Same topic, different lens. Turn what you know into what they want to watch.
AngleCraft is an AI skill that turns dry expert topics into high-engagement content by applying 7 core angles and 3 foundational lenses (10 angles total). Give it a topic - it returns hooks built to earn attention, with placeholders instead of invented facts.
Source of truth: SKILL.md. This README is the human-facing summary; the skill file is what an agent loads.
Most experts produce useful but dull content. They explain what and forget how to make it interesting. The difference between 200 views and 200,000 is rarely the topic. It is the angle: the perspective from which you enter the topic.
A topic is a building. You can photograph the facade, the rooftop, the reflection in a puddle, or the view from a helicopter. Same building, completely different photographs. Same topic, completely different content experiences.
AngleCraft systematizes that choice, then writes the hook. It stops at the angle layer - no padded 1,000-word drafts.
| # | Angle | Pattern | Best for |
|---|---|---|---|
| 1 | Myth vs Reality | "Everyone says X. In reality, Y." | Virality, debate, challenging dogma |
| 2 | Paradox | "The more X, the worse Y." | Cognitive hooks, pattern violation |
| 3 | Level Error | "You think the problem is X. The real cause is Y." | Reframing, intellectual depth |
| 4 | Price of Solution | "Yes, X works. But here is what you really pay." | Trust, credibility, objection handling |
| 5 | Conditions | "If A, do X. If B, do Y." | Self-identification, lead screening |
| 6 | Hidden Driver | "The real reason X happens is Y." | Insider insight, "aha" moments |
| 7 | Contrary Thesis | "Do not do X. Do the opposite." | Debate, contrarian trust |
Conditions is a decision structure, not a perspective. It only works when there are real branches.
| # | Lens | Pattern | Best for |
|---|---|---|---|
| 8 | Historical Lens | How did we get here? | Origin stories, authority |
| 9 | Future Lens | Where is this going in 5-10 years? | Urgency, vision |
| 10 | News Hook | Real event → bridge to the lesson | Immediate relevance |
Full patterns, examples, and close-angle discrimination: references/angles.md.
| Goal | Prioritize |
|---|---|
| Quick trust / credibility | Price of Solution, Hidden Driver |
| Virality / sharing / reach | Myth vs Reality, Paradox, Contrary Thesis |
| Self-identification | Conditions |
| Intellectual depth / authority | Level Error, Historical Lens, Future Lens |
| Immediate relevance | News Hook |
| Platform | Prioritize |
|---|---|
| Short-form (TikTok, Reels, Shorts) | Paradox, Myth vs Reality |
| Long-form (YouTube, blog, newsletter) | Level Error, Hidden Driver, Historical Lens |
| Ads / landing pages | Price of Solution, Conditions |
| LinkedIn / B2B | Contrary Thesis, Level Error, Hidden Driver |
- Reach: Myth vs Reality, Paradox, Contrary Thesis, News Hook
- Trust: Level Error, Hidden Driver, Price of Solution, Historical Lens
- Conversion: Conditions, Future Lens
git clone https://github.com/MADEVAL/AngleCraft.git ~/.agents/skills/anglecraft~/.agents/skills is read by OpenCode and by agents that follow the .agents convention. Other supported locations:
| Scope | Path |
|---|---|
| OpenCode global | ~/.config/opencode/skills/anglecraft/ |
| OpenCode project | .opencode/skills/anglecraft/ |
| Claude Code global | ~/.claude/skills/anglecraft/ |
| Claude Code project | .claude/skills/anglecraft/ |
Generic .agents |
~/.agents/skills/anglecraft/ or .agents/skills/anglecraft/ |
Create the directory if needed and copy the repository contents into it. The folder name becomes the skill ID, so keep it lowercase: anglecraft.
Give me content angles for: I sell handmade ceramic mugs.
Quick Mode answers without clarifying questions. If the topic is broad, it states one assumption and proceeds.
For a full strategy, ask explicitly: Need a content strategy for my B2B SaaS. That triggers Deep Mode.
See references/examples.md for complete Quick and Deep outputs.
Quick Mode: 3-5 rows, columns # | Angle | Hook | Direction. Hook is one sentence, 30 words max. Direction is 1-2 sentences, not the full piece.
Deep Mode: 5-10 rows, columns Topic | Angle | Hook | Tier | Platform, covering at least two funnel tiers.
Both modes add an Assumption: line when the topic was narrowed or the audience inferred, and a placeholder note when data is missing.
Every hook is scored on three criteria:
- Surprise - does it contradict a common assumption or reveal non-obvious information?
- Specificity - is it anchored in a concrete mechanism, trade-off, or decision rule?
- Curiosity gap - does the reader need to know what comes next?
Specificity never means inventing facts. If a hook needs a number that was not supplied, AngleCraft writes a [X] placeholder and tells you to replace it with verified data before publishing. No fake statistics, case studies, or testimonials.
AngleCraft works at the idea layer: angles, hooks, and outlines. It is not a full copy engine and not a research tool.
| Request | Use instead |
|---|---|
| SEO briefs, keyword maps, ranking audits | RankWise |
| Persuasion audits and cognitive-bias copy optimization | MindFluence |
| Rewriting AI text to sound human | HumanAI |
| Full article or script after the angle is chosen | Write it directly |
| Fact-checking or source research | A research/browsing tool - AngleCraft never invents sources |
- Works with any LLM that can load a Markdown skill.
- Tested paths: OpenCode, Claude Code, and the generic
.agents/skillsconvention. - The runtime skill is a single
SKILL.mdplus optional reference files; no scripts or API keys are required. - Supporting files (
references/,examples/,scenarios/) are read on demand, not loaded automatically.
AngleCraft/
├── SKILL.md # runtime skill (source of truth)
├── references/
│ ├── angles.md # full taxonomy, examples, discrimination
│ └── examples.md # worked Quick and Deep outputs
├── examples/ # ready outputs by niche
├── scenarios/ # input -> mode/handoff routing
├── evals/ # 20 test topics + rubric
├── tests/ # structure validator + dataset checker
├── run_tests.ps1 # Windows test runner
├── CHANGELOG.md
├── README.md
├── README.ru.md
├── LICENSE
└── .github/workflows/ci.yml
python tests/validate_skill.py
python tests/eval_dataset.pyWindows PowerShell: .\run_tests.ps1
validate_skill.py checks frontmatter, counts, references, README consistency, relative links, and forbidden strings. eval_dataset.py checks that every eval topic maps to valid angles and includes safety cases.
For manual output scoring, use evals/rubric.md.
AngleCraft refuses to generate content for:
- Pseudoscience presented as fact
- Hate speech, discrimination, or targeting of vulnerable groups
- Get-rich-quick, gambling, or predatory financial schemes
- Unproven medical, legal, or financial claims presented as professional advice
For sensitive topics (health, legal, finance, safety), AngleCraft produces educational framing only - never diagnosis, treatment, guaranteed returns, or legal outcomes - and expects a "not professional advice" disclaimer in the final content.
MIT - free to use, modify, and distribute. See LICENSE.
- RankWise - SEO content engine with 49 ranking factors
- MindFluence - Marketing copy using 20 cognitive biases
- HumanAI - Rewriting AI text to sound human-written (9 languages)
Pipeline: AngleCraft picks the angle and hook → MindFluence shapes the persuasive copy → HumanAI removes the AI voice → RankWise handles search visibility.