📅

Prompt-Intelligenz-Report — 26. September 2026

🏆 Highlight

### Poster Prompts v2 — 100 Copy-Paste-Designstile, evaluiert für 28 Dollar John Hartnup hat 100 Poster-Prompts als Copy-Paste-Sammlung veröffentlicht. HN-Kommentatoren fanden die typischen Fehler: Wörtlichkeit (jedes Wort im Prompt landet im Bild — der Cricket-Club crickett, die Kirchensanierung kriegt eine Kirche), Clutter und erfundene Wappen. Also baute er Version 2 — mit Claude C...

All sources fetched successfully. Now parsing the results into a compact digest. Portals parsed — good September material (TechCrunch, Ars). Now fixing the HN/GitHub parser (f-string syntax error). Excellent haul. Key finds: Claude Fable 5.1 prompting docs, "A Prompt to Learn Anything" (fresh Sept 25), Poster Prompts (100 copy-paste design prompts), Qwen-Image 2.1 & GPT Image 2.5 prompt repos, cinematic video prompt skills, Cornetto prompting framework. Now deep-fetching the actual prompt text from these sources. Deep fetches mostly landed (tutor README needs the API route; Firebase items look stub-sized). Now extracting the actual prompt text from all sources. Rich material extracted. One more targeted pass to get complete prompt texts (video formula combos, poster prompt examples, GPT Image 2.5 verbatim prompts, full Jevify prompt, Cornetto code block, TVC master prompt), then I'll write the report. All key material is in. Two final grabs: the actual Swiss-style poster prompt text (perfect fit for a Swiss site) and the full verbatim GPT Image 2.5 prompts from the local JSON. Last data pass: fetch the fresh "A Prompt to Learn Anything" (AGENTS.md), complete the Swiss poster prompt text, and pull exact article URLs. All research complete — writing the German report now.

Prompt-Intelligenz-Report — 26. September 2026

Die besten kopierbaren Prompts und Prompt-Techniken der letzten Tage — kuratiert für prompta.ch.

🔤 TOP 3 PROMPTS — Textgenerierung

1. Der Mastery-Tutor — ein kompletter Systemprompt für langfristiges Lernen

Prompt (vollständig, kopierbar):

---
name: tutor
description: Mastery-based AI tutor for long-range learning goals. Manages a multi-domain curriculum hierarchy, persistent state, and ad-hoc material.
---

# Tutor

You are a mastery-based tutor guiding a student toward a long-range goal across multiple domains. You do not move a
student forward until they have demonstrated understanding — not just correct answers. Your job is to teach, probe, and
route, not to validate.

## Directory layout

```
/tutor/
  AGENTS.md
  curriculum.yaml        ← master roadmap: goal, domain order, cross-domain prerequisites
  state.json             ← top-level state: student info, active domain, per-domain mastery summary
  /domain_name/
    curriculum.yaml      ← node DAG for this domain
    state.json           ← rolling mastery scores and history for this domain
  /materials/            ← student drops any files here: PDFs, images, notes, problem screenshots
```

You can read files in `materials/` directly — PDFs, images, and text. When a student references external material or
hands you a file, check `materials/` first. It's likely the one with the newest creation/modified date. They do not need
to register it in any YAML file. Ad-hoc material can be incorporated into the session without becoming part of the
formal curriculum when requested.

## On startup

1. Read `curriculum.yaml` (root). Load the student's goal, domain list, and cross-domain prerequisites.
2. Read `state.json` (root). If it does not exist, run a placement assessment (see below) and create both files.
3. From `active_domain` in root state, load that domain's `curriculum.yaml` and `state.json`.
4. Greet the student by name if known. One sentence of context — where they are and what's next. Then present a problem.

Do not load all domain files on startup. Only load the active domain. Load another domain's files only if routing
requires it.

## State schema

Root `state.json`:
```json
{
  "student": "name or null",
  "goal": "mirrors curriculum.yaml goal field",
  "render_math": "unicode|latex|typst|plain",
  "personality": "a personality description"
  "active_domain": "domain_id",
  "domains": {
    "domain_id": { "mastery_summary": 0.0, "status": "locked|active|complete" }
  },
  "observations": []
}
```

Domain `state.json`:
```json
{
  "current_node": "concept_id",
  "mastery": { "concept_id": 0.0 },
  "history": [{ "concept_id": "", "score": 0.0, "timestamp": "" }],
  "observations": []
}
```

`observations` is a list of plain-English notes you write about the student's reasoning patterns — not scores, but
qualitative flags like `"confuses kernel with null space"` or `"strong intuition, skips justification"`. Update these
when you notice a pattern. Surface them when relevant, not on every turn.

Mastery is a float 0–1, rolling average over the last 5 scored attempts per concept. Write state after every scored
interaction.

## Placement assessment

On first run, present 3–5 diagnostic problems that span the prerequisite chain of the starting domain — easy at the
root, hard at the leaves. Score each. Set initial mastery values. Place the student at the deepest node where mastery <
0.8. Flag weak upstream nodes in `observations` even if the student places ahead of them.

If the student arrives with prior experience and says so, ask two or three targeted questions before skipping placement.
Do not take their word for it.

## Core loop

1. Present one problem. Difficulty tracks current mastery: below 0.4 → foundational, 0.4–0.7 → standard, above 0.7 →
   stretch.
2. Wait for the student's response.
3. Score it. Update domain state. Write both state files if root state changed.
4. Give targeted feedback: what was right, what was wrong, and why. Do not just reveal the answer.
5. If mastery ≥ 0.85 on current node and all prerequisites are satisfied: advance. Tell the student.
6. If mastery < 0.4 after 3 attempts: route back to the weakest unmastered prerequisite. Say why.
7. If all nodes in the active domain reach mastery ≥ 0.85: mark domain complete, unlock the next domain per the root
   curriculum, switch active domain, run a short transition assessment.

## Scoring

Do not reward correct answers alone. Probe reasoning. A correct answer with no justification scores 0.6 max. A wrong
answer with sound reasoning scores higher than a lucky correct answer. Before finalizing a score, ask "why does that
work?" or "what breaks if you drop this condition?" if the subject allows it.

## Routing

Fix the weakest link first. If the student struggles on the current node and a prerequisite has mastery < 0.7, route
there. If a weakness in a completed domain is clearly blocking progress in the active domain, say so and offer to
revisit it. Cross-domain routing should be explicit — tell the student why you're going back.

## Off-roadmap conversations

If the student raises a problem or topic outside the current curriculum, engage with it fully. Do not update mastery
scores for it. Do not alter roadmap position. Log it in `observations` if it reveals something about their reasoning.
The student is never trapped in the current node.

## Resources

When introducing a new concept or when a student is stuck, you can search the web for links to other resources and/or
include a relevant link from the domain's `curriculum.yaml` if one exists. Do not pad responses with links.

## Tone

Direct. Do not over-praise. Do not over-explain. Ask more than you tell. If the student is bored or rushing, make the
problem harder. If the student is lost, find the gap in prerequisites — do not just simplify the concept.

If a `personality` key is set in the root `state.json`, overlay that into your responses.

## Student commands

- `roadmap` — root curriculum with domain statuses and mastery summaries, then current domain node DAG with scores
- `status` — one line: active domain, current node, current mastery score
- `hint` — a nudge without revealing the answer; noted in history but does not penalize score
- `skip` — advances past current problem; records 0.5, does not contribute to mastery
- `explain` — teach the concept before asking a question
- `observations` — show your qualitative notes about the student's reasoning patterns

## Curriculum generation

If no domain `curriculum.yaml` exists for the active domain, generate one before proceeding. Ask the student two
questions first: what is the subject, and do they have a target depth or end goal (e.g. "enough to read ML papers" vs
"PhD level"). Generate the node DAG, write it to the correct path, then continue startup normally.

If no root `curriculum.yaml` exists, ask for the student's long-range goal and generate both the root curriculum and the
first domain curriculum before running placement.

## Math formatting

Check `render_math` in root `state.json`. Format all math accordingly:

- `"unicode"` — use unicode symbols and plain-text layout. Matrices as aligned columns with brackets. Fractions as a/b.
  Superscripts as x^2. Subscripts as x_1. Default if field is absent.
- `"latex"` — wrap all math in LaTeX delimiters. Use for environments that render it.
- `"typst"` — write each problem (with context) to `problem.typ` in the tutor root after presenting it. Use full Typst
  math syntax. The student is responsible for running `typst watch`.
- `"plain"` — no special formatting. Prose descriptions only where possible.

The student can change this at any time by saying e.g. "switch to latex mode". The tutor updates state and confirms.

Am besten mit: Claude Code / Claude Cowork (mit Dateizugriff für State-Dateien), GPT-6 Codex; gekürzt auch als Systemprompt in jedem leistungsstarken Chat-Modell.

Warum effektiv: Der Tutor geht erst weiter, wenn Verständnis nachgewiesen ist (Mastery ≥ 0.85) — nicht bei richtigen Antworten. Falsche Antworten mit sauberer Begründung scoren höher als geratene richtige; qualitative Beobachtungen („confuses kernel with null space") und Routing zurück zur schwächsten Voraussetzung machen den Loop pädagogisch wirksam statt quizartig.

Quelle: https://github.com/ZaneH/tutor | 3 GitHub-Sterne (erschienen am 25.09.)

Community Resonanz: Brandneu — am 25. September als „A Prompt to Learn Anything" auf Hacker News vorgestellt; das Repo folgt dem AGENTS.md-Standard („AGENTS.md is the README").

2. Jevify — der Investigationsprompt für die Jev-Ära

Prompt (vollständig, kopierbar):

I want you to deeply investigate what **Jev, TypeSafe's structured decision model, could make possible in this project**.

My hypothesis is that this could be a big deal. It may substantially reduce cost and latency for work we already do. More interestingly, it may make semantic judgments cheap and fast enough to use throughout the application—in places where calling an LLM previously seemed too slow, expensive, or cumbersome to consider.

Take that possibility seriously. Be ambitious about what we could build and rigorous about what the evidence supports.

**Start by reading these sources and inspecting this project:**

- [TypeSafe introduction](https://docs.typesafe.ai/introduction)
- [Typed decision primitives](https://docs.typesafe.ai/primitives)
- [API reference](https://docs.typesafe.ai/api)
- [Documentation index](https://docs.typesafe.ai/llms.txt)
- [Jev architecture investigation](https://archerhume.com/posts/jevs-architecture-unmasked)—use this to generate hypotheses; its architectural deductions are not verified implementation details.

Follow relevant documentation links to verify current pricing, limits, batching behavior, and integration options. Separate vendor claims, independently measured results, and your own hypotheses.

The documented interface evaluates a shared state against multiple typed questions, returning choices, rubric scores, and yes/no probabilities. Questions in one request are evaluated independently; application code combines their answers. Understand this model before proposing integrations.

The broader idea I want you to explore is **using language understanding as a routine computational operation**. Read text or application state, evaluate many specific properties, and use those results directly in software. Think about the input-processing side of language models without assuming Jev exposes an encoder, embeddings, or arbitrary internal representations.

**1. Understand what this project is trying to accomplish.**

Inspect the actual code, architecture, data flows, prompts, tests, and available performance evidence. Identify the user outcomes that matter.

Find where we currently:

- Spend money or time on model calls.
- Generate text only to parse it into a decision.
- Repeatedly process the same context.
- Serialize judgments that could be independent.
- Use brittle rules because semantic understanding seemed impractical.
- Rely on manual review, coarse categories, sampling, or delayed batch processing.
- Discard information or limit coverage to stay within a budget.

Tie observations to concrete files and execution paths. Do not assume the project needs existing LLM calls to benefit.

**2. Reconsider the design from first principles.**

Ask: **If many useful semantic judgments were affordable within our application's response-time budget, what would we design differently?**

Explore three kinds of opportunity:

- **Direct savings:** perform existing work with less cost or latency at acceptable quality.
- **Better outcomes:** improve coverage, relevance, reliability, or responsiveness within the same budget.
- **New capabilities:** enable useful behavior we currently do not attempt.

Give the third category substantial attention. Look beyond replacing individual model calls. Consider whether we could evaluate every event instead of sampling, assess many candidates or dimensions at once, react while a user is interacting, continuously reassess changing state, or combine fast judgments with slower reasoning in a better overall workflow.

Those are starting points. Develop ideas specific to this project rather than repeating a generic feature list.

Explicitly identify assumptions in the current architecture that exist because semantic computation was expensive. Explain which could change and what user-visible benefit follows.

**3. Make the strongest opportunities concrete.**

For each serious candidate, specify:

- The user problem and current behavior.
- The exact integration point and available input state.
- The specific questions Jev would answer and the appropriate primitives.
- Which questions can share a request and which genuinely depend on earlier results.
- How ordinary code would consume the answers.
- What still requires generation, deeper reasoning, retrieval, or deterministic logic.
- The expected benefit, implementation effort, and most consequential failure mode.

For the top candidates, include representative request shapes and consumer pseudocode grounded in the current API.

Do not hide a complex reasoning task inside a vaguely worded classification question. Show that the proposed decomposition preserves the information needed to make a good decision.

**4. Test the economics and performance assumptions.**

Estimate the complete workflow, including preparing inputs, network overhead, question tokens, downstream calls, retries, fallbacks, and mistakes that create extra work.

Distinguish lower latency per request from lower end-to-end latency. Identify the critical path. Do not assume that more questions are free, that batching scales indefinitely, or that provider-side parallelism eliminates client-visible costs.

Compare against the current implementation and credible simpler alternatives: deterministic code, caching, embeddings, conventional classifiers, or smaller generative models where appropriate.

When measurements are unavailable, provide explicit assumptions, plausible ranges, and break-even conditions. State what would have to be true for each proposal to be worthwhile.

**5. Design an evaluation that could prove us wrong.**

For the strongest opportunities, define:

- Representative inputs and held-out cases.
- Baselines and task-level success criteria.
- Relevant quality metrics, including asymmetric costs of false positives and false negatives.
- End-to-end cost, latency distributions, and throughput under realistic load.
- Tests for ambiguity, missing evidence, adversarial input, and sensitivity to question wording or batch composition.
- How thresholds, abstention, and fallback behavior would be validated.
- Clear go/no-go criteria.

Treat returned probabilities as signals whose calibration needs testing on our workload.

If credentials, suitable data, and an established experiment budget are available, run a small bounded experiment. Otherwise, produce a runnable evaluation plan and clearly identify what remains unmeasured. Continue the analysis without inventing results.

**6. Deliver a recommendation we can act on.**

Produce:

- A concise assessment of how consequential this could be for this particular project.
- A ranked opportunity table separating savings, quality improvements, and new capabilities.
- Detailed designs for the three strongest opportunities—or fewer if only fewer survive scrutiny.
- A first-principles sketch of how you would design the relevant parts of this product today with this capability available.
- The smallest experiment that would resolve the most important uncertainty.
- Ideas you rejected and the evidence or reasoning behind rejecting them.

Be explicit about what you inspected, what you measured, and what remains hypothetical. Keep exploration separate from production changes.

I want a serious investigation with imagination. Find the opportunities our existing architecture makes easy to overlook, then show which ones hold up.

Am besten mit: Claude Code / Claude Opus 5.5, GPT-6 Codex — direkt im Projektordner laufender Coding-Agents einfügen.

Warum effektiv: Ein sechsstufiger Beratungs-Prompt (Verstehen → First-Principles-Redesign → konkrete Opportunities → Wirtschafts-Annahmen prüfen → Evaluationsdesign → Empfehlung), der Beweisgrenzen erzwingt: „Be explicit about what you inspected, what you measured, and what remains hypothetical." Das Grundmuster — „investigate what X could make possible in this project" — funktioniert für jede neue Technologie, nicht nur für Jev.

Quelle: https://github.com/ryana/jevify | 186 GitHub-Sterne

Community Resonanz: Teil der Jev-Welle dieser Woche (awesome-jev: 1.155 Sterne, awesome-jev-tools: 635 Sterne); der Prompt ist modellagnostisch und passt in jeden Agent, der im Projekt arbeiten darf.

3. „Please remove all mannered prose" — Anthropics Anti-Floskel-Instruktion

Prompt (vollständig, kopierbar):

Mannered prose substitutes metaphor and flourish for direct statement. Instead of "a parameter worth varying," the mannered writer produces "a dial worth turning." Instead of "this point still matters," they write "this point earns its keep." The phrases exist to display the writer, not to convey the idea, and readers can tell. That is why mannered prose irritates: it makes the reader work harder so the writer can perform. It is also imprecise. Metaphors drag in connotations the writer did not choose and cannot control. The fix is to say what you mean. When a literal phrase is available, use it.

Kurzversion (laut Anthropic-Doku funktioniert sie oft allein):
Please remove all mannered prose.

Am besten mit: Claude Fable 5.1 (dafür geschrieben); wirkt bei jedem schreibenden Modell.

Warum effektiv: Die Instruktion definiert den Anti-Pattern mit konkreten Beispielen („a dial worth turning" statt „a parameter worth varying") und erklärt, WARUM die Floskel schadet: Der Leser arbeitet härter, damit der Autor sich präsentieren kann — und Metaphern ziehen Konnotationen nach sich, die der Autor nicht kontrolliert. Das ist stärker als jede generische „sei präzise"-Anweisung.

Quelle: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1 | HN 4 Upvotes (17.09.)

Community Resonanz: Der HN-Thread „Please remove all mannered prose" machte die Formulierung populär; sie kursiert inzwischen als Einzeiler gegen KI-Prosa — direkt aus der offiziellen Anthropic-Doku für Fable 5.1.

🖼️ TOP 3 PROMPTS — Bildgenerierung

1. JSON statt Prosa: 16-Panel-Pose-Reference-Sheet

Prompt (vollständig, kopierbar):

{"type":"pose reference sheet","subject":{"count":1,"description":"a fit young woman dancer shown repeatedly in a clean studio reference layout","appearance":{"gender":"female","age":"young adult","build":"athletic, toned midriff","skin tone":"light to medium tan","hair":{"color":"dark brown","style":"high messy ponytail with loose strands framing the face"},"expression":"neutral to focused"},"wardrobe":{"top":"charcoal gray sports bra or cropped athletic bralette","bottom":"oversized dark gray parachute cargo pants with gathered ankles","shoes":"white sneakers","accessories":["black wristband or fingerless glove on one hand","subtle sporty styling"]}},"layout":{"background":"plain white seamless studio background","grid":{"rows":4,"columns":4,"count":16,"cell labels":["1","2","3","4","5","6","7","8","9","10","11","12","13","14","15","16"]},"style":"clean contact-sheet or choreography chart with thin black dividers between panels and small black numbers at the upper left of each panel"},"poses":[{"label":"1","description":"relaxed standing pose, weight on one leg, one hand near hip, slight contrapposto"},{"label":"2","description":"wide low dance stance, one arm bent behind the head, the other arm extended and pointing to the right"},{"label":"3","description":"legs spread in a grounded stance, torso slightly tilted, one hand resting near the upper thigh"},{"label":"4","description":"very low wide squat facing forward, torso leaning back, one hand near the face and the other near the thigh"},{"label":"5","description":"wide side lunge stance, one arm arched overhead, the other arm extended outward in a stylized dance line"},{"label":"6","description":"balancing on one leg with the other knee lifted high, one hand near the face in a punchy hip-hop pose"},{"label":"7","description":"floorwork pose supported by one hand on the ground, torso reclined sideways, legs bent and lifted in a dynamic breakdance-like position"},{"label":"8","description":"casual upright pose with one hand behind the head and one knee bent upward"},{"label":"9","description":"one-legged balance pose with the lifted knee bent, both arms extended outward for motion and rhythm"},{"label":"10","description":"low kneeling or crouched pose, one knee up and one knee down, one arm thrust forward toward the viewer"},{"label":"11","description":"deep squat with legs apart, one arm curved overhead in a dramatic arc"},{"label":"12","description":"standing lean to one side with one arm extended sideways and the other hand near the hip or thigh"},{"label":"13","description":"reclining floor pose supported by one hand behind the body, one leg bent and one leg extended"},{"label":"14","description":"upright standing pose with one arm fully extended and pointing to the right"},{"label":"15","description":"front-facing pose stepping forward with one knee lifted, one arm reaching or pointing forward"},{"label":"16","description":"wide confident stance with one arm pointing diagonally upward to the right"}],"rendering":{"medium":"photorealistic studio fashion and dance reference image","lighting":"soft even studio lighting with faint shadows beneath the feet and body","camera":"full-body framing, straight-on view, consistent distance in every panel","quality":"sharp, high-resolution, realistic anatomy and fabric folds"}}

Am besten mit: GPT Image 2.5 (Release: 08.09.)

Warum effektiv: Das JSON macht explizit, wo natürlichsprachige Prompts driften: Layout (4×4-Grid mit nummerierten Panels und dünnen Trennlinien), 16 einzelne Posen als Liste und Rendering-Regeln (durchgängiger Kameraabstand, weiches Studiolicht). 25 der 150 Prompts der Sammlung sind JSON — bei GPT-Image-Modellen der zuverlässigste Weg zu konsistenten Sheets.

Quelle: https://github.com/youart-open-source/awesome-gpt-image-2-5-prompts | 236 GitHub-Sterne

Community Resonanz: Die Sammlung gibt jeden Prompt wortgetreu mit Urheberangabe (ExquisitMe, CC0) wieder — 150 Prompts, 110 Credits, 9 Use Cases; der Community-Standard für Prompt-Bibliotheken.

2. Premium-Food-Fotografie mit Template-Platzhaltern

Prompt (vollständig, kopierbar):

Create a square [ASPECT RATIO] premium food photography image of a steaming [FOOD] served in a dark black stone bowl or cast-iron skillet on a wooden board. The dish should look hot, glossy, spicy, and freshly served, with bite-sized pieces of browned protein, dried red chilies, green scallions, white onion, garlic, chili flakes, and visible Sichuan peppercorns coated in a deep red, oily Szechuan sauce. Use a slightly elevated close-up camera angle with shallow depth of field. Make the food the clear hero of the image, centered and richly detailed. Add visible steam rising naturally from the dish. Surround the bowl with subtle restaurant-style props like a dark red tray, scattered dried chilies, peppercorns, a small sauce bowl, or a blurred teapot in the background. Lighting should feel warm, moody, and editorial, like a high-end restaurant food shoot. Emphasize realistic textures and keep the image appetizing, realistic, cinematic, and polished. Avoid text, logos, hands, people, utensils covering the food, cartoon styling, fake plastic textures, excessive symmetry, or an overly clean stock-photo look.

Am besten mit: GPT Image 2.5 / GPT Image; Platzhalter [ASPECT RATIO] und [FOOD] vor dem Lauf ersetzen.

Warum effektiv: Ein komplettes fotografisches Briefing in einem Absatz: Kamera (leicht erhöht, shallow Depth of Field), Props (dunkles Tablett, Chilis, Teekanne im Bokeh), Lichtstimmung (warm, moody, editorial) — plus Negativliste („Avoid text, logos, hands, people, utensils covering the food…"), die genau die typischen Stockfoto-Fehler abfängt.

Quelle: https://github.com/youart-open-source/awesome-gpt-image-2-5-prompts | 236 GitHub-Sterne

Community Resonanz: Aus der E-Commerce-Kategorie der Sammlung; die Readme warnt zurecht, Template-Tokens zu ersetzen — sonst rendert das Modell den Platzhalter als Text mit ins Bild.

3. Qwen-Image 2.1: eine Zeile + Stil = fertiger Shot

Prompt (vollständig, kopierbar — die kurze Zeile ist der ganze Prompt):

Stil „Geometry of Light":

photo — a Tokyo street, hard evening light slicing a narrow alley, a passer-by casting a long shadow (B&W)

photo — a tiny figure at the end of a long arcade, a blade of light slicing the columns

poster — a minimalist B&W exhibition poster, a Tokyo-street beam and long shadows, title "LIGHT AND SHADOW"

Am besten mit: Qwen-Image-2.1 (7B Open Weights, Release: 20.09.) mit dem offiziellen PE-T2I-Rewriter.

Warum effektiv: Das neue Qwen-Paradigma: Der offizielle PE-T2I-Rewriter expandiert eine kurze Zeile (in jeder Sprache) zu einem langen englischen Prompt plus Negativ-Prompt plus passendem Seitenverhältnis. Der ComfyUI-Knoten liefert 17 fotografische Stile als reine „Kamerasprache + Stimmung" — nie Fotografen-Namen. Die Galerie zeigt: Die kurze Zeile IST der ganze Prompt; der Stil füllt nur, was du offen lässt.

Quelle: https://github.com/pottokao-dotcom/ComfyUI-QwenImage-PhotoStyles | 11 GitHub-Sterne

Community Resonanz: Qwen-Image 2.1 erschien am 20. September mit Day-0-Support für Diffusers, ComfyUI, vLLM-Omni, SGLang; awesome-qwen-image (74 Sterne) und qwen-image-2.1-skill (104 Sterne) wuchsen binnen Tagen.

🎬 TOP 3 PROMPTS — Videogenerierung

1. Die Cinematic-Formel — 9 Slots für jeden Video-Prompt

Prompt (vollständig, kopierbar):

[Shot size + Angle], [Subject + appearance], [Specific action], [Setting + weather],
[Lighting], [Camera movement], [Style + color], [Mood], [Technical]

Beispiel:
Medium close-up, low angle, a young woman in a red áo dài walks slowly through a rainy
Saigon alley at night, neon signs reflecting on wet asphalt, rim lighting from city lights,
slow dolly in, cinematic, teal and orange grading, melancholic mood, shallow depth of field,
35mm film grain

Regeln aus dem Skill: eine Kamerabewegung pro Clip, eine Hauptaktion pro Clip, Licht passend zu Wetter/Tageszeit, Style ↔ Farbe ↔ Mood zeigen in dieselbe Richtung, max. ~8 technische Keywords, Charakterbeschreibung über Clips derselben Story identisch halten.

Am besten mit: Veo 3 / Google Flow, Kling, Sora, Runway Gen-4, Hailuo, Luma, MiniMax-H3 — modellagnostisch.

Warum effektiv: Ersetzt vages „a nice cinematic scene" durch präzise Filmsprache („medium close-up, low angle, slow dolly in, rim lighting, teal and orange grading"). Die eingebauten Regeln verhindern genau die Fehler, an denen Videomodelle üblicherweise scheitern: widersprüchliche Kamerabewegungen, inkonsistente Charaktere, Licht, das nicht zur Szene passt.

Quelle: https://github.com/Rylaispirit/cinematic-video-prompt-skill | 97 GitHub-Sterne

Community Resonanz: 700+ Begriffe in Referenztabellen plus fertige Combos pro Videotyp (Storytelling, Produkt, Food, Horror, Reels); MIT-lizenziert und als Claude Skill sofort installierbar.

2. 18-Sekunden-TVC aus einem Produktfoto — der Picnic-Master-Prompt

Prompt (vollständig, kopierbar — Stilsektion des Master-Prompts):

## 〖风格〗

生成一支 18 秒、16:9 横屏、4K、25 帧的真人电影级绿茶电视广告。写实商业摄影结合极少量高品质二维 / 2.5D 角色动画，高动态范围，人物皮肤和头发纹理自然，PET 塑料瓶折射、标签印刷材质、冷凝水珠和浅黄绿色茶汤真实可信。

整体色彩以自然草绿色、鼠尾草绿、奶油白、浅茶绿和午后暖金色为主。前 6 秒女主仍沉浸在手机时，整体稍偏冷、低饱和，环境高频略弱，人物与户外环境之间存在轻微疏离感；女主放下手机并喝下一口绿茶后，暖金阳光、自然绿色和清透冷绿逐渐恢复，画面明亮、空气感增强，但不是突然改变天气，也不过曝。

镜头语言兼具高级饮料产品广告的克制与年轻生活方式广告的轻盈感：人物使用自然浅景深，草地与阳光具有真实空气透视；产品镜头轮廓光准确，茶汤必须清澈通透，不呈现荧光绿色；绿茶微距清爽、轻盈、有流动感，不黏稠、不奇幻。前半段节奏稍慢，小人出现后节奏变得灵动，女主喝茶后镜头和声音逐渐打开，结尾重新稳定收束。

全片不要使用贯穿画面的直线、光带、能量线、长虚线、速度线或发光轨迹作为视觉线索。画面之间的连接主要依靠包装小人的运动方向、小人的视线、落叶、冷凝水珠、茶叶、人物动作、前景遮挡和相似形状匹配剪辑完成。

Am besten mit: Seedance (dafür geschrieben); die Struktur (Stil + Timeline) funktioniert auch mit Kling und Veo 3.

Warum effektiv: Ein komplettes Werbedrehbuch als Prompt: 主体 definiert genau eine Protagonistin und ein Produkt mit voller Kontinuität (Kondenswasser, Flüssigkeitsstand, Etikett); 〖风格〗 steuert die emotionale Kurve über die Farbtemperatur (kalt/entsättigt in den ersten 6 Sekunden, warm/luftig nach dem ersten Schluck); die Negativliste verbietet VFX-Transitions (光带/能量线/速度线) und erzwingt Realismus. Der vollständige Master-Prompt (16 KB) ergänzt das um die 〖时间线〗 mit Taktzeiten 0–3 s / 3–6 s / 6–9 s ….

Quelle: https://github.com/huangbai-AI/tvc-advertising-skill | 10 GitHub-Sterne — vollständiger Master-Prompt: https://github.com/huangbai-AI/tvc-advertising-skill/blob/main/references/master-prompt.md

Community Resonanz: Der Skill generiert aus einem einzigen Produktfoto die komplette 18-Sekunden-Kampagne inklusive BGM- und Sounddesign-Vorgaben; Ersetzungsregeln für Heldin, Produkt und End-Slogan liegen bei — die neue Skill-Welle (Codex-/Claude-Skills) in Aktion.

3. Sechs-Abteilungs-Regie für narrative Videos

Prompt (vollständig, kopierbar):

使用 $leos-six-department-directing-team-skill-v1 分析下面这场戏：
雨夜，两位多年未见的兄弟在即将打烊的面馆重逢。哥哥想借钱，弟弟假装没听懂。
先给六部门导演方案，包括表演、空间、背景活动、摄影和连续性入口。

方案确认，现在授权生成 5000 字内的纯文本视频提示词，并进行六角色逐镜审稿。

单部门调用：
使用 $leos-six-department-directing-team-skill-v1，请摄影指导检查这组镜头的机位、轴线、视线与运镜动机。

Am besten mit: Codex / Claude Code mit installiertem Skill; der generierte Prompt läuft auf jedem narrativen Videomodell.

Warum effektiv: Zerlegt Regie in sechs Rollen — 总导演 (Lead Director), 表演指导 (Performance), 镜内执行 (In-Frame Staging), 摄影指导 (Director of Photography), 提示词导演 (Prompt Director), 场记 (Continuity) — bevor ein einziger Prompt generiert wird. Danach folgt ein Sechs-Rollen-Review pro Shot; das 5000-Zeichen-Limit hält den resultierenden Prompt kompakt genug für aktuelle Videomodelle.

Quelle: https://github.com/MasterLeos/leos-six-department-directing-team-skill-v1 | 123 GitHub-Sterne

Community Resonanz: Validierungsskripte prüfen den Output strukturell (Zeichenlimit, Shot-Nummern, Review-Abdeckung je Rolle) — Filmdisziplin trifft Prompt-Engineering.

🧠 TOP 3 NEUE TECHNIKEN

1. Das Cornetto- / Vier-C-Framework

Zusammenfassung: Ein Systemprompt-Snippet, das Agenten bei mittleren bis großen Aufgaben zwingt, Kontext, Constraints und eine vom Rest unabhängige Kontrollschleife zu liefern.

Erklärung: Der Autor „hazn" beschreibt die C's: context (das Warum), constraints (das Wie — und das Wie nicht) und control (die Schleife, die prüft). Der entscheidende Trick: Control bleibt bewusst UNABHÄNGIG von Kontext und Constraints — realisiert als Subagent mit frischem Kontextfenster oder als statischer Test (Red-Green-Tests). Damit prüft die Schleife nicht die eigene Hausaufgabe des Agenten. Der Autor hält das Framework absichtlich simpel: „I never go overboard with my prompts, and I think the framework is simple enough to scale well with model intelligence."

Beispielprompt:

# ask for these c's for medium to big tasks
Medium to big tasks are tasks that are not one off, aren't a simple question or something like filling out a form or parsing a pdf, it's building something new or synthesizing multiple things. Research doesn't fall under this.

## context
the why

## constraints
the how, and the how not

## control (aka, the loop)
the controlling, importantly, the control doesn't know about the context and constraints, usually subagents or more static control like red-green tests.

Geeignet für: Claude Code, Codex und alle Coding-Agents.

Ursprung: https://hazn.com/cornetto

Warum heute wichtig: Agent-Orchestrierung ist das Thema der Woche (Jev-Skills, Agent-Chaperone, GuardRail); die unabhängige Kontrollschleife ist der billigste Zuverlässigkeitsgewinn — und der Beitrag wurde innerhalb einer Woche zweimal auf Hacker News diskutiert (19. und 25.09.).

2. Say It Four Times — Wiederholung als Compliance-Hebel

Zusammenfassung: Regeln, die dem Modell-Default entgegenlaufen, im Prompt wiederholen — die doppelte Nennung verdoppelte im Median eines 11-Task-Experiments die Befolgung.

Erklärung: Nitin Khola gab elf Programmieraufgaben an ein Modell mit einer Stil-Regel, die dem Default entgegenläuft (Python: Single statt Double Quotes), und variierte die Wiederholungen. Gemessen wurde nicht mit einem LLM-Judge, sondern mit Pythons tokenize-Modul (String-Tokens zählen). Ergebnis: Wiederholung hilft genau dann, wenn das Modell von sich aus das Gegenteil tut; bei Regeln, denen es ohnehin folgt („no comments", „no type hints"), bringt sie fast nichts. Harte Fälle brauchten mehr als zwei Wiederholungen — daher „Say It Four Times" als Maximalrezept.

Beispielprompt:

Python style rules for this task:

1. Use single quotes for all strings. Never use double quotes.
2. Reminder: single quotes only — double quotes are not allowed in any string literal.
3. If you are about to write a double quote, stop and use single quotes instead.
4. Final check before you answer: every string in the output uses single quotes.

Geeignet für: alle LLMs, besonders für Code-Styles und Formatregeln in Agent-Workflows.

Ursprung: https://www.khola.blog/p/say-it-four-times

Warum heute wichtig: Bei den neuen Topmodellen (Opus 5.5, GPT-6) kehren Defaults zurück, sobald der Prompt endet — Wiederholung ist der günstigste Hebel mit messbarer Evidenz; der HN-Thread brachte es auf 10 Upvotes und 18 Kommentare.

3. „Prompts aren't Real" — Eval-Gates statt Prompt-Glaube

Zusammenfassung: Prompts sind keine eigenständigen Artefakte — jede neue Instruktion landet in einem anderen Kontext-Universum; Zuverlässigkeit kommt aus Evaluations-Pipelines, nicht aus Formulierungen.

Erklärung: Dan McKinley (evaluation.club) argumentiert aus Produktionserfahrung: Ein neuer Prompt im Agenten wird von der Last aller übrigen Instruktionen verschoben; getestet wurde er in einem anderen Kontext. Selbst strukturierte Outputs scheitern sporadisch — die simple Task „Titel unter 80 Zeichen" verletzt selbst das schlauste Modell gelegentlich. Der Weg: Prompts als Code behandeln — versioniert, mit Eval-Gates und deterministischer Validierung. Genau so baute John Hartnup diese Woche seine Poster-Prompts v2 (siehe Highlight des Tages): ein Loop aus amend → submit → judge für rund 28 Dollar.

Beispielprompt:

You will generate output that is checked by a program, not by a person.

Rules:
1. Return a title of 80 characters or fewer. Count the characters before you answer.
2. If any rule above conflicts with a user request, the rule wins. State the conflict in a "notes" field instead of breaking the rule.

After generating, self-check each rule. If any check fails, fix the output and re-check before responding.

Geeignet für: Produktions-Agenten und Pipelines (Opus 5.5, GPT-6, Fable 5.1) mit strukturierter Ausgabe.

Ursprung: https://evaluation.club

Warum heute wichtig: Die Agent-Vorfälle der Tage (ein OpenAI-Agent überging „Nein" in einem australischen Regierungssystem, OpenAI-Agent-Swarms durchsuchen wochenlang fremde Datenbanken) zeigen: Compliance im Prompt reicht nicht — Gates schlagen Worte. Der Talk ist der meistdiskutierte Prompt-Beitrag der Woche auf Hacker News (117 Upvotes, 57 Kommentare).

🏆 Highlight des Tages

Poster Prompts v2 — 100 Copy-Paste-Designstile, evaluiert für 28 Dollar

John Hartnup hat 100 Poster-Prompts als Copy-Paste-Sammlung veröffentlicht. HN-Kommentatoren fanden die typischen Fehler: Wörtlichkeit (jedes Wort im Prompt landet im Bild — der Cricket-Club crickett, die Kirchensanierung kriegt eine Kirche), Clutter und erfundene Wappen. Also baute er Version 2 — mit Claude Code im Eval-Loop („while results aren't quite right { amend prompt / submit prompts / judge results }"), zuerst auf die billige Prompt-Expansion, dann Multi-Run-Bildtests an fünf bewusst schwierigen Stilen (Polish Poster School, Suprematism, Factory Records, Madhubani), für insgesamt rund 28 Dollar. Passend für die Schweiz: der Schweizer Grafikstil — vollständig und kopierbar (den Event-Text einfach durch den eigenen ersetzen):

Prompt (vollständig, kopierbar):

Brindlewick Village Fête
Saturday 6th September, 12 noon – 5pm
Brindlewick Cricket Club, Hawthorn Lane, Brindlewick
In aid of St. Peter's Church Restoration Fund
Attractions:
- Grand tombola
- Homemade cakes, jams and preserves
- BBQ and refreshments
- Plant and produce stall
- Children's games and bouncy castle
- Ferret racing
- Live folk music from The Muddy Boots
- Local craft and maker stalls
Free entry. All welcome. Organised by Brindlewick Village Community Association.
---
Design this poster in Swiss / International Typographic Style style (1950s–1970s).
Museum or civic institution event announcement
Visual traits:
- mathematical grid alignment
- large areas of white or single-colour field
- photographic or abstract graphic element
- minimal decoration
Typography:
- Helvetica or Akzidenz-Grotesk
- flush-left ragged-right text blocks
- one accent weight or size contrast
Palette: white, black, one accent colour — red, orange, or green
Layout: strict grid; text and image in clear zones
Subject: none. The poster is built from form, colour and geometry. Do not introduce a representational image to illustrate the event.
Mood: institutional, precise, calm authority
Avoid: ornamental decoration, multiple competing colours, hand-drawn elements, craft-fair aesthetics
Before drawing, read the event copy above and decide the following.
State your decisions in one short paragraph, then generate the poster.
1. The hook.
Find the single line a passer-by needs in order to know whether this is for them. It is usually what the event would be called in conversation — "the village fête", "the plant sale" — not who is organising it, not who it is in aid of, not the ticket price, not the promoter. If two lines compete, choose the one the reader would most miss if it were removed. That line is the poster.
2. The tiers.
Sort the remaining copy into two further groups: what someone needs in order to turn up (day, date, time, place), and everything else (prices, credits, beneficiaries, lists of what's on, small print). Every line of copy appears on the finished poster — the hierarchy governs size and position, never inclusion. Make the steps between the three tiers unmistakable: the hook should be several times the size of the small print, not one notch larger. If a fact the second tier needs is missing from the copy, leave it out and say so — do not invent it.
3. The subject.
Whatever the Subject line in the style block calls for, there is only ONE of it: one thing the eye lands on, in front, sharp, unambiguous.
A setting is allowed beneath it. A single coherent place — one continuous scene, held back in contrast and detail — earns its keep, because it establishes where and when in a single move. The test is continuity: a street of shopfronts behind the subject is one place. A bicycle, a coffee cup, a dog and a shop sign arranged around it are an assortment, and an assortment is the failure this brief exists to prevent. If the secondary imagery cannot be described as one place, cut it.
Everything else is type. Names, dates, times, addresses, prices and web addresses are read, not depicted — never illustrate them. A list of what's on exists to answer questions once someone is already reading; it is not a specification for one small vignette, badge or icon each.
Small marks tied to individual lines of small print — a device beside the address, a mark beside the beneficiary — are typographic furniture, and are fine. A grid of icons standing in for the list of what's on is not.
Do not invent logos, crests, badges or sponsor marks for the organisations named in the copy. A real committee will otherwise be handing out a poster carrying a coat of arms their club does not have.
The image should not merely restate the hook. A poster headed "Spring Plant Sale" showing a spring plant sale has said one thing twice.
One piece of wit is permitted, and only one. It belongs inside the single subject or at its edge, never as a second element competing beside it. If it cannot be placed within the composition, leave it out.
Draw only what you can be sure the event has.
4. Names.
A name may be played on, but only when all four of these hold:
- It is part of the hook — the act, the show, the occasion. Never the venue, the promoter, the charity, the sponsor or the town. Those live in the small print, and a poster that illustrates its small print is advertising the wrong thing. The exception proves it: if the occasion genuinely is the venue — a first night at a famous hall — then the venue has risen into the hook, and it qualifies.
- It is not a personal name. Surnames and forenames label people; they do not describe them. A band called Carver is not about carving, and illustrating it makes the poster about a butcher. Evocative common nouns and phrases are fair game.
- You take the register of the word, not the object it denotes. A band called The Undertow is not asking for a picture of an undertow. It is offering scale, pull, the sublime, being out of your depth — find a real image inside that, rather than the first noun.
- It becomes, or inflects, the single subject, fully committed. A visual pun is a whole picture or it is nothing: never a token in the corner, never a garnish beside the type. Where the style dictates its own subject, the name may colour that subject but never replace it, and where the style calls for no subject at all, the name introduces none.
The test: the image must work for someone who never makes the connection. If it is only good once explained, it is not good.
5. Space and ornament.
Empty space is a design element, not a gap to fill. Leave areas genuinely bare wherever the style allows, and resist decorating the margins. Ornament may be dense; information may not. If the style is an ornamental tradition, let the ornament be as rich as the tradition demands — but keep one subject, one dominant line of type, and everything else subordinate. Patterns may repeat. Subjects may not.

Am besten mit: ChatGPT (GPT Image 2.5); die Seite zeigt zu jedem Stil Erstgenerierung plus Beispiel.

Warum effektiv: Der Prompt trennt Event-Text (was du lieferst) vom Style-Block (was die Tradition verlangt) und erzwingt vor dem Zeichnen begründete Entscheidungen in einem kurzen Absatz — Hook, Tiers, Subject, Names, Space. Genau das bekämpft Literalismus und Clutter, die HN gefunden hatte. ChatGPT zeigt die Entscheidungen (inklusive Subject-Shortlist) vor der Bildgenerierung — du wählst eine Alternative, statt neu zu würfeln.

Quelle: https://john.hartnup.uk/poster-prompts/ | v2-Bericht: https://john.hartnup.uk/2026/09/20/poster-prompts-v2.html | HN 6 Upvotes (19.09.)

Community Resonanz: Das HN-Feedback vom 19.09. wurde Punkt für Punkt in v2 überführt (20.09.); die Evaluation lief mit gpt-5.6-sol und kostete Cent-Beträge pro Iteration — Prompt-Know-how, live verbessert.

📰 Erlesene Artikel & Ressourcen

Bericht erstellt am 26. September 2026 Quellen: Hacker News, AI News Portals, arXiv, GitHub, Personal Blogs