Quick answer: If you've ever finished a conversation and thought "I know that wasn't right, but I can't tell you why" — you don't have a vocabulary problem. You have a feedback loop problem. The fix isn't another grammar textbook. It's instant, contextual rule lookup at the exact moment a mistake happens. SpeakShark now ships a built-in Grammar Reference with 25 curated topics and a pattern detector that reads your past spoken sentences and surfaces the rules behind every error.

I've been speaking English for ten years and I still occasionally drop a final consonant or pick the wrong tense after "if." The reason isn't that I never learned the rule. It's that the moment I needed the rule, the rule wasn't there.

This post is about how we closed that gap — and why we think it changes what a grammar resource is supposed to do.

The two failure modes of every grammar resource I've used

There are two ways grammar resources fail learners. One is shallow. The other is deeper and most apps don't even notice it.

Failure mode 1: Generic drills. You buy a grammar workbook. Chapter 4 is "Past Perfect." You do twenty fill-in-the-blanks. You get 18/20. You close the book. Three days later, you say "When I arrived, she already left" in a real conversation. The "had" was supposed to be there. You knew the rule in isolation; you couldn't access it under pressure.

Failure mode 2: Decontextualized lookup. You're aware you said something wrong, so you Google "past perfect vs past simple." Twelve tabs open. Stack Exchange has three contradictory answers. Cambridge has a textbook explanation that doesn't address your specific sentence. By the time you parse it, the conversation is twenty minutes behind you and the moment is dead.

Both failures have the same underlying cause — the resource doesn't know what you just said, and you don't know what you don't know.

What we wanted to build

We wanted three things from a grammar resource:

Browse-ability for self-study — if a learner wants to refresh "Conditionals" before a meeting, they should be able to open the topic, read it in 90 seconds, and close it. No multi-page slogs.
Just-in-time lookup from real mistakes — when SpeakShark catches an error in a conversation, the user should be one tap away from the rule that explains it. Not a search. A tap.
No noise — no AI-generated content that's correct-on-average and wrong on edges. Every topic should be hand-written by someone who actually knows the rule, with examples that make the trap obvious.

We delivered all three in the Library → Grammar tab. Here's how each part works.

Part 1: 25 hand-written topics, browsable in seconds

The Grammar Reference ships with 25 curated topics, grouped into 3 mega-categories:

Verbs & Actions — All twelve tenses (broken into Present, Past, Future), Modal Verbs (Can/Could/May/Might + Should/Must/Have-to), Verb Forms (Gerund vs Infinitive), Passive Voice, Phrasal Verbs
Nouns & Structure — Articles (a/an/the), Quantifiers (much/many/few/little), Comparison (comparative + superlative), Subject-Verb Agreement, Prepositions of Time + Place
Sentences & Speech — Conditionals (Zero through Third), Relative Clauses, Question Formation, Reported Speech

Every topic follows the same compact shape:

[Title]
[Formula: e.g., "had + V3"]

WHEN TO USE
1. Rule with example sentence
2. Rule with example sentence
...

FORMATION
1. ...

COMMON MISTAKES
1. ✗ I had went → ✓ I had gone
2. ...

RELATED: Past Simple · Present Perfect · Reported Speech

We wrote each one by hand. No LLM. The reason is that LLMs are mostly correct, which means they're occasionally wrong in ways you can't predict — and a grammar reference that's "occasionally wrong" is worse than no reference at all, because it gaslights learners who don't have a way to fact-check.

The full topic for Past Perfect, for example, looks like this in the popup:

Past Perfect had + past participle (V3)

When to use

Earlier past action (before another past action). "When I arrived, she had already left."

Completed action before a specific past time. "By 2020, he had written three books."

In reported speech (Past Simple → Past Perfect). "He said he had finished."

Common mistakes

Using V2 instead of V3 after 'had'. ✗ I had went → ✓ I had gone

Using Past Perfect when only ONE past action. ✗ I had eaten breakfast yesterday → ✓ I ate breakfast yesterday

Related: Past Simple · Present Perfect · Reported Speech

That's it. No fluff. No three-paragraph build-up. The reader can be in and out in 30 seconds if they just need a refresh.

Part 2: The 4-level browser that scales to 100+ topics

Even 25 topics gets unwieldy in a flat list. We tried four different layouts before settling on the right one. The constraint: it had to work on a phone, and it had to stay clean even if we ship 100+ topics later.

The final shape is a 4-level drill-down:

Level 0: Single "Grammar Reference" card
  ↓ tap
Level 1: 3 mega-groups (Verbs / Nouns / Sentences)
  ↓ tap one
Level 2: Categories inside that group (~5 each)
  ↓ tap one
Level 3: Topics in that category
  ↓ tap one
GrammarPopup with full content

Two things make this work:

Breadcrumb back-nav. Once you've drilled in, the breadcrumb at the top lets you jump back to any earlier level without retracing your steps. Grammar > Verbs & Actions > Tenses — tap any segment to go back to that level.

Search bypasses every level. Type "past perfect" into the search bar at the top and you skip the whole drill — you get a flat list of matches immediately. The browse hierarchy is for not knowing what you need; the search is for knowing exactly what you need.

This was the part I wrestled with most. My first prototype was a 14-category accordion. It looked tidy on desktop but burned 700+ pixels of mobile vertical space before showing anything useful. The 4-level model takes ~120 px for the landing card. It scales linearly with topic count without ever turning into a wall of links.

Part 3: The thing nobody else does — pattern detection on past mistakes

Here's the part I'm most proud of, and the part I haven't seen in any other learning app:

When you finish a conversation in SpeakShark, the AI catches your mistakes and saves them with the exact sentence you said. The Vocab tab shows individual word errors. The new Grammar tab shows sentence-level mistakes — but it doesn't just show them as flat strings. It runs a pattern detector across each sentence and surfaces the underlying grammar rules.

Example. You said this in a session:

"I have went to the store yesterday."

A normal app would store the correction ("I went to the store yesterday") and that's that. Maybe a one-line explanation.

SpeakShark stores the correction AND detects:

have went matches the pattern have/has + V3 → Present Perfect topic
yesterday (with present perfect) → triggers the "no specific past time" rule from the same topic
the store → Articles topic

So when you open the past mistake card, you see three chips below the sentence:

[Present Perfect] [Articles] [Past Simple time markers]

Tap any chip — instant popup with the rule, the formula, and examples. AND a green banner at the top of the popup says "From your sentence: 'have went'" so you see the connection between your specific mistake and the abstract rule.

How does it work? Honestly, plain regex. We hand-wrote a pattern detector with ~20 rules covering tense markers (have/has + V3, had + V3, am/is/are + V-ing), modal verbs, conditional triggers (if + ...), comparison structures (more X than, the X-est), quantifiers, frequency adverbs, reported speech triggers, passive constructions, relative clauses, gerund/infinitive cues, articles, and time/place prepositions. No LLM call in this layer — just pattern matching on cached past sentences. It runs client-side, instantly, every time the user opens the Grammar tab.

The result: every past mistake becomes a launch pad into the rule that explains it. You stop having to figure out what kind of mistake you made before you can look it up. The app tells you.

Why this matters more than it sounds

The thing about grammar resources is that they assume you know which topic to look up. That's a learner-side problem most apps don't solve.

If you're a Vietnamese intermediate learner and you said "yesterday I have eaten pho", you don't know yet that what you need to learn is "Present Perfect vs Past Simple — incompatible time markers." You just know the sentence sounded off. Maybe you thought it was a vocabulary problem. Maybe you thought you needed to drill verbs.

The pattern detector closes that diagnostic gap. It looks at your sentence, recognizes the structural marker, and hands you the exact topic to study. No diagnosis from you required.

This is the part that, in my opinion, separates passive grammar references (Cambridge online, dictionary apps) from active ones. A passive reference waits for you to come find it. An active one comes to you the moment you make a mistake.

How to actually use it

Three workflows I'd recommend.

Workflow 1 — Just-in-time after a session.

Finish a Daily Talk or Challenge conversation.
Open Library → Grammar tab.
Scroll to "Your past mistakes" below the Reference section.
Find a card with the sentence that bothered you.
Tap a chip. Read the rule. Close. Done.

This takes about 90 seconds and locks in the rule while the mistake is still fresh.

Workflow 2 — Pre-session refresher.

Before a meeting where you're going to speak English, open Library → Grammar.
Drill into the category you're nervous about (Conditionals, Reported Speech, Articles, etc.).
Tap one topic. Skim the "Common mistakes" section in 30 seconds.
Close the tab. Go into your meeting with the trap surfaced in working memory.

The point isn't to memorize the rule. It's to have it primed so you catch yourself in real time.

Workflow 3 — Free exploration.

Tap "Related" chips at the bottom of any topic.
Drift between connected concepts (Past Simple ↔ Past Perfect ↔ Reported Speech).
Stop when you've absorbed enough.

This is the closest thing the app has to a "rabbit hole" — but it's a structured one, where every jump connects to a real grammar concept.

What we're not doing

A few things this Grammar Reference deliberately doesn't do, and probably shouldn't:

It's not a textbook. Each topic is ~5–8 rules. If you want the full treatment of subjunctive mood across English's history, go read Quirk's Comprehensive Grammar of the English Language.
It doesn't auto-quiz you. There's a separate part of SpeakShark for active practice — that's what Daily Talk and Challenges are for. The Reference is for understanding.
It doesn't generate examples on the fly. Every example is hand-written. If we ever expand the topic count to 100+, we'll add LLM-assisted draft generation with mandatory human review — but only as a draft tool, never as the final published content.

What's next

We're tracking which topics get opened most often. If "Conditionals" or "Articles" or "Phrasal Verbs" turn out to be the heaviest, we'll deepen those first — adding sub-topics, more examples, IELTS-band-specific tips.

We're also considering tying the pattern detector to the Daily Talk live transcript. So as you're speaking, the bar at the top would show the rule for the construction you're using right now. Realtime grammar feedback without breaking the flow of conversation. Still prototyping.

If you want to try this — the Grammar Reference ships today in SpeakShark Free. No paywall. 3 free conversations per day forever. The Reference is unlimited.

Tap a mistake. Read the rule. Stop guessing.