update

templates/Prompt/Artefakt-Check.md

@@ -0,0 +1,21 @@
+You are the quality checker for the learning artefacts of ONE block of the topic "{topic}". Another agent produced question patterns and worked examples. Check both.
+
+GROUND TRUTH — the supported facts (measure only against these):
+{facts}
+
+QUESTION PATTERNS (rows: (subblock) question):
+{table}
+{fehlend}
+WORKED EXAMPLES (numbered; PROBLEM / STEPS / RESULT):
+{examples}
+
+TASKS:
+1. **pattern** — return the CLEANED-UP final version of ALL patterns: each question EXACTLY ONE question mark, one thing, 1–2 sentences, neutral, hits the subblock's core, answerable from the facts without invented assumptions. Keep good ones unchanged; rephrase violations. Exactly one pattern per subblock; carry `block`/`subblock` over unchanged. Questions in GERMAN.
+2. **pattern_ergaenzt** — for each subblock listed above as MISSING a question: create ONE pattern (same rules). Empty list if none are missing.
+3. **examples_probleme** — object to an example ONLY if clearly faulty: calculation error, wrong inference, contradicts/invents beyond the facts, or would imprint a wrong path. Recompute yourself; conservative — when in doubt, keep. Give 1-based `index`.
+
+Reply with ONLY the JSON — no code fences, no other text. Format:
+{{"pattern": [{{"block": "…", "subblock": "…", "question": "…"}}],
+ "pattern_ergaenzt": [{{"block": "…", "subblock": "…", "question": "…"}}],
+ "examples_probleme": [{{"index": 2}}]}}
+{extra}

templates/Prompt/Artefakt-Generate.md

@@ -0,0 +1,25 @@
+Build the learning artefacts for the subblocks of ONE block of the topic "{topic}": per subblock ONE question pattern and ONE flashcard, plus a worked example where it carries.
+
+BLOCK WITH SUBBLOCKS AND THEIR FACTS (process EVERY subblock):
+{blocks}
+
+1. **pattern** — exactly ONE question pattern per subblock (exam question pool):
+- ONE concrete, concise question about the subblock's CORE concept, phrased neutrally (difficulty is set later by the learner's level). Exactly one question mark, one thing — no multi-parter, no "name three …". At most 1–2 sentences, no scenario build-up.
+- Answerable from the supplied core points and cited facts. Assume NOTHING that is not in the facts (no "shown" code/diagrams, no invented values).
+
+2. **cards** — exactly ONE flashcard per subblock (active recall):
+- `question`: brief recall question testing exactly one key point (≤ 15 words). `answer`: the short, precise answer from the supported facts (≤ 25 words). No prose, no lead-in, invent nothing.
+
+3. **examples** — a worked example ONLY where it carries understanding:
+- `problem` (1 sentence) → `steps` (2–5 followable steps, right order) → `result` (1 sentence). Rely on `example_idea` and the supported facts; compute cleanly. A pure definition or meta-knowledge gets NO forced example — leave it out.
+
+Common rules:
+- `block`/`subblock` are exactly the titles above.
+- **Mathematics ALWAYS as LaTeX**, never raw characters: inline `$…$`, longer set off as `$$…$$`. NO Unicode math substitutes (`≤`, `≥`, `Σ`, `δ`, `∈`, `⊆`, `×`) and no bare `_`/`^` — use `$\le$`, `$\ge$`, `$\Sigma$`, `$\delta$`, `$\in$`, `$\subseteq$`, `$\times$`, `$n^d$` instead. The supplied facts often contain raw Unicode math — convert it. Real code/paths/identifiers in backticks.
+- ALL content in GERMAN.
+
+Reply with ONLY the JSON as your final message — no code fences, no other text. EXACTLY this format:
+{{"pattern": [{{"block": "…", "subblock": "…", "question": "…"}}],
+ "cards": [{{"block": "…", "subblock": "…", "question": "…", "answer": "…"}}],
+ "examples": [{{"block": "…", "subblock": "…", "problem": "…", "steps": ["…"], "result": "…"}}]}}
+{extra}

templates/Prompt/Artifact-Example-Check.md

@@ -1,25 +0,0 @@
-You are a correctness checker for worked examples on the topic "{topic}". Another agent produced them — they are meant to demonstrate a **correct** solution to the learner. A wrong example imprints a wrong solution path. Find the faulty ones.
-
-GROUND TRUTH — the supported facts (measure only against these):
-{facts}
-
-EXAMPLES TO CHECK (numbered; PROBLEM / STEPS / RESULT):
-{examples}
-
-Object to an example if ANY of these applies:
-- **Calculation error** in a step (a number, a transformation, a unit).
-- **Wrong inference:** a step does not follow from the previous one; the RESULT does not follow from the steps.
-- **Not covered:** a step/value contradicts the facts or invents something not stated there.
-- **Incomplete/misleading:** the path would lead the learner to a wrong conclusion.
-
-Rules:
-- Measure ONLY against the facts + internal logic. Recompute it yourself.
-- **Conservative:** object only to what is **clearly** wrong. When in doubt, keep it.
-- Give the 1-based number (`index`) of each faulty example.
-
-Reply with ONLY the JSON — no other text, no code fences — one of the two:
-{{"ok": true}}
-{{"problems": [{{"index": 2}}, {{"index": 5}}]}}
-
-Output no other text.
-{extra}

templates/Prompt/Artifact-Example.md

@@ -1,24 +0,0 @@
-Build a worked example for the subblocks of the topic "{topic}" — a fully worked-through case that carries understanding.
-
-BLOCKS WITH SUBBLOCKS AND THEIR FACTS (go through EVERY subblock):
-{blocks}
-
-A good worked example:
-- **problem**: a concrete, small task/question about the subblock (1 sentence).
-- **steps**: 2–5 followable steps from the problem to the solution. Each step a brief sentence, in the right order.
-- **result**: the final result / the insight (1 sentence).
-- Rely on `example_idea` and the supported facts of the subblock. Compute cleanly; invent no values that contradict the facts.
-- **Only where it carries:** if a subblock cannot be sensibly shown with an example (a pure definition, meta-knowledge), LEAVE IT OUT — no forced example.
-- **Mathematics ALWAYS as LaTeX**, never as raw characters: inline `$…$` (e.g. `$T_A(n)$`, `$\Sigma^*$`, `$|x| = 5$`, `$O(n^d)$`, `$q_a$`), longer/worked-through formulas set off as `$$…$$`. NO Unicode math substitutes (`≤`, `≥`, `Σ`, `δ`, `∈`, `⊆`, `×`) and no bare `_`/`^` — use `$\le$`, `$\ge$`, `$\Sigma$`, `$\delta$`, `$\in$`, `$\subseteq$`, `$\times$`, `$n^d$`, `$q_a$` instead. Real code/paths/identifiers (not math) in backticks.
-- **The supplied facts often contain math as raw Unicode — convert it to LaTeX, do not take it over raw.**
-
-Write `problem`, the `steps` and `result` in GERMAN.
-
-Reply with ONLY the JSON (all examples) as your final message — no code fences, do NOT write a file. EXACTLY like this:
-{{"examples": [
-  {{"block": "<exact block title>", "subblock": "<exact subblock title>",
-    "problem": "…", "steps": ["…", "…"], "result": "…"}}
-]}}
-
-Output no other text.
-{extra}

templates/Prompt/Artifact-Flashcard.md

@@ -1,23 +0,0 @@
-Build ONE flashcard (question→answer) for active recall for each subblock of the topic "{topic}".
-
-BLOCKS WITH SUBBLOCKS AND THEIR FACTS (process EVERY subblock):
-{blocks}
-
-A good flashcard:
-- **question**: a brief recall question that tests exactly one key point (not "explain everything"). One question, one thing.
-- **answer**: the short, precise answer — based on the supported facts/key points of the subblock. Invent nothing extra.
-- Rely on the supplied facts. Where a supported fact exists, the answer must match it.
-- Brief: question ≤ 15 words, answer ≤ 25 words. No prose, no lead-in.
-- **Mathematics ALWAYS as LaTeX**, never as raw characters: inline `$…$` (e.g. `$T_A(n)$`, `$\Sigma^*$`, `$|x| = 5$`, `$O(n^d)$`, `$q_a$`), longer calculations set off as `$$…$$`. NO Unicode math substitutes (`≤`, `≥`, `Σ`, `δ`, `∈`, `⊆`, `×`) and no bare `_`/`^` — use `$\le$`, `$\ge$`, `$\Sigma$`, `$\delta$`, `$\in$`, `$\subseteq$`, `$\times$`, `$n^d$`, `$q_a$` instead. Real code/paths/identifiers (not math) in backticks.
-- **The supplied facts often contain math as raw Unicode — convert it to LaTeX, do not take it over raw.**
-
-Write `question` and `answer` in GERMAN.
-
-Reply with ONLY the JSON (all cards) as your final message — no code fences, do NOT write a file. EXACTLY like this:
-{{"cards": [
-  {{"block": "<exact block title>", "subblock": "<exact subblock title>",
-    "question": "…", "answer": "…"}}
-]}}
-
-Output no other text.
-{extra}

templates/Prompt/Facts-Check.md

@@ -1,26 +0,0 @@
-You are the fact-checker for the learning facts of the topic "{topic}". Another agent has extracted facts per subblock. Check ONLY the **backed facts** for truth — not the examples.
-
-{source}
-
-FACTS TO CHECK (per subblock):
-{facts}
-
-Check per subblock:
-1. **Evidence fidelity**: Does each `cited_facts` entry appear that way in the source (accurate in substance)? Is the source citation correct? With source material (folder or excerpts): check against it. Without a source: is it established standard knowledge?
-2. **Factual correctness**: Are formulas, values, definitions, signatures technically correct? A wrong formula/value is a defect.
-3. **Fact vs. example**: Is a self-computed/invented example wrongly declared as `cited_facts`? That is a defect — it belongs in `example_idea`.
-4. Do NOT check examples (`example_idea`) for source evidence — they are generative.
-
-Note only REAL defects (wrong fact, wrong source, example disguised as a fact). Not matters of taste.
-
-**Discard vs. correct** — decide per defect:
-- `verwerfen: true` — the subblock is substantively **not backable**: an invented claim, a bound/formula/assertion not findable in the material, or simply technically wrong. The subblock is then REMOVED. Be sure — when in doubt, `false`.
-- `verwerfen: false` — the core is right, only a fact/value/source is imprecise and correctable.
-
-Reply with ONLY the JSON — no other text, no code fences.
-
-Format — all in order:
-{{"ok": true}}
-Otherwise (subblock title EXACTLY as above):
-{{"problems": [{{"subblock": "<exakter Subblock-Title>", "problem": "…", "discard": true}}]}}
-{extra}

templates/Prompt/Facts-Research.md

@@ -1,30 +0,0 @@
-Extract the learning facts for each subblock of the topic "{topic}". These facts are the binding basis from which the guide text, the levels, and the exam questions are later created — they must be **correct**.
-
-{source}
-
-BLOCKS WITH SUBBLOCKS (process EVERY subblock):
-{blocks}
-
-Collect per subblock ONLY the essentials (superfluous material harms learning); write all field content in GERMAN (technical terms/code identifiers stay original):
-- **key_points**: 1–3 concise statements — what must one understand about this subblock?
-- **prerequisites**: what must one know beforehand (a half-sentence)? Empty if nothing.
-- **hurdles**: typical beginner misconception (a half-sentence). Empty if none.
-- **cited_facts**: hard facts (definitions, formulas, values, names, signatures) — **only what you can back up**. Each with a source:
-  - With a source file/script: cite **accurately in substance** and give the location (e.g. „Skript Def. 6.3, Z.66"). Invent no values, compute nothing yourself.
-  - Without a source (pure topic): only established standard knowledge; verify uncertain points via web search; source = „allgemein" or the URL.
-- **example_idea**: ONE example that carries understanding — freely phrased. **Here** is where self-formed sentences, mini-scenarios, worked examples belong. Empty if an example adds nothing.
-
-HARD SEPARATION — important:
-- `cited_facts` = only backable material from the source/established knowledge. **NEVER** output a self-computed or invented example as a backed fact.
-- A worked example, an invented sentence, a constructed case → belongs in `example_idea`, not in `cited_facts`.
-- When in doubt: better to leave out than to claim falsely.
-
-Reply with ONLY the JSON as your final message — no code fences, no prose around it. Do NOT write any file; your tools are for research only. EXACTLY this format:
-{{"facts": [
-  {{"block": "<exakter Block-Title>", "subblock": "<exakter Subblock-Title>",
-    "key_points": ["…"], "prerequisites": "…", "hurdles": "…",
-    "cited_facts": [{{"text": "…", "source": "…"}}], "example_idea": "…"}}
-]}}
-
-Output no other text.
-{extra}

templates/Prompt/Facts-Supplement.md

@@ -1,25 +0,0 @@
-You are filling **gaps**. For each subblock of the topic "{topic}", the facts already captured are listed below. A first pass extracted them — and experience shows backable material slips through. Your task: add **only what is missing** from the source.
-
-{source}
-
-SUBBLOCKS WITH FACTS ALREADY CAPTURED:
-{blocks}
-
-Rules:
-- Output **only NEW** points per subblock that are backed by the source and still **missing** above.
-- **Repeat nothing** from what is already "captured" — not even rephrased. If it's already there → leave it out.
-- **Only backable material.** With a source file/script: cite accurately in substance + give the location. Invent no values, compute nothing yourself. Without a source file (pure topic): only established standard knowledge.
-- **HARD SEPARATION:** `cited_facts` = only backable material. A self-computed/invented example belongs in `example_idea`, NEVER in `cited_facts`.
-- Add `key_points` only if an **essential** aspect is missing — no duplications, no trivialities.
-- **If you find nothing new for a subblock → leave it out.** If you find nothing at all → empty list. Better nothing than fabrication.
-- Write all field content (key_points, prerequisites, hurdles, cited_facts text, example_idea) in GERMAN, matching the source material (technical terms/code identifiers stay original).
-
-Reply with ONLY the NEW facts as ONE JSON in your final message — no code fences, no file writing (tools are for research only). EXACTLY this format:
-{{"facts": [
-  {{"block": "<exakter Block-Title>", "subblock": "<exakter Subblock-Title>",
-    "key_points": ["…"], "prerequisites": "…", "hurdles": "…",
-    "cited_facts": [{{"text": "…", "source": "…"}}], "example_idea": "…"}}
-]}}
-
-Output no other text.
-{extra}

templates/Prompt/Levels-Mapping.md

@@ -1,21 +0,0 @@
-Several agents have classified the subblocks of the topic "{topic}" by learning-path position. On some they disagree. Decide the final level for each disputed subblock.
-
-DISPUTED SUBBLOCKS (with the votes cast):
-{disputed}
-
-Level = learning-path position (WHEN you need the point), NOT difficulty:
-- **beginner**: foundation — what you need first/at all; a prerequisite for the rest. Even if conceptually demanding.
-- **advanced**: builds on it; common variants, typical application.
-- **expert**: subtleties, special cases, niches, rare details — even if substantively simple (e.g. an obscure flag, a version detail).
-
-Rules:
-- A simple niche detail is `expert`; a fundamental complex concept is `beginner`.
-- Weigh the votes, decide on the merits by learning-path position.
-- **`advanced` is not a fallback choice.** Don't pick the middle because the votes are scattered — decide by the criterion "when do you need this?". Hit `beginner`/`expert` clearly where they apply.
-- Exactly one level for EACH disputed number.
-
-Reply with ONLY the JSON — no other text, no code fences.
-
-Format (no other text):
-{{"levels": {{"1": "beginner", "4": "expert"}}}}
-{extra}

templates/Prompt/Levels-Research.md

@@ -1,33 +0,0 @@
-Assign each subblock of the topic "{topic}" to its **learning-path position**: beginner, advanced, or expert. The level controls when a learner sees the point (beginners see only „beginner", advanced learners more).
-
-SUBBLOCKS (grouped by block; many state their core points with „Kern:"):
-{subblocks}
-
-Use the core points for a well-founded classification — judge on the content, not just the title.
-
-DECISIVE — the level measures WHEN you need the point in the learning path, NOT how hard it is to understand:
-- **beginner**: foundation. What you need first/at all to understand the block in the first place. A prerequisite for the rest. Even if the concept is conceptually demanding — if it is the foundation, it belongs here.
-- **advanced**: builds on the foundation. Common variants, typical application, the next step.
-- **expert**: subtleties, special cases, edge cases, niches, rare details. What you need only once you've mastered the basics — **even if it is substantively simple** (e.g. an obscure flag, a version detail).
-
-IMPORTANT — difficulty is NOT the criterion:
-- A **simple niche detail** (rarely needed, e.g. „seit 6.4.12.0") is `expert`, not `beginner`.
-- A **fundamental but complex concept** (needed early) is `beginner`, not `expert`.
-- Ask: "when in the learning path do you need this?" — not "how hard is it?".
-
-Rules:
-- Level RELATIVE within each block: what is foundation here, what is build-up, what is fine detail?
-- Not every block needs a `beginner` point. Some topics presuppose prior knowledge and are entirely `advanced`/`expert` — that is intended. Level honestly by learning-path position, without inventing an artificial entry point.
-- Only judge — invent nothing, change no subblocks.
-
-Don't flee to the middle:
-- **`advanced` is not a fallback choice.** Don't pick the middle because you're unsure.
-- Decide each level by the question **"when in the learning path do you need this?"** — not by the safest middle option.
-- **Differentiate within the block:** What is foundation (beginner), what is build-up (advanced), what is fine detail (expert)? Not everything is the middle.
-- Hit `beginner` and `expert` clearly where they apply — no bonus for the middle.
-
-Reply with ONLY the JSON — no other text, no code fences.
-
-Format (exactly one level for EACH number; no other text):
-{{"levels": {{"1": "beginner", "2": "advanced", "3": "expert"}}}}
-{extra}

templates/Prompt/Question-Pattern-Critique.md

@@ -1,23 +0,0 @@
-You are the quality reviewer for the question pool of a learning exam on the topic "{topic}". Another agent created question patterns for **several blocks** — exactly one pattern per subblock. Review them and return a **cleaned-up** final version.
-
-PATTERNS TO REVIEW (grouped by block; rows: (subblock) question):
-{table}
-
-REVIEW AND CLEAN UP AGAINST THESE CRITERIA (per block, separately):
-- **Style:** each question is EXACTLY ONE question, one question mark, one thing. No "and"/"as well as", no enumeration. At most 1-2 sentences. Violations → rephrase.
-- **Unambiguously answerable:** clearly answerable from the block's knowledge, with no invented assumptions. Unanswerable ones → fix.
-- **Core hit:** the question targets the subblock's core concept, phrased neutrally (not deliberately easy or hard — the difficulty is set later by the learner's level). If it misses the core → rephrase.
-- **Exactly one per subblock:** exactly one pattern remains per subblock. Duplicates → keep one.
-- Carry over the `block` and `subblock` of the entries unchanged (only `question` may change); do not invent new ones.
-
-Keep good patterns unchanged. Change only what genuinely violates the criteria.
-
-Keep every question in GERMAN (the questions are for German-speaking learners), even though these instructions are in English.
-
-Reply with the cleaned-up final version of ALL blocks as ONE JSON — no other text, no code fences — EXACTLY in this format:
-{{"pattern": [
-  {{"block": "<exact block title>", "subblock": "<exact title>", "question": "<one concrete question>"}}
-]}}
-
-Output no other text.
-{extra}

templates/Prompt/Question-Pattern-Research.md

@@ -1,28 +0,0 @@
-You are building the question pool for a learning exam on the topic "{topic}". Create **question patterns** — **exactly one** concrete example question **per subblock**. Each pattern is later turned, at exam time, into a slightly varied question — matched to the learner's level (beginner to expert). So the difficulty is NOT in the pattern; it only comes in at exam time. Therefore ask about the **core concept** of the subblock, phrased neutrally.
-
-You are working on **several blocks**. Each block has its own subblocks — many with core points, cited facts, and an example:
-{blocks}
-
-TASK (for **each** of the blocks above):
-- For **each** of its subblocks, create **exactly one** pattern — a concrete, concise question about that subblock's core concept.
-- Each pattern is ONE concrete question (1-2 sentences, exactly one question mark, one thing) — no multi-parter, no enumeration.
-- The question must be answerable from the block's knowledge. Do not invent extra assumptions.
-- Base the question on the supplied **core points and cited facts** (the reliable foundation). Hit the CORE CONCEPT of the subblock, not a detail to be memorized. Assume NOTHING that is not in the facts: no "shown" code snippets/diagrams, no concrete values/lists/numbers from some later guide text.
-- The question hits the **core** of the subblock — the central point one must have understood. Not too narrow on a detail, not too broad across the whole block.
-- `block` is exactly one of the block titles above. `subblock` is exactly one of the subblock titles of the respective block.
-
-HARD STYLE RULES PER QUESTION:
-- Exactly ONE question, one question mark, one thing. No "and"/"as well as", no "name three …".
-- At most 1-2 sentences, no preamble, no scenario build-up.
-- Phrased neutrally — not deliberately easy, not deliberately hard. The difficulty is set by the exam via the learner's level.
-- Direct address, clear German.
-
-Write every question in GERMAN (the questions are for German-speaking learners), even though these instructions are in English.
-{extra}
-
-Reply with ONLY the JSON (all patterns of all blocks) as your final message — no code fences, do NOT write a file. EXACTLY this format:
-{{"pattern": [
-  {{"block": "<exact block title>", "subblock": "<exact subblock title>", "question": "<one concrete question>"}}
-]}}
-
-Output no other text.

templates/Prompt/Relevance-Mapping.md

@@ -1,18 +0,0 @@
-Several agents have rated the subblocks of the topic "{topic}" by relevance (relevant/peripheral). For some they disagree. Decide the final relevance for each disputed subblock.
-
-DISPUTED SUBBLOCKS (with the votes cast):
-{disputed}
-
-Relevance:
-- **relevant**: core/standard knowledge that carries the topic — central concepts, definitions, properties. For theoretical topics, also central theorems with no practical use.
-- **peripheral**: a side note, niche/special case, "nice to know", deprecated/exotic, purely historical details, concrete numeric examples, mere cross-references. NOT peripheral: a central theorem or a fundamental property. For the complete guide only.
-
-Rules:
-- Weigh the votes and decide by the criterion **core vs. peripheral within ITS OWN block**. `peripheral` is a genuine category — mark peripheral items deliberately as such. Only true core concepts/central theorems are never `peripheral`.
-- Exactly one value for EACH disputed number.
-
-Reply with ONLY the JSON — no other text, no code fences.
-
-Format (no other text):
-{{"relevance": {{"1": "relevant", "4": "peripheral"}}}}
-{extra}

templates/Prompt/Relevance-Research.md

@@ -1,24 +0,0 @@
-Rate each subblock of the topic "{topic}" by relevance: relevant or peripheral. The relevance controls whether a point goes into the focused guides or only into the complete guide.
-
-SUBBLOCKS (in brackets the parent block of the number; many state their core points with "Kern:"):
-{subblocks}
-
-Use the core points for a well-grounded rating — judge on the content, not just the title.
-
-Relevance:
-- **relevant**: core/standard knowledge that CARRIES the topic — central concepts, definitions, properties, common practice. For practical topics: what you need in actual use. For theoretical topics: central theorems/properties, EVEN without practical use. What you genuinely should know.
-- **peripheral**: a side note, special/niche case, "nice to know", deprecated/exotic — PLUS purely historical details (who/when), concrete numeric examples of a general principle, mere cross-references. NOT peripheral: a central theorem or a fundamental property just because it seems "theoretical". Belongs only in a complete guide.
-
-Rules:
-- These subblocks **all already** belong to the topic — the question is not *whether* but *how central*. Judge **core vs. peripheral within its own block**, one at a time.
-- **The basic syntax / core rule of a block — what the block TITLE promises — is NEVER peripheral.** A block "Headings" without its heading syntax marked relevant is broken. Check this per block before anything else.
-- **When the BLOCK itself is a fringe/extension feature** (non-standard, niche, an add-on rather than core of the topic), only its basic syntax and purpose are relevant — its detail rules and special cases are `peripheral`. A fringe feature must not be decomposed deeper than the core topics (measured: a non-standard block carried 22 "relevant" detail rules while core blocks had 6).
-- `peripheral` is a **genuine category** for clear cases: nice-to-know, special/niche cases, detail/numeric examples, historical notes, cross-references, style tips.
-- `peripheral` items are CUT from the focused guides — a wrong `peripheral` loses content, a wrong `relevant` only adds a paragraph. **When in doubt → relevant.**
-- Only judge — invent nothing, change no subblocks.
-
-Reply with ONLY the JSON — no other text, no code fences.
-
-Format (exactly one value for EACH number; no other text):
-{{"relevance": {{"1": "relevant", "2": "peripheral", "3": "relevant"}}}}
-{extra}

templates/Prompt/Subblock-Fix.md

@@ -0,0 +1,22 @@
+You are correcting and completing the subblocks of ONE block of the topic "{topic}" — strictly from the source excerpts.
+
+BLOCK: {block}
+
+{source}
+
+TASKS:
+{auftraege}
+
+Rules:
+- Work ONLY from the excerpts. Invent no values, compute nothing yourself.
+- For corrections: return the subblock with its EXACT title and the complete corrected fact set.
+- For gaps: ONE new subblock per gap, atomic, GERMAN title (max ~10 words), fully backed — if the excerpts don't support it, leave it out entirely.
+- Field rules as always: key_points 1–3; cited_facts only backable with location; example_idea free; level beginner|advanced|expert; relevance relevant|peripheral. All content in GERMAN.
+
+Reply with ONLY the JSON — no code fences, no other text. EXACTLY this format (empty list if nothing is backable):
+{{"subs": [
+  {{"title": "…", "level": "…", "relevance": "…",
+    "key_points": ["…"], "prerequisites": "…", "hurdles": "…",
+    "cited_facts": [{{"text": "…", "source": "…"}}], "example_idea": "…"}}
+]}}
+{extra}

templates/Prompt/Subblock-Generate.md

@@ -0,0 +1,35 @@
+Break the block below (topic "{topic}") into its SUBBLOCKS — the individual learnable sub-points — and extract per subblock the learning facts from the material. A later guide writes a short text PER subblock; the facts are the binding basis for guide text, levels, and exam questions — they must be **correct**.
+
+BLOCK:
+{block}
+
+{source}
+
+What a subblock is:
+- A single sub-point you must learn for the block. One statement, one aspect, one pitfall.
+- Examples: block `<img>` → `src` (Bildquelle), `alt` (Alternativtext), empty `alt` for decorative images. Block `<p>` → text paragraph as a block, allowed inline children, no nesting.
+
+DECISIVE — the count follows the difficulty:
+- A simple, trivial block has FEW subblocks (1–2). Do NOT inflate it.
+- A complex, rich block has MANY subblocks. Leave out nothing essential the material covers.
+- Each subblock is atomic (one sub-point), backed by the excerpts — a point the excerpts do not support is left out.
+- Subblock titles in GERMAN (code identifiers stay original), max. ~10 words, one statement.
+{seeds}
+Per subblock, collect ONLY the essentials (all field content in GERMAN; technical terms/code identifiers stay original):
+- **level**: beginner | advanced | expert — difficulty within this block.
+- **relevance**: relevant (core of the topic) | peripheral (edge knowledge).
+- **key_points**: 1–3 concise statements — what must one understand?
+- **prerequisites**: what must one know beforehand (a half-sentence)? Empty if nothing.
+- **hurdles**: typical beginner misconception (a half-sentence). Empty if none.
+- **cited_facts**: hard facts (definitions, formulas, values, names) — **only what the excerpts back**, each with the location (e.g. „Skript Def. 6.3, Z.66"). Invent no values, compute nothing yourself.
+- **example_idea**: ONE example that carries understanding — freely phrased. Empty if an example adds nothing.
+
+HARD SEPARATION: `cited_facts` = only backable material. A worked example, an invented sentence, a constructed case → `example_idea`, NEVER `cited_facts`. When in doubt: better to leave out than to claim falsely.
+
+Reply with ONLY the JSON as your final message — no code fences, no other text. EXACTLY this format:
+{{"subs": [
+  {{"title": "…", "level": "beginner|advanced|expert", "relevance": "relevant|peripheral",
+    "key_points": ["…"], "prerequisites": "…", "hurdles": "…",
+    "cited_facts": [{{"text": "…", "source": "…"}}], "example_idea": "…"}}
+]}}
+{extra}

templates/Prompt/Subblock-Konsolidierung.md

@@ -1,29 +0,0 @@
-Du prüfst die Zerlegung EINES Lernbausteins in Subbausteine für das Thema "{topic}". Jeder Subbaustein wird später ein eigener Absatz im Lernguide. Ziel ist eine 100%-Zerlegung: keinen Eintrag kann man weglassen, ohne dass eine Lücke entsteht — und keinen ergänzen, ohne dass Dopplung entsteht.
-
-BLOCK: {block}
-
-SUBBAUSTEINE (nummeriert, mit ihren Kernpunkten):
-{subs}
-
-## Auftrag (vier Urteile)
-1. **gruppen — Dopplungen**: Einträge, die DIESELBE Aussage treffen (anders formuliert) oder reine TEILMENGE eines anderen sind. Pro Gruppe: `haupt` = der GRUNDLEGENDSTE, allgemeinste Eintrag (Basis vor Detail, Grundsyntax vor Sonderfall, nie ein inhaltsleerer Stub).
-2. **kataloge — Aufzählungs-Bündel**: Einträge, die REINE Nachschlage-Listen gleicher Art sind (z. B. Symbol-, Befehls-, Namenslisten ohne eigene Verhaltensregel). Solche Listen bündelst du zu EINEM Katalog-Eintrag mit sprechendem Titel. Einträge mit eigener Regel oder eigenem Verhalten gehören NICHT in ein Bündel.
-3. **fremd — Weglass-Test**: Einträge, deren Aussage nicht zum Thema "{topic}" gehört (anderes Fachgebiet, Grundlagen einer anderen Technologie) — auch wenn die Aussage stimmt und belegbar ist. Ließe man sie weg, fehlte dem Thema nichts.
-4. **luecken**: Ein KERN-Aspekt des Blocks fehlt — ohne ihn ist die Zerlegung UNVOLLSTÄNDIG (die Grundregel oder ein Pflichtbestandteil fehlt). Katalog-, Detail- und Randwissen ist KEINE Lücke. Was man ergänzen KÖNNTE, ist keine Lücke — nur was fehlen DARF nicht. Im Zweifel: keine Lücke. Meist ist die Liste leer.
-
-## Regeln — der häufigste Fehler zuerst
-- **Verschiedene Schreibweisen, Marker, Befehle oder Syntaxen sind IMMER eigene Einheiten** — auch wenn sie derselben Kategorie angehören oder im selben Satz erklärt werden könnten. „Gehört zusammen" oder „ähnliches Thema" ist KEIN Gruppierungs-Grund. Nur „sagt DASSELBE" zählt.
-- Zwei Einträge zum gleichen Konzept mit VERSCHIEDENEN Facetten (Grundregel vs. Sonderfall vs. Randbedingung) sind KEINE Dopplung.
-- Eine benannte Variante oder ein Spezialfall ist NIE Dopplung seiner Basis.
-- Einträge mit gegensätzlicher Aussage (Negation) nie gruppieren.
-- Ein inhaltsleerer oder generischer Eintrag, dessen Titel ein anderer Eintrag vollständig abdeckt, ist Teilmenge → gruppieren, der konkrete Eintrag ist `haupt`.
-- Im Zweifel: NICHT gruppieren, NICHT bündeln, NICHT als fremd markieren.
-
-Antworte NUR mit JSON, ohne weiteren Text:
-{{"gruppen": [{{"haupt": 1, "weitere": [4]}}], "kataloge": [{{"titel": "Symbolkatalog: Operatoren und Relationen", "mitglieder": [2, 5, 9]}}], "fremd": [7], "luecken": ["fehlender Kernaspekt"]}}
-
-- gruppen: pro Gruppe haupt + weitere (Nummern). Keine Dopplungen → [].
-- kataloge: pro Bündel Titel + Mitglieds-Nummern. Keine → [].
-- fremd: Nummern themenfremder Einträge. Keine → [].
-- luecken: fehlende Kernaspekte in Stichworten. Keine → [].
-{extra}

templates/Prompt/Subblock-Mapping.md

@@ -1,27 +0,0 @@
-Below, for each block of the topic "{topic}", are the subblocks found. Return the **cleaned** final list per block.
-
-{source}
-
-FOUND SUBBLOCKS (two groups per block):
-{blocks}
-
-The groups:
-- **Consensus (≥2 finders):** found independently multiple times — keep by default. Remove only fabrications/duplicates.
-- **Uncertain (1×):** named by only one finder — **scrutinize strictly**. Include an uncertain entry ONLY if it is **clearly backed by the source AND a standalone point**. When in doubt, leave it out.
-
-Rules:
-- **Evidence check (important):** Check each subblock against the source. **Discard whatever is NOT backable in the material or clearly invented** — fabricated bounds, formulas, values, or claims not actually in the source. With source material (folder or excerpts): check against it. Without a source (pure topic): keep only established standard knowledge, drop the doubtful/false.
-- **Merge duplicates:** Subblocks that state the same point in other words are ONE. Keep the clearest, drop the rephrasings. (e.g. „Broadcast Mode verfügbar" and „Broadcast-Modus aktivieren" → one.)
-- Keep all technically **DISTINCT** and backable sub-points in full — leave out nothing essential.
-- Discard whatever is too fine-grained, at the edge of the topic, or technically doubtful.
-- Each point atomic (one statement). Copy kept points VERBATIM, do not rephrase, invent nothing.
-- The count follows the difficulty: better few distinct than many redundant points.
-
-Reply with ONLY the final lists — no other text, no code fences. One block marker per block (title EXACTLY as above), with the final subblock list below it:
-
-<!-- block: Exact block title -->
-- Subblock
-- Subblock
-
-Output the marker line exactly like this. Every block must appear. No text outside the blocks.
-{extra}

templates/Prompt/Subblock-Research.md

@@ -1,27 +0,0 @@
-Break each assigned block of the topic "{topic}" into its SUBBLOCKS — the individual learnable sub-points the concept is made of. A later guide will write a short text PER subblock.
-
-Assigned to you — binding: every block must appear, invent no additional ones:
-{assignment}
-{material}
-What a subblock is:
-- A single sub-point you must learn for the block. One statement, one aspect, one pitfall.
-- Examples: block `<img>` → `src` (Bildquelle), `alt` (Alternativtext), empty `alt` for decorative images, `width`/`height` against layout shifts, void element without a closing tag. Block `<p>` → text paragraph as a block, allowed inline children, no nesting.
-
-DECISIVE — the count follows the difficulty:
-- A simple, trivial block has FEW subblocks (1–2). Do NOT inflate it.
-- A complex, rich block has MANY subblocks. Leave out nothing essential.
-- There is NO target count. `<img>` must have noticeably more subblocks than `<p>`.
-
-Rules:
-- Each subblock is atomic: exactly one sub-point. No two aspects in one point.
-- Only what THIS block yields (scope). Don't inflate anything mentioned only in passing.
-- {backing}
-- Subblock titles in GERMAN (code identifiers stay original), max. ~10 words, one statement.
-
-Reply with ONLY the subblock lists in this exact marker format — one block marker per block (title EXACTLY from the assignment), the subblocks as a list below it. No code fences, no text outside the blocks:
-
-<!-- block: Exact block title -->
-- First subblock
-- Second subblock
-{known}
-{extra}

templates/Prompt/Subblock-Verify.md

@@ -0,0 +1,26 @@
+You are auditing the decomposition of ONE block of the topic "{topic}" into subblocks. Below are the numbered subblocks with their captured facts, then source excerpts. Apply the 100%-decomposition test: no entry removable without a gap, none addable without duplication — and verify the facts.
+
+BLOCK: {block}
+
+SUBBLOCKS (numbered; key points indented):
+{subs}
+{unsicher}
+{source}
+
+Judge ALL of the following (use the numbers):
+1. **gruppen** — entries that state the SAME thing or where one is a subset of the other: group them and name the number that should remain (`haupt` = the base statement, not the detail).
+2. **kataloge** — pure enumeration entries of ONE kind (e.g. five option rows): bundle them under ONE short GERMAN collective title.
+3. **fremd** — entries off-topic for the topic "{topic}" (not this block — the TOPIC).
+4. **luecken** — essential aspects of THIS block that the excerpts cover but no entry captures (short German phrases). Only real gaps, no nice-to-haves.
+5. **uebernehmen** — for each UNSICHER-numbered entry: "ja" if the excerpts back it and it fills a real spot, else "nein".
+6. **facts_probleme** — entries whose facts contain something wrong or unbackable: `discard: true` ONLY if the entry as a whole is unsupportable in substance; otherwise `discard: false` with a short `hinweis` what to correct.
+7. **levels** / **relevanz** — ONLY entries whose level (beginner/advanced/expert) or relevance (relevant/peripheral) is clearly wrong: number → correct value.
+
+Judge ONLY from the excerpts. If everything is fine, return empty lists/objects.
+
+Reply with ONLY the JSON — no code fences, no other text. Format:
+{{"gruppen": [{{"haupt": 1, "weitere": [4]}}], "kataloge": [{{"titel": "…", "mitglieder": [2, 5]}}],
+ "fremd": [7], "luecken": ["…"], "uebernehmen": {{"9": "ja"}},
+ "facts_probleme": [{{"nr": 3, "discard": false, "hinweis": "…"}}],
+ "levels": {{"2": "expert"}}, "relevanz": {{"5": "peripheral"}}}}
+{extra}