yrzhe/claude-skills/seed

Session Seed (auto-recorded, on-demand synthesized)

A Stop hook silently records every turn of every Claude Code session into a per-session markdown file. When the user wants to tweet about what they just did, you read that file and help them turn raw activity into a draft.

Architecture

~/.claude/
├── hooks/capture-session-seed.sh    # Stop hook → appends turn record
├── skills/seed/
│   ├── SKILL.md                     # (this file)
│   ├── scripts/
│   │   ├── extract_turn.py          # called by the hook
│   │   └── shot.py                  # /seed shot — screenshot + session binding
│   └── state/
│       ├── sessions/{session_id}.md        # raw turn logs, one per session
│       ├── sessions/{session_id}/shots/    # screenshots bound to that session
│       └── archive/                        # optional: consumed seeds moved here

The hook is mechanical — it extracts the last user prompt + all assistant text/tool_use blocks for that turn and appends a markdown block. No scoring, no synthesis, no LLM call. De-duped by user-prompt uuid so repeat Stops on the same turn don't double-write.

When User Invokes

/seed — read-and-synthesize flow

1. Get current session's file:

• Read $CLAUDE_SESSION_ID if exported, otherwise ask the user (the ID is the jsonl filename in ~/.claude/projects/.../<session-id>.jsonl)
• Path: ~/.claude/skills/seed/state/sessions/{session_id}.md

1. Read the full session log. It's a flat list of turns, each with:

• Timestamp
• User prompt
• Tools used (name + short input summary)
• Full assistant output

1. Synthesize: scan for the genuinely tweet-worthy moments — specific decisions, non-obvious observations, concrete data points, real problems solved. Skip routine back-and-forth.
2. Present 2–3 draft tweet angles tied to actual evidence from the log. Each draft must:

• Reference a real action from the log (tool used, file touched, problem faced)
• Include specific detail (number, tool name, timing)
• Lead with reframe or observation, not agreement
• Sound human: no AI-slop phrases ("The real X isn't Y, it's Z", "Happy to...", "Makes me think...")
• 200–350 chars for replies; longer is fine for standalone originals

1. Ask which draft the user wants refined.

/seed <one-line description> — manual override

User wants to capture something that ISN'T in the log (e.g., a non-Claude moment). Save a manual entry with their description; generate drafts as above.

/seed done — archive/delete

After user posts, move the session file to state/archive/{YYYY-MM-DD}_{session_id}.md (or just delete if the user prefers). Reduces clutter and marks the raw material as consumed.

/seed list — show all current session logs

Print filenames + first turn timestamp + file size. Useful when a user wants to pick material from a past session they remember.

/seed shot — capture a screenshot and bind to current session

Two modes, delegated to scripts/shot.py:

# Interactive: cursor turns into a camera, user clicks the target window
python3 ~/.claude/skills/seed/scripts/shot.py --session-id $SID
 
# Headless: screenshot a URL
python3 ~/.claude/skills/seed/scripts/shot.py https://example.com --session-id $SID
python3 ~/.claude/skills/seed/scripts/shot.py localhost:3000 --session-id $SID --note "dashboard after latency fix"

Always pass `--session-id` with the CURRENT session's id. You (the agent) know which session you're in — the transcript path is ~/.claude/projects/.../{session_id}.jsonl, and $CLAUDE_SESSION_ID is often set. Whichever session triggered the shot is the one it should bind to. Without --session-id, shot.py falls back to "most recently modified session file" — which is wrong when multiple Claude Code windows are running in parallel.

Output goes to state/sessions/{session_id}/shots/{timestamp}.png and appends a **📸 Shot** marker to the session's .md file. During /seed synthesis, read those markers and reference the image paths when a tweet draft would benefit from a screenshot.

Uses:

• Interactive mode when the tweet-worthy thing lives in a native app or needs multi-window framing (no Accessibility permission needed — screencapture -w does it). macOS only lets one screencapture -w own the cursor at a time, so parallel sessions can't collide.
• URL mode for dev servers, public pages, or anything Chrome can render headlessly (--headless=new at 1280×800 by default; override with --width/--height)

When the user says "拍一下 / 截图 / shot this / grab a screenshot" during a work session, run the appropriate mode unprompted.

Synthesis Guidelines

• Pure facts in, judgment on request out. The log is raw. Your job at /seed invocation is to add the judgment the hook intentionally skipped.
• Verify before citing. If you draft "fixed a naive-datetime bug in notion_check_pending.py", confirm that's what the log actually shows (tool = Edit on that file, not just mentioning it).
• Prefer one good draft over three generic ones. If only one angle from the session is actually tweet-worthy, say so.
• Skip if nothing fits. If the session was routine (config edits, doc reads, no real discoveries), tell the user plainly — don't manufacture a tweet.

Format of Raw Session File (hook output)

Example of what the hook writes per turn:

## 2026-04-20 22:59:52
<!-- turn-uuid: abc-...-xyz -->
**cwd:** `/path/to/project`
 
**User:**
 
扫一下
 
**Tools:**
-`Bash` — python3 ~/.claude/skills/tweet-monitor/scripts/scan.py ...
-`Read` — /Users/.../notion_check_pending.py
 
**Assistant output:**
 
(full markdown body of what Claude said, untruncated)
 
---

The turn-uuid marker is how the hook dedupes repeat invocations.

Tweet Draft Rules (re-applied during synthesis)

• Lead with the result, not the process
• Include a specific detail (time, number, tool name)
• Write as "I" not "we"
• No "The real X isn't Y, it's Z" pattern
• Must sound like a human talking, not AI writing
• Reinforce identity: "AI PM shipping with Claude Code"

What the Hook Does NOT Do

• It does NOT score / rank / filter turns
• It does NOT summarize — full assistant output is preserved
• It does NOT touch Google Drive (old location is deprecated)
• It does NOT invoke any LLM — zero cost, zero latency

If you want LLM-level filtering, that's what /seed invocation is for. The hook stays dumb by design.