Anthropic Research Fetch

---
name: anthropic-research-fetch
description: Recurring pipeline for fetching Anthropic research articles and turning reusable methodologies into ai_collection skills. Triggers on scheduled/cron Anthropic research fetches, "fetch anthropic research", extracting methods from anthropic.com/research, or syncing new skills to the ai_collection project. Covers fetch → article extraction (with the web_extract-blocked workaround) → skill creation → ai_collection sync (skills + INDEX.md) → kg.db update → Obsidian notes.
license: Complete terms in LICENSE.txt
---

# Anthropic Research Fetch → Skill Pipeline

Recurring cron task: pull articles from `https://www.anthropic.com/research`, extract reusable methodologies, and persist them into the `ai_collection` skill library + knowledge graph + Obsidian.

## Pipeline steps

1. **Fetch**: run the fetch script.
   ```
   python3 /Users/hiyenwong/.openclaw/workspace/scripts/fetch_anthropic_research.py
   ```
   Saves to `/Users/hiyenwong/.openclaw/workspace/obsidian/anthropic_research.json` with keys: `fetch_date`, `source_url`, `total_items`, `items[]` (each has `title`, `url`, `category`, `date`, `summary`).

2. **Extract article bodies** — see the pitfall below (web_extract is BLOCKED for anthropic.com). Use `browser_navigate` + `browser_snapshot` instead.

3. **Select skill candidates**: prefer articles with a clearly reusable CROSS-DOMAIN methodology (alignment training, interpretability technique, capability-control architecture, agent-reliability pattern). Skip domain-specific or economic/usage analyses unless they surface a transferable method.

4. **Create skills**: follow `skill-creator` format — write `SKILL.md` with frontmatter (`name`, `description` incl. triggers) + body (numbered steps, pitfalls, **Activation** keyword list). Save to `~/.hermes/skills/ai_collection/{slug}/SKILL.md`.

5. **Sync to ai_collection project**:
   - Copy skill dir → `/Users/hiyenwong/ai_github/ai_collection/collection/skills/{slug}/`
   - Append a dated section to `/Users/hiyenwong/ai_github/ai_collection/INDEX.md` (see references/INDEX_FORMAT.md for the exact template — `## YYYY-MM-DD - Anthropic Research (Cron Job)` header, `[[slug]]` wikilink entries, bullet "Core point:" lines, and a `**Activation**:` keyword line).
   - Commit + push:
     ```
     cd /Users/hiyenwong/ai_github/ai_collection
     git add collection/skills/{slug}/ INDEX.md
     git commit -m "feat: add {slug} from Anthropic research"
     git push
     ```
   - Only `git add` the files you created. Do NOT commit unrelated untracked dirs left by other cron runs (e.g. other skill folders). Verify with `git status --short` before committing.

6. **Update kg.db** (workspace copy at `/Users/hiyenwong/.openclaw/workspace/scripts/kg.db`): insert `skill` entities (columns: `name, type, description, metadata`) with metadata JSON containing `skill_name`, `source`, `date`, `topic`; link each to an "Anthropic research" `concept` entity via `relates_to`. Use INSERT-or-ignore (UNIQUE(name,type)) to stay idempotent.

7. **Save Obsidian notes**: write `/Users/hiyenwong/.openclaw/workspace/obsidian/YYYY-MM-DD-anthropic-research-summary.md` with per-article methodology, key results, skill links, and "Skills Created Today" list. Match prior summary format (see `2026-06-02-anthropic-research-summary.md`).

## Pitfalls

- **web_extract is BLOCKED for anthropic.com** (returns "Blocked: URL targets a private or internal network address" — a false positive from the security scanner). Do NOT waste calls on it. Use `browser_navigate(<url>)` then `browser_snapshot(full=true)` to read article content. This is the single most common time-sink in this pipeline — go straight to the browser.
- **Skill slugs must be class-level and collision-safe.** If a prior run already created a similarly-named skill (e.g. `teaching-claude-why`), use a distinct, more specific slug (e.g. `teaching-claude-why-alignment`) rather than overwriting. Search `collection/skills` before creating.
- **Don't commit other crons' untracked files.** `git status` often shows stray dirs from parallel runs; scope your `git add` to your 4 skills + INDEX.md.
- **web_extract may also be blocked for other vendor domains** (e.g. openai.com, deepmind.google) with the same false error — default to browser_navigate there too.

## Support files
- `references/INDEX_FORMAT.md` — exact INDEX.md section template + a worked example from the 2026-07-14 run.
- `references/web_extract_blocked_workaround.md` — full detail on the blocked-web_extract issue and the browser_navigate sequence.

## Activation keywords
anthropic research fetch, fetch anthropic research, anthropic.com/research extraction, ai_collection skill sync, web_extract blocked anthropic, create skill from anthropic research, cron research pipeline