts, carry privacy responsibility, complete an exam that forbids outside help, or decide whether you actually learned. Do not paste customer data, unpublished documents, identity records, medical privacy, or unauthorised third-party content.

11. Make Ability Move Across Conditions

Immediate performance describes one condition. Change at least one for transfer:

Keep	Change	Observe
Task	Topic	Do structure and strategy still work?
Topic	Audience	Do register, background, and explanation adjust?
Information	Mode	Does listening, reading, speaking, and writing support one another?
Genre	Time	Can the line and limits survive pressure?
Tool	Scaffold	Can work continue after the cue closes?

Change one main condition at a time. If everything changes, a lower score cannot tell you whether difficulty, background, or method caused it. Transfer does not require identical tasks; it requires the core action to remain callable.

12. Seven Days, Thirty Days, Twelve Weeks

Seven days: let a first version appear

• complete a first version of one real task;
• run three complete learning loops;
• classify one recurring error;
• remove the latest cue and perform a related task on day seven;
• choose only one condition to change next week.

Thirty days: make rhythm sustainable

• preserve one comparable output each week;
• alternate an old task with a parallel task;
• delete activities that create familiarity without use;
• record sleep, capacity, time, and real response;
• repeat under similar conditions on day thirty and state what cannot be inferred.

Twelve weeks: let ability enter life

• preserve parallel samples in weeks 1, 4, 8, and 12;
• add an unfamiliar topic, real audience, or time pressure midway;
• complete a meeting, exam simulation, documentation check, email delivery, or public artifact;
• let the result decide whether to continue, downgrade, change a variable, or stop.

13. One-page General Diagnostic

For a complete diagnostic:

1. choose one real topic and four tasks;
2. record conditions and complete listening, reading, speaking, and writing first versions;
3. name one task-critical error per skill;
4. choose one evidence card and one targeted repair per skill;
5. complete a parallel task on day seven and repeat one real delivery on day thirty;
6. state what changed, what did not, and what remains unsupported.

Do not average four scores into a personality judgment. Someone with B2 reading and A2 speaking does not need one “true level” first; they need the next task and evidence card that fit each component.

Sources and Boundaries

• Dunlosky et al. (2013), Improving Students' Learning With Effective Learning Techniques: reviews ten techniques and their generalisability across learning conditions, learners, materials, and criterion tasks; practice testing and distributed practice receive high-utility assessments, while rereading and highlighting are more conditional.
• Roediger & Karpicke (2006), Test-Enhanced Learning: compares testing with repeated study using prose materials and reports delayed-retention differences; sample, material, and measure do not guarantee every task.
• Cepeda et al. (2006), Distributed Practice in Verbal Recall Tasks: reviews and quantitatively synthesises distributed practice in verbal recall; the useful interval depends on retention goal, material, and task.
• Hattie & Timperley (2007), The Power of Feedback: analyses differences in feedback type, timing, and delivery; “feedback works” is not a reason to add undirected comments.
• Rasch & Born (2013), About Sleep's Role in Memory: reviews sleep and memory-consolidation research; individual needs, health, and real schedules still require separate judgment.
• Pashler et al. (2008), Learning Styles: finds no adequate evidence for matching instruction to fixed visual/auditory labels while cautioning that untested variants cannot all be declared false.
• Research differs in sample, material, measure, and delay. Important work should return to current facts, primary sources, real audiences, and named accountable owners.

Related entry points: CEFR Goals and Self-check | Vocabulary | Grammar | Listening | Reading | Speaking | Writing | English Diagnostic | Evidence Chain

Closing: Give the Method to the Next Attempt

The temptation of a learning method is that it looks like a shortcut. A method begins to belong to you only when the material is closed, the cue is gone, the task has changed, and it can still help you finish something.

You do not need to memorise every researcher's name or turn life into continuous measurement. One plain loop is enough: attempt the task alone and let difficulty appear; obtain specific feedback; return after an interval; change one condition and see whether the ability travels with you.

A good method does not make you better at explaining how hard you have worked. It makes you more able to complete something when no one is reminding you. What remains is not the name of the method, but the knowledge of how to begin the next unfamiliar problem.