GitHub Radar

An open-source intelligence engine for AI PMs. Four modes, one Layer analysis framework.

Language Selection

This skill ships with both English and Chinese versions. The agent should automatically match the user's language:

• - If the user speaks English -> Use skill.md, agents/analyzer.md, references/layer_model.md, templates/*.html, and generate English reports
• If the user speaks Chinese -> Use skill_cn.md, agents/analyzer_cn.md, references/layer_model_cn.md, templates/*_cn.html, and generate Chinese reports
• Scripts (scripts/) and config (config/) are language-neutral and shared by both versions

When to Use

• - "What's worth looking at today?" / --pulse -> Mode 1
• "Find me GitHub projects related to [topic]" -> Mode 2
• "Monitor anomalous signals" / --watch -> Mode 3
• "Analyze the ecosystem around [repo]" -> Mode 4
• Any need involving GitHub project discovery, trend analysis, or paradigm assessment

File Structure

CODEBLOCK0

Dependencies

Dependency	Requirement	Check Command
gh CLI	>= 2.40.0, authenticated	INLINECODE12
Python

>= 3.9 | python --version | | Extra Python packages | None, stdlib only | — | | API quota | 5,000 requests/hour when authenticated | python scripts/check_rate_limit.py |

Common Prerequisites

Must be completed before running any mode:

CODEBLOCK1

Determine the execution strategy based on the returned mode field:

• - full -> Normal execution, including star history fetching
• INLINECODE17 -> Skip fetchstarhistory.py, use basic data only
• INLINECODE18 -> Run search scripts only, skip detail API calls

---

Mode 1: Proactive Exploration (Radar Pulse)

Trigger: --pulse or "What's worth looking at today?"

Execution Steps

CODEBLOCK2

Filtering Rules

1. 1. Label each candidate's Layer
2. Remove L1 (model-level, too low-level) and L5 (wrapper/demo, noise)
3. PM value weighting: L2 x 1.5, L3 x 1.3, L4 x 1.0
4. Take Top 3-5, deep-dive into 1-2

Output Format

CODEBLOCK3

Report saved to: output/radar-pulse_{date}.md

---

Mode 2: Direction Search

Trigger: User provides a technical direction or keywords

Execution Steps

Step 1: Check Quota

CODEBLOCK4

Step 2: Keyword Expansion + Layer 1 Relevance Review

1. 1. Understand the topic: State the core concept of the user's search in one sentence
2. Expand keywords: Generate 8-15 search keywords around the topic, covering:

- Synonymous expressions (swarm -> fleet, colony) - Scenario-specific terms (swarm observability, coding agent swarm) - Adjacent concepts (coordination, monitoring — things adjacent to swarm)

1. 3. Layer 1 self-review: Review each keyword; the criterion is "would most results returned be about the same category of thing?" — exact match to every word is not required:

- Keep: Results are different angles on the same topic - Remove: Most results fall into a broader category where the topic is only a small subset - Example — topic "agent swarm": swarm orchestration keep (same topic); multi-agent framework remove (swarm is a subset of multi-agent; most results won't be about swarms) - Example — topic "Agent-human collaboration": human-in-the-loop agent keep (same topic); AI assistant remove (assistant != human-agent collaboration)

1. 4. Present to user for confirmation: List kept and removed keywords with rationale; proceed with search only after user confirms

Step 3: Search

CODEBLOCK5

Step 3.5: Dynamic Strategy When Recall Is Low

If deduplicated results < 50, do not silently expand. Instead, present the current situation to the user with three options:

Searched {n} keywords, only {m} unique results after deduplication. Possible reasons and options:
A. This direction hasn't formed a distinct category yet — The relevant capabilities may be embedded as features within larger frameworks rather than existing as standalone projects. Recommend abandoning the search; this finding is itself valuable.
B. Keyword coverage is insufficient — Current keywords may be missing expressions commonly used by the community. I suggest adding the following keywords: {list}. Will continue after confirmation.
C. Proceed with existing results — Although {m} results is a small set, if quality is sufficient, we can go straight to analysis. Suitable for a quick overview of the landscape.

Heuristics for recommending an option:

• - Most keywords return 0 results -> lean toward A (category doesn't exist)
• Only the main keyword has results, expanded terms return nothing -> lean toward B (missing community terminology)
• Few results but highly relevant -> lean toward C (small category with clear signal)

Step 4: Layer 2 Result Relevance Classification

After raw search results are returned and before analysis, classify each repo's relevance:

Classification	Criteria	Handling
high	This project is directly working on the topic	Include in competitive landscape analysis
medium

Related to the topic, but not its primary focus | Include based on quality |
| low | Keyword match was coincidental; project is actually about something else | Filter out, list under "Filtered Out" |

Judgment basis: repo name + description. Ask yourself, "Would this project's author consider themselves working on {user's topic}?"

Step 5: Star History + PM Analysis

CODEBLOCK6

Output Structure

CODEBLOCK7

Reports generated in both HTML and MD: output/search_{keyword}_{date}.html/.md

---

Mode 3: Anomalous Signal Monitoring (Signal Watch)

Trigger: --watch or "Monitor anomalous signals"

Known blind spot: Currently only detects growth anomalies in new projects (created within 90 days). Detecting sudden surges in established projects requires persistent storage for differential comparison — reserved for future iteration.

Execution Steps

Step 1: Check Quota

CODEBLOCK8

Step 2: Candidate Discovery

CODEBLOCK9

The script returns a candidate list (sorted by rough velocity descending), each containing:

• - stars, forks, created, INLINECODE30
• INLINECODE31 = stars / age_days (rough velocity)
• INLINECODE32 = forks / stars (adoption depth signal)

Step 3: Initial Screening + Growth Curve Fetching

1. 1. Exclude obviously irrelevant items: Check descriptions, exclude games, tutorials, awesome-lists, and other non-technical projects
2. Fetch star history for remaining candidates (full mode only):

CODEBLOCK10

Returned growth metrics:

Metric	Meaning
INLINECODE33 / INLINECODE34	Average daily growth
INLINECODE35

7d avg / 30d avg; >1 means accelerating |
| trend_direction | Last 3 days avg / prior 4 days avg; indicates current trend |
| consecutive_growth_days | Consecutive days of growth |
| peak_recency | Days since peak; 0 = today |
| burst_ratio | Peak day / 7d avg; high = spike-type growth |
| recent_7_days[] | Daily breakdown; used to assess growth shape |

Step 4: Growth Pattern Classification

Examine the shape of recent_7_days[] to determine the growth type:

Pattern	Characteristics	PM Implication	Signal Quality
sustained	INLINECODE42 + INLINECODE43	Organic growth, real demand	High
accelerating

trend_direction > 2 + consecutive > 5 | Currently surging, act fast | Highest |
| spike+decay | burst_ratio > 5 + trend_direction < 0.5 | One-time launch burst, likely noise | Low |
| step | Single-day spike + stable before and after | Event-driven (influencer repost) | Medium, watch follow-through |

Step 5: Three-Tier Assessment + PM Analysis

Read agents/analyzer.md and assess each candidate holistically:

• - Worth deep-diving: sustained/accelerating pattern + L2/L3 layer
• Watch: Has growth signal but pattern is unclear, or step-type awaiting follow-through
• Ignore: spike+decay + L5 wrapper / tutorial / fork_ratio < 0.02

Output Structure

CODEBLOCK11

Report saved to: output/signal-watch_{date}.html

---

Mode 4: Deep Link Analysis

Trigger: User provides a repo URL or owner/repo name

Execution Steps

CODEBLOCK12

Output Structure

CODEBLOCK13

Output Style

• - CSS uses the --bg/--surface/--border/--accent/--muted variable system, consistent across modes
• PM insights use .pm-box card component (white background + border), not inline INLINECODE52
• Layer positioning uses .layer-box component with badge + reasoning list + "why not X"
• Paradigm assessment uses .paradigm component (blue background + border)
• Competitor candidates go in <details> collapsed sections
• All technical metrics include plain-language explanations (plain language principle)

Report saved to: output/deep-link_{owner}_{repo}_{date}.html

---

Seed List Customization

Edit config/seed_list.json to add or remove developers you follow:

CODEBLOCK14

The default list currently includes 76 important AI builders/orgs, covering 17 categories including labs, agent frameworks, coding agents, inference, platforms, and more.