Your mistakes are the syllabus: making corrections actually stick

Every learner has one. The gender of a noun you have looked up thirty times. The auxiliary verb you always pick wrong. The particle you know the rule for and still get backwards under time pressure. It has been corrected repeatedly, you understand the rule perfectly when it is explained, and it comes out wrong again the next time you are composing at speed.

The intuitive diagnosis is that you need to be corrected more. The evidence points somewhere less obvious: the problem is usually not the number of corrections, but what the correction asked you to do — which, most of the time, is nothing.

The argument about whether correction works at all

It is worth knowing that this was a genuine fight in the field, because it explains a lot of the contradictory advice you will read.

In 1996 John Truscott published an argument that grammar correction in writing classes was ineffective and should be abandoned — that it did not improve accuracy and cost time better spent elsewhere. Dana Ferris responded that the studies he drew on were mostly not designed to answer that question. What came out of the following two decades of studies — Bitchener and Knoch’s work among them — is a more useful picture than either extreme:

• Focused feedback works better than comprehensive feedback. Targeting one or two error types produces measurable gains on those types. Marking everything in red produces overload, and it is the format most likely to be skimmed.
• Some errors respond much faster than others. Ferris’s distinction between treatable and untreatable errors is the practical one: treatable errors are rule-governed — verb agreement, tense formation, case endings, article use in some languages — and you can look the rule up and apply it. Untreatable errors are idiomatic: word choice, collocation, the preposition a particular verb just takes. Those improve mainly through exposure, and expecting them to yield to a rule is how people conclude that correction does not work.
• The feedback has to be noticed to do anything. Which brings us to the central mechanism.

Noticing is the bottleneck

Roy Lyster and Leila Ranta’s classroom research on feedback types is the clearest illustration. In immersion classrooms, by far the most common teacher move was the recast: the learner says something malformed, the teacher repeats it correctly and the conversation moves on. It is polite, it does not interrupt, and it is the most natural thing to do.

It also produced the least learner uptake of the feedback types they coded. From the learner’s seat, a recast is ambiguous: it can be heard as a confirmation rather than a correction. You said a thing, an adult said the thing back, the conversation continued. Nothing signalled that anything was wrong.

The moves that produced far more self-repair were the ones that pushed the work back onto the learner: elicitation, clarification requests, metalinguistic hints — “that verb takes the dative” rather than a rewritten sentence.

A correction that does the work for you is a correction you can nod at and forget.

This is the practical core of Schmidt’s noticing hypothesis applied to feedback. A form you did not consciously register as different from what you produced does not change what you will produce next time.

Four things that make a correction stick

1. It has to be explicit about what changed

Not just a corrected sentence handed back, but a visible difference and a named reason. “bin, not habe — movement verbs take sein in the perfect” does something that a silently rewritten sentence does not. If your feedback source only gives you the clean version, do the naming step yourself; it takes ten seconds and it is most of the value.

2. You have to produce the corrected form yourself

The single biggest lever, and the most commonly skipped. Reading a correction is recognition — the same easy direction a flashcard runs in. Retyping it while looking at it is barely better. What transfers is closing the correction and generating the sentence again from the meaning, with the error site still live.

This is why prompts beat recasts in the classroom data, and it is reproducible on your own: after every correction, write one sentence that would need the same form, without looking.

3. It has to come back later

A re-attempt thirty seconds after the correction is working from short-term memory. The same re-attempt three days later is a real retrieval — and retrieval at a delay is what builds durable memory, exactly as it does for vocabulary.

In practice this means an error you made needs to be recorded somewhere and re-surfaced, not just fixed in place. A correction that exists only in a document you will never reopen has a half-life of about a day.

4. It has to be narrow

If eleven things were wrong in your paragraph, you are not going to fix eleven things. Pick the one or two that recur across sessions and work on those; let the rest be corrected and pass. Focused feedback is the version of correction that the evidence actually supports, and it is also the version that does not make you want to stop writing.

Why errors fossilise

Fossilisation — an error that stops responding to instruction and stays for years — is usually described as a natural endpoint. The mechanism is less mysterious than that word suggests. Every time you produce a form, you strengthen it. If you produce the wrong form a hundred times and the right form twice, both after being corrected, the arithmetic is not subtle.

Which means the antidote is arithmetic too: raise the number of times you produce the correct form, in real composition, spread out over time. Not re-reading the rule — producing the form. A drill built from your own wrong sentence is the most direct version of this, because the retrieval cue is the exact context in which you got it wrong.

What not to bother correcting

Being selective is part of the method, not a compromise on it. Three categories to let go:

• One-off word choice. If a word was wrong once and you have no reason to think you will reach for it again, note it and move on. Untreatable errors improve through reading, not through drilling.
• Errors above your level. Being corrected on a construction you have not studied yet produces a correction with nothing to attach to. Note that it exists; come back when the rest of the scaffolding is there.
• Style, early on. “A native speaker would phrase this differently” is useful at B2 and demoralising at A2, where the goal is a sentence that is correct and understood.

Putting it together

The loop that works is short and slightly tedious, which is why almost nobody runs it manually:

1. Write something real, at the edge of what you can do.
2. Get explicit, visible corrections — what changed and why.
3. Name the rule for each one, in your own words.
4. Immediately produce one new sentence using the same form, unaided.
5. Have that error come back in a few days, as a drill, in a fresh context.
6. Track which error types keep recurring, and let that list — not a textbook’s table of contents — decide what you study next.

Step 6 is the one that changes the shape of your studying. Your recurring errors are a syllabus that is already sequenced by usefulness: they are, by definition, the things you try to say often and get wrong. No curriculum written for a general audience can be that well targeted at you.

The reason this loop is rare is that steps 2, 5 and 6 are expensive by hand — they need someone to correct every sentence, remember every error, and bring them back on a schedule. That is the part worth automating, and the daily routine is where it fits: eight minutes of writing, five of processing, and a weekly look at what keeps coming back. If you want to know which errors to expect in your particular language, the language pages list the ones that show up most.