# Claude In-App Browser Workflows — Riley Brown Video **Source:** [Codex and Claude Shipped Browser Updates. This Changes Everything.](https://www.youtube.com/watch?v=juPDqb89dew) — Riley Brown Riley demos the new in-app browser in Claude Code / Claude Desktop and Codex (multi-tab support, non-localhost navigation, annotation tool). His workflows are Codex-heavy; everything below is translated to my Claude stack (Claude Code CLI, Claude in Chrome, MCPs, skills library). ## 🧠 Core insight > "I've shifted from telling agents what to do to asking them what to do and pulling the best thread." — Jack Dorsey (quoted in video) The pattern: agent evaluates all connected context → recommends the task → then **stages the entire workspace** (tabs, docs, references) so work starts instantly. The staging step is the piece my current setup lacks. ## 🖥️ Ideas mapped to my workflows ### 1. Multi-tab briefing sessions - Use Claude in Chrome's MCP tab group (`tabs_create_mcp`, `browser_batch`) to batch-open a working set: Inside Airo AI Builder YouTube Studio analytics, WordPress.org plugin dashboard (Outpost & Co suite), Obsidian daily note (via URI), open Trac tickets. - Chrome extension batching avoids the one-tab-at-a-time lag Riley complained about. ### 2. "Ask, don't tell" morning triage → workspace staging - Chief-of-staff plugin (daily-note + week-state-machine + proactive-intelligence + things-integration) + Gmail/Calendar MCPs already does the *triage* half. - **To build:** extend `/chief-of-staff:daily-note` with a "stage my workspace" step — after Big 3 selection, fire Chrome tab-opens for every doc/link/context item the chosen task needs. ### 3. Email draft skill with tab review - Gmail MCP `create_draft` → Claude in Chrome opens each draft URL in its own tab → I review and send manually. - Wrap as a Claude Code skill; pair with the humanizing-AI-artifacts skill so drafts sound like me. Nothing auto-sends. ### 4. B-roll tab staging per script line - Feed a video script → Claude identifies 15+ visual moments → opens a Google Images or vidIQ `generate_broll` tab per moment **in script order**. - vidIQ MCP goes further than Riley's setup: stock clips, thumbnail scoring, outlier research in the same session. - Fits directly into airo-content-engine and airo-for-wp-content-engine video production. ### 5. Pre-filled social compose tabs - Claude writes 3 caption variants → opens each in a pre-filled X compose tab (`x.com/intent/post?text=...`) for @godaddypro → I pick and send manually. - Human-in-the-loop preserved for brand accounts. ### 6. Parallel research via sub-agents - Three simultaneous Claude Code sub-agent sessions, e.g.: (a) outlier hooks via vidIQ, (b) WordPress release tie-in research, (c) competitor/community signals through devrel-engine framing — each dumping links into tabs + a summary doc. ### 7. Bake tab-opening into existing skills ("anytime X happens, open X") - **airo-content-engine:** Phase 1 SEO research auto-opens keyword + competitor tabs. - **church-bulletin-to-obsidian:** open the finished note + Bible Gateway passages. - **PR review workflow:** open PR, diff, and related issues as tabs before rendering verdict. ### 8. Annotation tool for doc review - Claude's browser annotation (Excalidraw-style): select a region on any live page, comment, lands in chat. - Use for: plugin directory asset reviews, blog QA passes (Yoast readability on the rendered page instead of describing issues in prose). ## ⚠️ Caveats - Browser update was days old at recording — test which surface (Desktop chat vs. Claude Code CLI) supports non-localhost tabs on my build. - Tab-heavy skills belong in Claude Code CLI config (14+ servers), **not** Desktop chat (5-MCP lean rule). ## ✅ Implementation order (biggest daily payoff first) 1. "Stage my workspace" extension to daily-note skill 2. Email draft + tab review skill 3. Tab-opening steps in airo-content-engine (SEO phase + B-roll staging) 4. Parallel research sub-agent pattern 5. Social compose tabs, annotation review habit --- ## 🚧 Implementation status (2026-07-18, via Claude chat + Desktop Commander) Implemented on the **old laptop** (`/Users/crobertson`) — Desktop Commander was connected there: - ✅ `chief-of-staff/skills/daily-note/SKILL.md` — new "Stage My Workspace" section: batched tab staging after a confirmed triage pick, 8-tab cap, link-list fallback for Desktop Chat - ✅ `~/.claude/skills/email-draft-review/SKILL.md` — **new skill**: Gmail triage → drafts in my voice → drafts opened in tabs; hard never-send rule - ✅ `~/.claude/skills/airo-content-engine/SKILL.md` — "Browser tab staging" section: SEO-phase evidence tabs + script-order B-roll staging with manifest - ✅ `~/.claude/skills/church-bulletin-to-obsidian/SKILL.md` — optional finish: open sermon note + BibleGateway passage tabs - ✅ `~/.claude/skills/pr/SKILL.md` — review tab staging: PR + Files Changed + linked issues (+ CI if failing) - ✅ `~/.claude/skills/social-content/SKILL.md` — pre-filled X intent compose tabs + LinkedIn manifest pattern, brand-account guardrails ### ⏭️ Still to do - [ ] Mirror all six edits to the **new laptop** (`/Users/courtneyrobertson`) — skills there weren't reachable this session - [ ] `devrel-engine` — parallel sub-agent research pattern (skill lives on new laptop only) - [ ] `airo-for-wp-content-engine` — same SEO/B-roll staging sections as the Airo engine (new laptop only) - [ ] Test which surface supports non-localhost in-app browser tabs on current builds (Claude Code CLI vs Desktop) - [ ] Annotation-tool review habit for plugin directory assets + blog QA (no skill change needed — just try it) - [ ] Next up: Fedica integration exploration ## Links - [[AI Agent Safety MOC]] — home MOC; every workflow here (draft-but-never-send, pre-fill-but-human-picks, tab staging for review) is an agent computer-use pattern governed by this MOC's control-layer concerns - [[Bounded Action Spaces Not Better Models Make Agents Safe to Run]] — applies-in; the recurring 'nothing auto-sends / human-in-the-loop' rule for email drafts, social compose tabs, and brand accounts is a deliberately bounded action space — the agent stages, the human commits - [[An Ops Agent Is Only as Capable as Its System Access]] — extends; the workspace-staging pattern only works because Chrome, Gmail, and vidIQ MCPs give the agent real read/write access to the systems where the work lives — capability follows access, not persona