article

> **${var}** — Selector: `[angle:arg] [--visual]`. The **angle** prefix picks the article type; append **`--visual`** (or `visual`) anywhere to also generate a Replicate hero image.
>
> - **empty** → `standard` general long-form article on an auto-selected trending topic. If the resolved topic is a single explainable mechanism, it becomes a technical explainer instead.
> - **`<topic>`** (no recognized prefix) → `standard` article on that topic.
> - **`repo:<owner/repo>`** → `repo` thesis-driven article about that repo. `repo:<angle>` (e.g. `repo:architecture`) or bare `repo:` uses the repo from `memory/watched-repos.md` with that angle / an auto-selected angle — this preserves repo-article's original input.
> - **`lens:<topic>`** → `lens` project-through-a-lens essay framed by that lens (e.g. `lens:unix philosophy`). Bare `lens:` auto-selects the lens.
> - **`--visual`** appended to any of the above → after the body is written, generate a Replicate hero image (optional `REPLICATE_API_TOKEN`; ships text-only if absent).
>
> Examples: `""`, `"entropy trajectory reasoning --visual"`, `"repo:aeonfun/aeon"`, `"repo:roadmap"`, `"lens:regulation wave --visual"`.

Today is ${today}. Write a high-quality, publication-ready article. No placeholders.

## Shared preamble (every run)

1. Read `memory/MEMORY.md` for context on what topics/articles have been covered recently.
2. Read the last 3–7 days of `memory/logs/` for recent activity — and **don't re-report** something already covered.
3. **Parse `${var}` into `angle` + `visual`:**
   - Detect a standalone `--visual` or `visual` token anywhere in `${var}`; if present set **`visual = true`** and strip that token. Otherwise `visual = false`.
   - From what remains: if it starts with `repo:` → `angle = repo`, `arg =` the rest. If it starts with `lens:` → `angle = lens`, `arg =` the rest. Otherwise → `angle = standard`, `arg =` the whole remaining string (empty ⇒ auto-select).
4. Dispatch to the matching angle section below. If `visual = true`, run the **Visual add-on** after the article body is written, regardless of angle.

---

## Angle: standard — long-form article / technical explainer

A single long-form article. It takes one of two structures depending on the topic:

- **General article** — a broad trend, development, or event. 600–800 words.
- **Technical explainer** — a single explainable mechanism, technique, algorithm, or system. 600–1000 words, using the explainer structure below.

### Topic selection (standard)

- If `arg` (the topic) is set, use it verbatim. If it clearly names a single mechanism/technique/system → **technical explainer** structure; otherwise → **general article** structure.
- If `arg` is empty, pick deterministically — first hit wins:
  1. **Explainer candidate:** a single most non-obvious mechanism inside the newest file in `output/articles/` from the last 3 days; else the newest "Paper Pick" in `memory/logs/` from the last 7 days (its headline mechanism); else a specific technique/algorithm/system surfaced in the last 7 days of logs. If a strong single-mechanism candidate exists → **technical explainer** on it. Reject any candidate broader than a single mechanism (e.g. "AI agents" — too vague; "MCP tool-routing via vector search" — usable).
  2. **General candidate:** otherwise search the web for the most interesting recent development in AI, crypto/DeFi, or consciousness research — pick whichever has the most compelling story today (WebSearch) → **general article**.

### Voice (technical explainer)

If a `soul/` directory exists, read the soul files for voice calibration: `soul/SOUL.md` (identity, worldview, opinions), then `soul/STYLE.md` (writing style, sentence structure, anti-patterns). This is *you explaining a mechanism to a smart friend* — more precision than a general article, same voice. No textbook tone, no "let's explore." If `soul/` is empty, default to clear, direct, neutral.

### Research

**General article:** read 2–3 source articles with WebFetch to gather facts and quotes.

**Technical explainer:** run **three distinct WebSearch queries** so you triangulate rather than echo one source:
1. `"<topic>" how it works` — mechanism explanations
2. `"<topic>" benchmark OR results OR latency OR cost` — concrete numbers
3. `"<topic>" limits OR criticism OR fails OR doesn't work` — failure modes and pushback

If the topic is from a paper, also fetch the paper metadata and abstract:
```bash
curl -s "https://api.semanticscholar.org/graph/v1/paper/search?query=TOPIC&limit=5&fields=title,authors,abstract,url,publicationDate,openAccessPdf" \
  || echo "curl failed — use WebFetch on https://www.semanticscholar.org/search?q=TOPIC instead"
```
Use **WebFetch** to read the 2–3 best sources in depth. **At least one source must be primary**: a paper (arXiv / OpenReview / Semantic Scholar), official documentation, the project's own README, or a code repo. Blog summaries alone are not enough — they often mangle the mechanism.

Extract:
- The **single core mechanism** — the one move that, once you grok it, makes the rest fall into place.
- A **vivid analogy** for the mechanism, and the precise place where the analogy breaks down (the breakage is the interesting part).
- **3–5 specific numbers** — benchmarks, latencies, costs, error rates, training compute, parameter counts. Each number gets a source URL.
- **What would falsify this** — what result, if observed, would mean the mechanism doesn't work as claimed. If you can't name one, the explanation isn't sharp enough — keep digging.

### Write

**General article** — 600–800 words in Markdown. Include:
- A compelling title
- A short intro hook
- 3–4 substantive sections
- Cited sources (with URLs) at the bottom

**Technical explainer** — 600–1000 words. Structure (every section required):
```
# <Title>

**Key idea in one sentence:** <one-sentence claim about the mechanism>

## The Setup
2-3 sentences. What problem does this solve? Why now?

## The Intuition Pump
A vivid analo