# Conversation
Archived: Jun 2, 8:43 PM
---
**User**:
Script output:
{
"updatedRepos": [
{
"name": "anthropics/cwc-workshops",
"updated": "2026-06-02T16:58:32Z",
"desc": null,
"url": "https://github.com/anthropics/cwc-workshops"
},
{
"name": "anthropics/claude-for-legal",
"updated": "2026-06-02T16:57:59Z",
"desc": "A suite of plugins for legal workflows",
"url": "https://github.com/anthropics/claude-for-legal"
},
{
"name": "anthropics/skills",
"updated": "2026-06-02T16:57:37Z",
"desc": "Public repository for Agent Skills",
"url": "https://github.com/anthropics/skills"
},
{
"name": "anthropics/financial-services",
"updated": "2026-06-02T16:55:54Z",
"desc": null,
"url": "https://github.com/anthropics/financial-services"
},
{
"name": "anthropics/knowledge-work-plugins",
"updated": "2026-06-02T16:54:06Z",
"desc": "Open source repository of plugins primarily intended for knowledge workers to use in Claude Cowork",
"url": "https://github.com/anthropics/knowledge-work-plugins"
},
{
"name": "anthropics/defending-code-reference-harness",
"updated": "2026-06-02T16:52:37Z",
"desc": "Skills for threat modeling, scanning, triage, patching, plus an autonomous scanning harness you can /customize",
"url": "https://github.com/anthropics/defending-code-reference-harness"
},
{
"name": "anthropics/claude-code-action",
"updated": "2026-06-02T16:50:29Z",
"desc": null,
"url": "https://github.com/anthropics/claude-code-action"
},
{
"name": "anthropics/claude-plugins-official",
"updated": "2026-06-02T16:50:13Z",
"desc": "Official, Anthropic-managed directory of high quality Claude Code Plugins.",
"url": "https://github.com/anthropics/claude-plugins-official"
},
{
"name": "anthropics/h...
**Nano**: Let me check the existing wiki structure and fetch the most relevant repo READMEs in parallel.
**Nano**: Good, many repos already have wiki pages. Let me check the existing pages for the key ones and fetch a few more READMEs to check for updates.
**Nano**: Now let me fetch READMEs for repos that may be newer or have less coverage.
**Nano**: Two repos don't have wiki pages yet: `agent-sdk-workshop` and `riv2025-long-horizon-coding-agent-demo`. Let me create those and update the index and log.
**Nano**: Now update the wiki index and log.
**Nano**: Done. Sending the summary to James.
Sweep #86 done — 30 repos updated, 2 new ones worth adding:
**[[Agent SDK Workshop]]** (`anthropics/agent-sdk-workshop`) — Workshop curriculum for the Claude Agent SDK. Python-based, no code writing required. 4-stage guided demo: bare chat → tools → subagents → memory (each stage adds one SDK primitive). Then 6 breakout scenarios assembled from 19 pre-built tools and 6 sub-agents. Good on-ramp if you want to see how the Agent SDK primitives stack up progressively.
**[[Long-Horizon Coding Agent Demo]]** (`anthropics/riv2025-long-horizon-coding-agent-demo`) — Archived AWS re:Invent 2025 demo. An autonomous agent that reads GitHub issues and builds full-stack apps: GitHub issue → Bedrock AgentCore (hosted container) → Claude Agent SDK build loop → CloudFront preview → human approval gate before promotion. The GitHub Actions + approval-gate pattern is directly relevant to qip-qien's ticket → expert review flow.
Everything else in the 30-repo sweep was already in the wiki (cwc-workshops, claude-for-legal, knowledge-work-plugins, defending-code-reference-harness, claude-agent-sdk-demos, etc.) or skipped (infra forks: terragrunt, argo-cd, connect-rust, orjson, etc.).
**User**:
Script output:
{
"updatedRepos": [
{
"name": "anthropics/skills",
"updated": "2026-06-02T19:00:31Z",
"desc": "Public repository for Agent Skills",
"url": "https://github.com/anthropics/skills"
},
{
"name": "anthropics/claude-agent-sdk-python",
"updated": "2026-06-02T18:57:30Z",
"desc": null,
"url": "https://github.com/anthropics/claude-agent-sdk-python"
},
{
"name": "anthropics/claude-cookbooks",
"updated": "2026-06-02T18:56:43Z",
"desc": "A collection of notebooks/recipes showcasing some fun and effective ways of using Claude.",
"url": "https://github.com/anthropics/claude-cookbooks"
},
{
"name": "anthropics/knowledge-work-plugins",
"updated": "2026-06-02T18:56:05Z",
"desc": "Open source repository of plugins primarily intended for knowledge workers to use in Claude Cowork",
"url": "https://github.com/anthropics/knowledge-work-plugins"
},
{
"name": "anthropics/claude-code",
"updated": "2026-06-02T18:53:25Z",
"desc": "Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.",
"url": "https://github.com/anthropics/claude-code"
},
{
"name": "anthropics/claude-code-base-action",
"updated": "2026-06-02T18:52:43Z",
"desc": "This repo is a mirror of the contents of base-action in https://github.com/anthropics/claude-code-action.",
"url": "https://github.com/anthropics/claude-code-base-action"
},
{
"name": "anthropics/claude-code-action",
"updated": "2026-06-02T18:52:29Z",
"desc": null,
"url": "https://github.com/anthropics/claude-code-action"
},
{
"name": "anthrop...
**Nano**: All 18 repos in this sweep are already covered — another CI/plugin-bump wave. Let me log it and update James.
**Nano**:
Checked GitHub — nothing new worth ingesting. Sweep #87 (17:19–19:00 UTC): 18 repos, all CI/plugin-bump automation. Everything already in the wiki.
**User**:
Script output:
{
"updatedRepos": [
{
"name": "anthropics/knowledge-work-plugins",
"updated": "2026-06-02T21:00:03Z",
"desc": "Open source repository of plugins primarily intended for knowledge workers to use in Claude Cowork",
"url": "https://github.com/anthropics/knowledge-work-plugins"
},
{
"name": "anthropics/financial-services",
"updated": "2026-06-02T20:58:40Z",
"desc": null,
"url": "https://github.com/anthropics/financial-services"
},
{
"name": "anthropics/claude-cookbooks",
"updated": "2026-06-02T20:58:27Z",
"desc": "A collection of notebooks/recipes showcasing some fun and effective ways of using Claude.",
"url": "https://github.com/anthropics/claude-cookbooks"
},
{
"name": "anthropics/claude-quickstarts",
"updated": "2026-06-02T20:54:10Z",
"desc": "A collection of projects designed to help developers quickly get started with building deployable applications using the Claude API",
"url": "https://github.com/anthropics/claude-quickstarts"
},
{
"name": "anthropics/claude-code",
"updated": "2026-06-02T20:53:03Z",
"desc": "Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.",
"url": "https://github.com/anthropics/claude-code"
},
{
"name": "anthropics/claude-plugins-official",
"updated": "2026-06-02T20:51:58Z",
"desc": "Official, Anthropic-managed directory of high quality Claude Code Plugins.",
"url": "https://github.com/anthropics/claude-plugins-official"
},
{
"name": "anthropics/skills",
"updated": "2026-06-02T20:51:58Z",
"desc": "Public repository for Agent Skills",...
**Nano**: Another CI wave — all 19 repos already covered.
**Nano**:
Checked GitHub — nothing new. Sweep #88 (19:03–21:00 UTC): 19 repos, all CI automation. Everything already in the wiki.
**User**:
Script output:
{
"videosToIngest": [
{
"videoId": "Dcvl4mseZ6c",
"title": "Launches are still terrifying for me…what if zero people care?",
"channel": "Behind the Craft (Peter Yang)",
"publishedAt": "2026-06-02T15:00:17Z",
"url": "https://www.youtube.com/watch?v=Dcvl4mseZ6c"
},
{
"videoId": "bjrmYsFae_g",
"title": "The Problem Solvers | Max Junestrand at Legora",
"channel": "Claude (Code with Claude workshops)",
"publishedAt": "2026-06-02T14:03:39Z",
"url": "https://www.youtube.com/watch?v=bjrmYsFae_g"
},
{
"videoId": "rJ3tkHW-0pY",
"title": "AI Race: Anthropic's New Opus 4.8 Strategy #shorts",
"channel": "Authority Hacker Podcast",
"publishedAt": "2026-06-02T10:57:27Z",
"url": "https://www.youtube.com/watch?v=rJ3tkHW-0pY"
},
{
"videoId": "OlPnb9HhQKc",
"title": "The skill that makes AI catch the bugs it missed",
"channel": "Behind the Craft (Peter Yang)",
"publishedAt": "2026-06-01T14:51:05Z",
"url": "https://www.youtube.com/watch?v=OlPnb9HhQKc"
},
{
"videoId": "2QOKsTcaQmA",
"title": "ChatGPT 5.5 vs. Opus 4.8: Which Chatbot Researches Better? #shorts",
"channel": "Authority Hacker Podcast",
"publishedAt": "2026-06-01T15:57:46Z",
"url": "https://www.youtube.com/watch?v=2QOKsTcaQmA"
},
{
"videoId": "PllVotreJNE",
"title": "Opus 4.8: Honest AI That Admits When It Doesn't Know #shorts",
"channel": "Authority Hacker Podcast",
"publishedAt": "2026-06-01T10:48:31Z",
"url": "https://www.youtube.com/watch?v=PllVotreJNE"
},
{
"videoId": "GdxLaeyu33c",
"title": "The Exact AI Skills This Solo Founder Uses to Build 5 Apps at Once | Josh Pigford",
"channel": "Behind the Craft (Peter Yang)",
"publishedAt": "2026-05-31T13:...
**Nano**: Let me check how the monitor script accesses YouTube, since it clearly can reach the API.
**Nano**: The monitor script uses the timedtext API too and it worked. Let me retry with the same approach the script uses, plus the YouTube API key for the caption data.
**Nano**: Captions exist — the timedtext rate limit may have cleared. Let me retry and also try via the OneCLI proxy.
**Nano**: The timedtext API is still rate-blocked from this IP. Let me use the browser to get transcripts from the key full-length videos.
**User**: Base directory for this skill: /home/node/.claude/skills/agent-browser
# Browser Automation with agent-browser
## Quick start
```bash
agent-browser open # Navigate to page
agent-browser snapshot -i # Get interactive elements with refs
agent-browser click @e1 # Click element by ref
agent-browser fill @e2 "text" # Fill input by ref
agent-browser close # Close browser
```
## Core workflow
1. Navigate: `agent-browser open `
2. Snapshot: `agent-browser snapshot -i` (returns elements with refs like `@e1`, `@e2`)
3. Interact using refs from the snapshot
4. Re-snapshot after navigation or significant DOM changes
## Commands
### Navigation
```bash
agent-browser open # Navigate to URL
agent-browser back # Go back
agent-browser forward # Go forward
agent-browser reload # Reload page
agent-browser close # Close browser
```
### Snapshot (page analysis)
```bash
agent-browser snapshot # Full accessibility tree
agent-browser snapshot -i # Interactive elements only (recommended)
agent-browser snapshot -c # Compact output
agent-browser snapshot -d 3 # Limit depth to 3
agent-browser snapshot -s "#main" # Scope to CSS selector
```
### Interactions (use @refs from snapshot)
```bash
agent-browser click @e1 # Click
agent-browser dblclick @e1 # Double-click
agent-browser fill @e2 "text" # Clear and type
agent-browser type @e2 "text" # Type without clearing
agent-browser press Enter # Press key
agent-browser hover @e1 # Hover
agent-browser check @e1 # Check checkbox
agent-browser uncheck @e1 # Uncheck checkbox
agent-browser select @e1 "value" # Select dropdown option
agent-browser scroll down 500 # Scroll page
agent-browser upload @e1 file.pdf # Upload files
```
### Get information
```bash
agent-browser get text @e1 # Get element text
agent-browser get html @e1 # Get inner...
**Nano**: No transcript available for Josh Pigford — captions not yet generated. Let me check the description and try the other videos, especially the Legora one which had a "standard" caption type.
**Nano**: Good — got the description. Let me now check the Legora video which has a standard (human) caption track, then get descriptions for the other videos.
**Nano**: Transcripts unavailable without auth — the monitor script has a detection bug (rate-limit HTML is >200 chars). Let me get video descriptions via browser, then fix the script and re-queue the videos.
**Nano**: Good — I have everything I need. The shorts gave complete content in their descriptions. Now let me close the browser, write all wiki pages in parallel, fix the monitor script, and handle the GitHub sweep.
**Nano**: Good — 2:44 promotional video, full story in the description. Now let me write all wiki pages in parallel and fix the monitor script.
**Nano**: Now update the wiki index, log, fix the monitor script, and re-queue the full episodes.
**Nano**: Now fix the monitor script and update the state to re-queue full-length videos, then update the log.
**Nano**: Now re-queue the full-length videos (31:32, 41:10 episodes) that don't have transcripts yet — the shorts/clips are fine as processed since we got complete content from their descriptions.
**Nano**: Now update the log for both the YouTube ingest and GitHub sweep.
**Nano**:
**YouTube ingest done** — partial. Here's what happened and what was extracted:
**Bug found & fixed:** The YouTube monitor script was fooled by Google's rate-limit error page (HTML, also >200 chars) into thinking transcripts were available. Fixed — now requires the response to actually be JSON. The two full-length episodes (Josh Pigford 31:32, Opus 4.8 41:10) have been re-queued; next monitor run will retry when transcripts are genuinely available.
---
**What got ingested from descriptions + clips:**
🔧 **Josh Pigford's AI Skills Stack** (Behind the Craft, May 31) — Solo founder building 5 apps with agents. His 4-skill stack:
- `/build` — 4-step structured build workflow
- `/adversarial-code-review` — pits Claude Opus against GPT on the same code
- `/but-for-real` — forces AI to self-audit after completing work ("you almost certainly screwed some stuff up, go back over it again" → finds 3-5 bugs, simulates PR code review)
- `/learnings` — accumulates lessons over time to improve output
New concept page: **[[But For Real Skill]]** — practical, low-overhead self-audit pattern. No extra tooling needed — just a well-framed follow-up prompt.
🏛️ **Legora** (Code with Claude, June 2) — Max Junestrand's company. "AI-native operating system for legal teams" — autonomous planning, execution, review, and completion of legal work end-to-end. Built on Claude. Their thesis: every Claude model release "raises the tide," their job is to "build boats for everyone else."
⚡ **Opus 4.8** (Authority Hacker shorts, May 29) — 4.7 regression is fixed. Key changes: more honest (admits uncertainty), won't claim completion if implementation is partial, 4x fewer logic errors in code. Authority Hacker ran their real marketing skills (Meta ads, LinkedIn copy, one-shot website) and found it meaningfully better than 4.7.
---
**GitHub sweep #89** (9pm task): 20 repos, all CI automation. Nothing new.
**User**:
I’ve seen mentions of notion, linear, tutti, and other such tools. I’m thinking that there is value in being able to trigger the codex and Claude code from a kanban board. Which of these do you think we should use and the very next thing is what are we putting in it and how do we prioritize? I’ve got various repos and projects and internal infrastructure tasks, trying to figure out a system to keep my sanity, get stuff done, prioritize well, utilize the 5h and 7d usage windows well. Need help and advice.
https://podcasts.apple.com/us/podcast/authority-hacker-podcast-ai-automation-for-small-biz/id1073349789?i=1000757265570
**Nano**:
Two parts: (1) kanban tool recommendation + system design, (2) queue the Authority Hacker podcast episode.
On "Tutti" — I don't recognize this name. Could be a newer tool, a misspelling, or something I haven't ingested. I should be honest about that.
Key factors:
- James is solo, GitHub-native, wants to trigger Codex/Claude Code from the board
- Has repos already + non-repo infrastructure tasks
- Wants to optimize for 5h/7d usage windows
My recommendation: **Linear** over GitHub Projects and Notion for this specific use case. Notion is too freeform for triggering agents; GitHub Projects is fine but worse UX for mixed (repo + infra) work. Linear has clean API, good prioritization, and integrates with GitHub.
For the podcast link — queue it, don't immediately ingest.
On the podcast link — queued for overnight ingest, not processing it now.
**Nano**: Let me save the podcast to the ingest queue, then give James my full answer.
**Nano**:
Podcast queued for overnight ingest.
On the kanban question — here's my honest take:
---
**Tool: Linear**
Not Notion (too freeform, becomes a mess without discipline, not built for agent triggering), not Tutti (not familiar with it — can you share a link?), not GitHub Projects (fine but worse UX for mixing repo work with infra tasks, and weaker prioritization).
Linear wins because:
- Clean priority system that actually makes you decide P0/P1/P2/P3
- GitHub sync is native — issue ↔ PR ↔ commit linkage works out of the box
- Webhook → trigger Claude Code Action on label change (e.g. label "claude-run" → fires the agent)
- API is simple and well-documented for automation
- Handles mixed work (code + infra + planning) in one place without needing a database schema
The "trigger from kanban" flow with Linear specifically looks like: create issue → add label → GitHub Action fires `claude-code-action` against the linked repo. That's it.
---
**What to put in it**
One board per *context*, not per repo. Your repos naturally cluster:
| Board | What goes in it |
|---|---|
| **qip-qien** | Product features, agent pipeline improvements, Ramón collaboration items |
| **Autonomous agents** | congressional-trading, pool-monitor — improvements, monitoring fixes, alert tuning |
| **iOS apps** | DashSmash, Speaker Listens — features, bug fixes |
| **Platform/Nano** | This agent's infrastructure, MCP servers, scheduled tasks, second brain |
| **Arduino/home** | Home automation — low urgency, occasional |
Add a **"claude-runnable"** tag for tasks you can hand entirely to an agent vs. tasks needing your direct involvement. This becomes your queue management signal.
---
**How to use the 5h/7d windows**
The 5h rolling window is your **session budget** — one focused agent task per session. The 7d window is your **sprint budget** — plan the week's agent work to stay under it.
Practical system:
- **Monday:** tag this week's P1 items as "In Window"...
**User**:
Tutti https://github.com/nutthouse/tutti
**Nano**:
This changes the whole picture. Tutti is not a kanban board — it's the **orchestration layer that sits on top of** Linear/GitHub Issues/Jira and runs agents against them. These tools aren't competing. They're a stack:
```
Linear or GitHub Issues ← where you prioritize and kanban
↓
Tutti ← claims work, runs agents, manages PRs
↓
Claude Code + Codex ← actual execution
```
What Tutti does specifically:
- Defines your "agent team" in `tutti.toml` (roles: implementer=claude-code, reviewer=codex, etc.)
- Spawns multiple agents in separate git worktrees — no merge conflicts, true parallelism
- Orchestrates the full SDLC loop: intake issue → implement → test → open PR → collect review → enforce merge gates
- Web dashboard at `:4040` — factory floor view of all agents with live terminal, token usage, git diff
- GitHub issue claim leases — exclusive locks so two agents don't race on the same ticket
- Resilience: auto-recovers from auth failures, rate limits, provider outages
The part most relevant to your window question: Tutti's API-direct mode has **cost tracking and policy gates** built in — you can set budget limits per workflow run. That's the usage window management system you're looking for.
**What I'd actually recommend for you:**
Use **GitHub Issues** as the kanban (not Linear — it's already where your repos are and Tutti has native intake for it). Use **Tutti** as the ops layer.
Day-to-day flow:
1. Open a GitHub issue on whichever repo → tag it with a priority label
2. Tutti claims it, picks the right agent role, runs the workflow in an isolated worktree
3. Agent opens a draft PR when done
4. You review → approve gate → merge
The `tutti.toml` is versioned — you define the pipeline once per repo, fork it for similar ones (qip-qien, congressional-trading share similar patterns).
**One caveat:** Tutti is still early (`cargo install tutti`, requires Rust + tmux). If you want production-...
**User**:
Script output:
{
"updatedRepos": [
{
"name": "anthropics/knowledge-work-plugins",
"updated": "2026-06-03T00:58:56Z",
"desc": "Open source repository of plugins primarily intended for knowledge workers to use in Claude Cowork",
"url": "https://github.com/anthropics/knowledge-work-plugins"
},
{
"name": "anthropics/claude-code",
"updated": "2026-06-03T00:58:45Z",
"desc": "Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.",
"url": "https://github.com/anthropics/claude-code"
},
{
"name": "anthropics/skills",
"updated": "2026-06-03T00:56:49Z",
"desc": "Public repository for Agent Skills",
"url": "https://github.com/anthropics/skills"
},
{
"name": "anthropics/claude-plugins-official",
"updated": "2026-06-03T00:54:48Z",
"desc": "Official, Anthropic-managed directory of high quality Claude Code Plugins.",
"url": "https://github.com/anthropics/claude-plugins-official"
},
{
"name": "anthropics/financial-services",
"updated": "2026-06-03T00:49:17Z",
"desc": null,
"url": "https://github.com/anthropics/financial-services"
},
{
"name": "anthropics/claude-code-action",
"updated": "2026-06-03T00:41:58Z",
"desc": null,
"url": "https://github.com/anthropics/claude-code-action"
},
{
"name": "anthropics/prompt-eng-interactive-tutorial",
"updated": "2026-06-03T00:41:51Z",
"desc": "Anthropic's Interactive Prompt Engineering Tutorial",
"url": "https://github.com/anthropics/prompt-eng-interactive-tutorial"
},
{
"name": "anthropics/claude-cookbooks",
"updated": "2026-06...
**Nano**:
Checked GitHub — nothing new. Sweep #90 (23:33–00:58 UTC): 14 repos, all CI automation.
**User**:
Yes please check traction and reviews on Tutti
**Nano**: The agent found something important — Claude Code itself has docs on orchestrating agent teams. Let me check that before responding.
**Nano**: Good finds on both. Let me add the Claude Code agent teams page to the wiki and then give James the full picture.
**Nano**:
**Tutti verdict: real but invisible**
**GitHub stats:** 37 stars, 3 forks, v0.10.1 (released May 4, 2026 — about a month old). Rapid iteration from v0.2.0 to v0.10.1 in ~6 weeks, so they're actively building. One main maintainer (nutt-adam). "dogfood-now" internal label suggests they use it themselves.
**Community:** Zero. No HN posts, no Reddit threads, no blog coverage, no user testimonials. It exists almost entirely in stealth. Either intentionally low-profile while maturing, or genuinely hasn't broken through yet.
**Assessment:** Not vaporware — the code is real and architecturally solid. But you'd be an early adopter with limited community support, expect breaking changes before 1.0, and you'd be on your own if something breaks.
---
**Bigger find while researching it:** Claude Code has native agent teams built in.
`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`
Team Lead + Teammates model — teammates share a task list, claim work autonomously, communicate directly with each other, support plan approval before implementing. Two display modes: in-process (any terminal) or split-pane (tmux/iTerm2).
**The key difference from Tutti:**
| | Native Agent Teams | Tutti |
|---|---|---|
| Claude Code only | ✓ | Runs Codex/Aider too |
| Worktree isolation | ✗ | ✓ (separate git worktree per agent) |
| Versioned config | ✗ | `tutti.toml` |
| Web dashboard | ✗ | Factory-floor view at :4040 |
| Intake from GitHub/Linear | ✗ | ✓ |
| Cost gates | ✗ | ✓ (API-direct mode) |
| Stability | Experimental | Pre-1.0 |
**My actual recommendation for you:**
Start with native agent teams — it's built into what you already have, Anthropic backs it, no new tooling. Use it on qip-qien first (triager + researcher + drafter as a team makes natural sense).
Only move to Tutti if you specifically need: (1) Codex running alongside Claude Code, (2) the factory-floor dashboard, or (3) versioned `tutti.toml` as auditable ops config. Those are real reasons — but th...