AI Capsule — Personal AI News Value Evaluator#

Scores and ranks your daily AI news feed based on what you actually care about. Tell it your role and familiar areas — it scores every article against your profile and ranks them so the most relevant ones appear first. Nothing is filtered out: every article makes it into the digest, just in the right order. Sources include HuggingFace Papers, OpenAI, Anthropic, DeepMind, Simon Willison, GitHub Trending, HN, Reddit, and more (17 total), plus X/Twitter accounts.

Trigger words: run the full digest by saying daily, daily mode, 每日模式, 每日摘要, or 日报. To avoid firing during ordinary conversation, the bare word daily only triggers as a short command — the message must start with a trigger word and contain little else (≤ ~6 words / ≤ 5 tokens after the trigger). daily appearing mid-sentence (e.g. "I read the daily news", "今天的 daily report 还没看") does NOT trigger the digest. When ambiguous, ask the user to confirm before fetching.

First time: run bash $SKILL_DIR/scripts/setup-env.sh once, then say daily — Claude will walk you through a short setup (role, familiar areas, output language).

Initialization#

First time only: run bash $SKILL_DIR/scripts/setup-env.sh to set up the Python venv. Config is created on first run via guided setup.

On every run, resolve the skill directory, Python interpreter, and config path:

bash

SKILL_DIR=$(cd "$(dirname "$(readlink -f "${BASH_SOURCE[0]:-$0}" 2>/dev/null || echo "$0")")" && pwd)
CONFIG_FILE="${AI_CAPSULE_CONFIG:-$HOME/.ai-capsule/config.yaml}"

# Prefer venv; fall back to system python3 for spawned/headless agents (OpenClaw, CI)
if [ -x "$SKILL_DIR/.venv/bin/python" ]; then
  PYTHON="$SKILL_DIR/.venv/bin/python"
else
  PYTHON="$(command -v python3 || command -v python)"
  echo "WARN: venv not found, using system Python ($PYTHON). Run: bash $SKILL_DIR/scripts/setup-env.sh"
fi

echo "SKILL_DIR: $SKILL_DIR"
echo "PYTHON: $PYTHON"
echo "CONFIG_FILE: $CONFIG_FILE"
cat "$CONFIG_FILE" 2>/dev/null || echo "initialized: false"

If initialized: false (or config file missing) → enter the guided setup flow (Step -1), otherwise jump to Step 0.

---

User Configuration#

Config file: ~/.ai-capsule/config.yaml — lives outside the skill directory, survives updates.

Override: set env var AI_CAPSULE_CONFIG=/path/to/config.yaml to use a different file.

Built-in identity profiles:

• engineer: Is it actionable in a real engineering context?
• pm: Does it affect product direction, UX, or business decisions?
• researcher: Does it advance understanding of a technical principle?
• learner: Is this direction worth investing time to learn systematically?
• founder: Does it affect tech stack choices, product direction, or competitive positioning?

Built-in purpose profiles (control scoring weights):

• learn: D×0.35, N×0.25, R×0.2, U×0.1, WPS×0.1
• solve: U×0.45, R×0.3, D×0.1, N×0.1, WPS×0.05
• scout: W×0.25, P×0.2, N×0.2, S×0.15, R×0.15, U×0.05

---

Usage#

• Single article: paste title + body directly
• Batch: paste a JSON array [{"id":"1","title":"...","content":"..."},...]
• URL: paste a link starting with http or https
• Daily: say "daily" or "daily mode"
• Reconfigure: say "reconfigure" → Claude resets initialized: false and walks through setup again

---

Execution Steps#

Step -1: Guided Setup (first run only)#

Runs when initialized: false or config file is missing.

Question 1 — Your role

What role do you primarily take when consuming AI news?

• A) Application engineer / full-stack engineer → engineer
• B) AI researcher / ML engineer → researcher
• C) Product manager / founder → pm or founder
• D) Student / self-learner → learner
• E) Other (describe freely)

Question 2 — Your familiar areas

e.g. LLM application development, RAG, Agent frameworks, model fine-tuning…

Question 3 — Your primary purpose

• A) Follow the technical frontier (learn)
• B) Solve concrete problems at work (solve)
• C) Scan market trends (scout)
• D) Context-dependent (default: learn)

Question 4 — What you don't want to see

e.g. pure marketing fluff, trend analysis with no code…

Question 5 — Output language

What language should card text, value analysis, and scoring reasons be written in?

• A) Chinese / 中文 (default)
• B) English
• C) Other (describe freely, e.g. "日本語")

Proper nouns (model names, framework names, GitHub repo names) always stay in their original form.

Question 6 — Data directory

Where should daily reports, history, and dedup files be stored?

• A) ~/.ai-capsule/data (default — survives skill updates)
• B) Custom path (describe freely)

After collecting answers, write the config:

bash

mkdir -p ~/.ai-capsule
cat > "${AI_CAPSULE_CONFIG:-$HOME/.ai-capsule/config.yaml}" << 'EOF'
initialized: true
identity: [user description]
role_desc: [user description]
familiar_areas:
  - [area 1]
default_identity: [engineer/pm/researcher/learner/founder]
default_purpose: [learn/solve/scout]
output_language: zh  # zh = Chinese / en = English / other language name
data_dir: ~/.ai-capsule/data
EOF

---

Step 0: Mode Detection#

• Input starts with http(s):// → URL mode
• Input starts with [ and looks like a JSON array → Batch mode
• Input contains an explicit multi-word trigger — daily mode, 每日模式, 每日摘要, or 日报 — → Daily mode (matches anywhere in the message)
• Bare daily → Daily mode only if it is a short command: the message starts with daily and has ≤ 5 tokens after it (roughly ≤ 6 words total). daily mid-sentence (e.g. "I read the daily news", "今天的 daily report 还没看") does NOT trigger.
• Ambiguous → ask the user to confirm before fetching

Step 1: Content Retrieval#

Fetch tool mapping — use whichever tool your agent runtime provides:

RuntimeToolClaude CodeWebFetchCodex / shell agentsBash: curl <url> + python parseTavily agentstavily_extractFallbackAsk user to paste content

URL mode: Fetch and extract main text. For WeChat links, ask user to paste manually.

Batch mode: Parse JSON, extract title and content/ text fields, score each article.

Daily mode: Read and execute $SKILL_DIR/sections/daily-mode.md.

Single article mode: Go directly to Step 2.

Step 2: Dedup Check#

Dedup file: {data_dir}/dedup-titles.txt (one title per line).

Daily mode: handled centrally in sections/daily-mode.md Step 3 — never judge manually.

Single / URL / Batch mode:

bash

grep -Fx "article title" "$(
  "$PYTHON" -c "
import os, pathlib, yaml
cfg = yaml.safe_load(open(os.environ.get('AI_CAPSULE_CONFIG', str(pathlib.Path.home()/'.ai-capsule/config.yaml'))))
print(cfg.get('data_dir','~/.ai-capsule/data').replace('~', str(pathlib.Path.home())))
")/dedup-titles.txt"

If output is non-empty, skip and say: "This article has already been scored (title: XXX)."

Step 3: Scoring#

Read default_identity, default_purpose, and output_language (default: zh) from $CONFIG_FILE.

Scoring framework: see $SKILL_DIR/sections/scoring.md (Read the file before every run).

Runtime mode switches:

• "learn" / "学习" → PURPOSE = learn
• "solve" / "解决问题" → PURPOSE = solve
• "scout" / "找灵感" → PURPOSE = scout
• "pm perspective" / "engineer perspective" / etc. → switch IDENTITY

Always output the score_json block first:

text

<score_json>
{
  "title": "article title",
  "source": "hf | openai | anthropic | hn | reddit | github | producthunt | blog | unknown",
  "url": "taken directly from pending.json or fetch result — never construct from title",
  "identity": "engineer",
  "purpose": "learn",
  "R": 8, "U": 7, "N": 6, "D": 8,
  "W": 5, "P": 6, "S": 4,
  "total": 7.2,
  "F": 9,
  "final_score": 7.2,
  "action": "READ",
  "summary": "article entity + core value",
  "reasons": {
    "R": "specific relevance reason",
    "U": "specific utility reason",
    "N": "specific novelty reason",
    "D": "specific depth reason",
    "W": "specific wow reason",
    "P": "specific perspective reason",
    "S": "specific cross-domain reason"
  }
}
</score_json>

F: Personal Fit (0–10), assessed against familiar_areas and dislikes in config. final_score = total × (0.7 + 0.3 × F/10) — sorting only, never displayed. action: TRY (runnable code/tool) / READ (deep content) / SCAN (brief/low score).

Step 4: Display Output#

Output format: see $SKILL_DIR/sections/output-format.md (Read the file before every run).

Language rule: all card text follows output_language. Always keep in original form: article titles, technical terms, model/framework names, GitHub repo names, URLs, Action tags (TRY/READ/SCAN).

Batch / Daily mode: output detailed cards only, sorted high-to-low by score — no summary table.

Step 5: Append Records#

Append to {data_dir}/history.jsonl (with timestamp) and {data_dir}/dedup-titles.txt.

---

Data Source Management#

Source configs: $SKILL_DIR/sources/{industry}.yaml

View sources:

bash

"$PYTHON" "$SKILL_DIR/sources_extract.py" --markdown --industry ai

Add a source:

bash

bash $SKILL_DIR/scripts/add-source.sh --industry ai --type rss     --name "Name" --url https://example.com/feed
bash $SKILL_DIR/scripts/add-source.sh --industry ai --type webfetch --name "Name" --url https://example.com --note "hint"
bash $SKILL_DIR/scripts/add-source.sh --industry ai --type tavily  --name "Name" --query "search terms" --domains "reddit.com"
bash $SKILL_DIR/scripts/add-source.sh --industry ai --type twitter --name "Name" --accounts "handle1,handle2"

Source schema:

yaml

# RSS — fetched automatically by fetch.py
- name: Source Name
  type: rss
  url: https://...
  limit: 5
  handler: huggingface    # optional: huggingface | hacker_news

# URL fetch — agent fetches in daily Step 2
- name: Source Name
  type: webfetch
  url: https://...        # use YYYY/M/D placeholder for date-based URLs
  limit: 5
  note: "hint for the agent"

# Tavily search — agent searches in daily Step 2
- name: Source Name
  type: tavily
  query: "search terms"
  include_domains:
    - reddit.com
  time_range: day         # day | week | month
  limit: 5

# X/Twitter — fetched automatically by fetch.py (requires Chrome login)
- name: X/Twitter
  type: twitter
  accounts:
    - handle1
    - handle2
  limit: 5

New industry: create $SKILL_DIR/sources/{industry}.yaml using ai.yaml as template.

---

Source: github.com/WebPudge/ai-capsule · Feedback / Bug report → Issues