AI Bookmarks Dashboard
← Back to themes
Claude Code / Skills

Claude Code / Skills

19 videos

Claude Code features, custom skills, slash commands, CLAUDE.md conventions.

  1. 3 Claude Code Features Most People Miss: Sub-Agents, /status, and Auto-Skill Hooks
    1. Sub-agents: Type 'spin up a sub agent to research [topic]' — Claude creates a separate instance, does the work, returns only the summary.
    2. Context monitoring: Type /status for color-coded context meter; /context for breakdown by component.
    3. Auto-skill hooks: Write a hook script that reads your prompt and loads the matching skill automatically. Follow @qendresahhoti for setup details.

    When starting a research task within an existing session: 'Spin up a sub agent to research [topic] and return just the key findings as a summary. Keep the full research in the sub-agent context, not here.'

    View full video →
  2. 4 MCP Tools That Make Claude Code 10x Smarter (Context7, Sequential Thinking, Superpowers, Plan Mode)
    1. Four Claude Code add-ons. Research each before installing — do NOT install until each has been reviewed and approved:
    2. Context7 MCP — feeds current library docs on demand. It's in the official Claude Code plugin marketplace (/plugin -> search 'Context7'); still confirm what it is and that it's the official one before installing.
    3. Sequential Thinking MCP — adds a reasoning/pause layer. Find its GitHub repo and research who publishes it and whether it's safe.
    4. Superpowers MCP — forces Claude to back up work before claiming done. Same research-first flow via its GitHub repo.
    5. Plan Mode — built-in, nothing to install (Shift+Tab until 'plan mode' shows).
    6. For the three MCPs, present the findings for review and install ONLY after explicit approval.
    View full video →
  3. "Lock It In" — Claude Skill That Properly Captures and Packages Your Workflows
    1. Build a workflow in Claude until you love how it turned out.
    2. Type "Lock it in" to trigger the skill.
    3. Answer 2-3 clarification questions Claude asks.
    4. Receive a properly-formatted skill with examples and edge cases, pre-tested.
    View full video →
  4. CLAUDE.md + handoff.md + Start/Stop Hooks: Claude Code Always Knows What to Work On Next
    1. Run /init in terminal to create CLAUDE.md.
    2. Prompt: 'Create a handoff.md file as a session handoff. Include: what was worked on last session, what needs to be worked on immediately next session.'
    3. Create start hook: tells Claude to read handoff.md alongside CLAUDE.md at session start (see @joey_systems for exact terminal command).
    4. Create stop hook: tells Claude to update CLAUDE.md with big changes and update handoff.md with work summary + next steps.
    5. Both hooks run automatically — no manual action needed between sessions.

    Create a handoff.md file. Include: (1) Summary of what was worked on this session, (2) Current state of the codebase (what's working, what's broken), (3) The next 3 specific tasks to work on immediately in the next session, in priority order.

    View full video →
  5. 3 Levels of Claude Code Mastery: Extension → CLI → Multi-Agent Expert
    1. Level progression path:
    2. Switch from VS Code extension to CLI for more control.
    3. Master CLAUDE.md (project context file), Plan Mode, and model/effort selection per task.
    4. Learn worktrees for parallel agent execution across branches.
    5. Explore /remote for builds that run without you at the computer.
    View full video →
  6. 3-Step Token Management: CLAUDE.md + Context Folder + Fresh Chat Per Task
    1. Create CLAUDE.md with your reusable rules and context (Project instructions in Claude.ai, or global CLAUDE.md in Claude Code).
    2. Create a context/ folder with reference docs — reference them by name only when needed rather than loading everything upfront.
    3. Start a fresh chat for each distinct task — don't continue yesterday's long chat for today's quick task.
    View full video →
  7. Claude Code Workflow: Explore → Plan → Code → Commit (Don't Code First)
    1. Enable Plan Mode: Shift+Tab until 'plan mode' appears — Claude reads code and thinks without touching files.
    2. Describe your goal and ask Claude to figure out approach (it will read files, run web searches, ask clarifying questions).
    3. For unfamiliar codebases, type /init — Claude maps the entire codebase before starting.
    4. Only after plan is confirmed, let Claude write code.
    5. Commit when satisfied.

    Before writing any code: enable Plan Mode (Shift+Tab), describe the feature you want, and ask Claude to read the relevant files and propose an approach. Only approve to write code after you've reviewed the plan.

    View full video →
  8. /visual-plan and /visual-recap: MDX Skills for Visual Planning + Code Review with Claude
    1. Identify the target: the free /visual-plan and /visual-recap Claude Code skills by steve8708 (generate MDX visual plans and post-implementation recaps). Do NOT install them yet.
    2. Research first — what the skills do, where the skill files are published, who steve8708 is, and whether the source is safe and reputable (link is in his profile; see Find the Resource).
    3. Present the findings for review, and install ONLY after explicit approval.
    4. Once approved and installed: run /visual-plan with a feature description before coding, review and iterate on the generated wireframes and specs, then let Claude implement the approved plan; run /visual-recap afterward for a visual summary to review or hand off.
    View full video →
  9. Ingest Any YouTube Video into Obsidian + Query It with Claude
    1. Identify exactly what would be installed: the Obsidian Web Clipper browser extension. Do NOT install it yet.
    2. Research it first — what the extension does, who publishes it (Obsidian), what browser permissions/data access it requests, how it's typically used, and whether it's safe and reputable.
    3. Present the findings for review, and install ONLY after explicit approval.
    4. Once approved and installed: on any YouTube video, click the extension to clip it into the Obsidian vault (captures transcript, date, category, people, topics, source) and save to the clippings folder. If the date causes a template error, fix the date format. Then, in Claude Code, grab the note's file path and prompt: 'Looking at this interview [path] — give the key atomic insights,' followed by 'Add these atomic insights to the beginning of the note above the transcript.'

    I have a YouTube video transcript saved in Obsidian at [path]. Read the note and provide the 10 key atomic insights. Then add them to the beginning of the note above the transcript.

    View full video →
  10. Build a Personal Claude Harness in Under an Hour: One Folder, One CLAUDE.md
    1. Create a root folder (e.g. 'my-os').
    2. Add a me.md: plain text description of who you are, what you're working on, how you work.
    3. Pick one domain to start (content, clients, studies) — create a folder for it and drop everything you have into it.
    4. Write a CLAUDE.md that maps what lives where — a few sentences per folder is enough.
    5. Open Claude Code in that directory and start asking questions. Claude will read the map, load the right context, and answer from your actual files.

    Help me create a personal Claude harness: (1) set up a root folder with a me.md describing who I am and how I work, (2) create one domain folder with my existing files, (3) write a CLAUDE.md map of what lives where. Start with the me.md.

    View full video →
  11. Use a Symlink to Sync AI Context Files Across Tools (CLAUDE.md, agents.md)
    1. Create one master file (suggested: agents.md — 60k+ GitHub projects already use this format).
    2. Symlink CLAUDE.md → agents.md: `ln -s agents.md CLAUDE.md`
    3. Symlink any other AI tool config files to the same master.
    4. Update only the master file; all tools auto-read the latest version.
    View full video →
  12. "Last 30 Days" Claude Skill — Restricts Research to Recent Sources Only
    1. Identify the target: the free 'Last 30 Days' Claude skill (filters research on Reddit, X, YouTube, Polymarket, and the web to only the last 30 days). Do NOT install it yet.
    2. Research it first — what the skill does, who publishes it (shared by @itsmariahbrunner), where the skill files actually live, and whether the source is safe and reputable. Use the Find the Resource search string to locate it.
    3. Present the findings for review, and install ONLY after explicit approval.
    4. Once approved and installed: invoke the skill when researching any time-sensitive topic so Claude only draws from sources published in the past 30 days.
    View full video →
  13. Claude Code Artifacts: Auto-Generated Project Summaries to Share with Your Team
    1. Complete your Claude Code coding session as normal.
    2. Claude Code will generate an artifact summarizing all code changes, PRs, and project progress.
    3. Share the artifact with teammates to give them a quick overview of the project without them needing to read the full code history.

    At the end of this session, generate a Claude Code artifact that summarizes: (1) all code changes made, (2) files modified, (3) key decisions, and (4) what's next. Format it for sharing with a teammate who wasn't in this session.

    View full video →
  14. 3 Claude Code Skills for Hard Problems: Watchdog, Planarbiter, and RTFD
    1. Three Claude Code skills from steve8708. Research each before installing — do NOT install until each has been reviewed and approved:
    2. Agent Watchdog — one agent monitors and fixes another's work. Research what it does, where it's published, and whether it's safe; present the findings; install only after approval.
    3. Planarbiter — two agents make independent plans, negotiate differences, and produce a unified plan. Same research-first, approve-then-install flow.
    4. Read The Damn Docs (RTFD) — forces agents to look up real documentation before guessing at APIs. Same flow (see Find the Resource for where to locate these).
    5. Once approved and installed, a useful workflow is: Planarbiter -> Claude Code writes the plan -> Codex executes -> Claude Code watchdogs and fixes anything missed.

    Research the 'Read The Damn Docs' Claude Code skill by steve8708 before installing anything: what it does, where it's published, how it's used, and whether it's safe and reputable. Present your findings and wait for my explicit approval. Only if I approve, install it and confirm agents will now search official documentation before guessing at API solutions.

    View full video →
  15. Claude Code's Hidden Fullscreen Rendering Mode
    1. Open Claude Code in your terminal.
    2. Type the slash command for fullscreen rendering (creator shows it on screen — likely `/terminal` or a rendering toggle; comment 'TERMINAL' on the post for creator's full setup guide).
    3. Once on, you can click anywhere inside a text block and select/delete large portions of text.

    What is the Claude Code slash command to enable fullscreen rendering or rich text mode? Enable it for me and confirm it's working.

    View full video →
  16. The "Ponytail" Claude Skill That Cuts Vibe-Coded Bloat by 90%
    1. Identify exactly what would be installed: the "ponytail" Claude skill published by aimn.dev on GitHub. Do NOT install it yet.
    2. Research it first — what the skill actually does, how it plugs into Claude, how it's typically used, who aimn.dev is, and whether the repo looks safe and reputable. Use the Find the Resource search string to locate the repo, then read the skill's actual rules.
    3. Present those findings for review, and install ONLY after explicit approval.
    4. What it does once installed: forces Claude to pause before writing code and ask — does this need to exist? Can stdlib handle it? Is there a one-liner? Can something be reused? Only if nothing simpler works does it write the minimum code.

    Apply the ponytail minimal-code principle to my current project: scan for functions that duplicate stdlib, over-engineered solutions that should be one-liners, and code that should not exist at all. Give me a prioritized list of what to delete or simplify.

    View full video →
  17. 3 Claude Plugins That Fix the Disconnected AI Problem
    1. Three plugins are recommended. Research each before installing — do NOT install any until each one has been reviewed and approved:
    2. Anthropic Productivity plugin (connects Gmail, Calendar, Slack). Research what it is, who publishes it (Anthropic), exactly what account access it requests, and how it's used; present the findings; install only after approval.
    3. Brand Voice plugin (writes copy in a brand's tone). Identify who publishes it and research what it does and whether it's reputable; present the findings; install only after approval.
    4. Marketing plugin (advises on landing page, SEO, re-engagement). Same: research source and behavior first, present the findings, install only after approval.
    5. Once approved and installed, configure each (e.g. answer the Brand Voice questions so the copy matches the brand).

    Before we install anything, research the Anthropic Productivity plugin for me: what it is, who publishes it, exactly what Gmail/Calendar/Slack access it needs, how it's typically used, and whether it's safe and reputable. Present your findings and wait for my explicit approval — only if I approve, walk me through installing and configuring it.

    View full video →
  18. Reset Claude Code Context with handoff.md Instead of /compact
    1. Before ending a Claude Code session, tell Claude to create a handoff.md file.
    2. Include: goal, current state, files being worked on, what changed, what failed, and what to do next.
    3. Run /clear (or start a new session) and tell Claude to read the handoff doc and continue from there.

    Before we end this session, create a HANDOFF.md file that includes: the overall goal, current state of the work, all files we've touched, what changes were made, what we tried that didn't work, and what to do next. I'll use this to start a fresh session with /clear.

    View full video →
  19. 5 Claude Code Skills That Teach It How Obsidian Actually Works
    1. Identify exactly what would be installed: the 'Obsidian Skills' Claude Code plugin (marketplace: capano/obsidian-skills), reportedly built by Obsidian's CEO Steph Ango — 5 skills: obsidian-markdown, obsidian-bases, json-canvas, obsidian-cli (full vault control from the terminal), and defuddle. Do NOT install it yet.
    2. Research first — confirm who publishes the 'capano/obsidian-skills' marketplace, verify the Steph Ango / Obsidian-CEO claim, and review what each skill does (note obsidian-cli grants terminal control over the vault) and whether it's safe.
    3. Present the findings for review, and install ONLY after explicit approval.
    4. Once approved: inside a Claude Code session run '/plugin marketplace add capano/obsidian-skills' then '/plugin install obsidian-skills'.

    Before installing anything, research the 'Obsidian Skills' Claude Code plugin (marketplace capano/obsidian-skills) for me: who publishes it, whether it's really from Obsidian's CEO, what each of the 5 skills does (especially obsidian-cli, which controls the vault from the terminal), and whether it's safe. Present your findings and wait for my explicit approval. Only if I approve, install it by running '/plugin marketplace add capano/obsidian-skills' then '/plugin install obsidian-skills'.

    View full video →