AMA #1 with Thariq

Ask Claude to read a site or code base and compress its visual language into a single HTML design-system file reusable across projects.

• ·Prompt used: 'look at the code and create a design system HTML file that fully represents the Anthropic design system.'
• ·Output contained colors, typography, radiuses, shadows, buttons, toggles, chips, badges, and feedback states in one file.
• ·The HTML file is then copied into a new project (e.g., a Remotion video repo) so the new build inherits the aesthetic.
• ·Positioned as a hack for non-designers starting new projects.

Append 'interview me in depth about this using the ask-user-a-question tool' to vague prompts so Claude drives spec clarification.

• ·Used in place of plan mode when the user hasn't fully thought through requirements.
• ·Claude surfaces new ideas during interview (e.g., suggesting a mouse component Thariq hadn't considered).
• ·Thariq mentions this could become a /interview skill, though productizing it is tricky since not every user wants to be interviewed.

Run a scheduled loop (e.g., every 20 minutes) that calls a since-timestamp fetch script and processes only new data.

• ·Pair with a tight script like get-messages-since so the loop does nothing when there's no new data.
• ·Example structure: every 20 minutes, check for new messages using get-messages-since, then invoke a process-messages skill.
• ·Warning: can burn tokens if the script isn't tight.
• ·For rapidly-changing, highly interrelated state (e.g., hospital data), a managed agent is suggested instead.

A skill that reads existing git changes and tries to pare down unnecessary code.

• ·Mentioned as Thariq's go-to when Claude writes more code than necessary.
• ·Works against the current diff rather than the whole repo.

Give Claude domain-specific deterministic checks so it can verify its own work beyond generic linting or screenshots.

• ·Generic linting is a starting point — push toward the most specific form of linting for the domain.
• ·Remotion example: check that video frames don't overlap as a programmatic verification.
• ·Playwright is useful but screenshots alone usually aren't enough information for Claude to act on.
• ·For high-risk outputs (e.g., payments), build a testing environment where Claude iterates and writes tests before the real send.

An opinionated custom memory skill that ingests email and iMessage to maintain contacts, to-dos, and events.

• ·Thariq mentions he's been working on open-sourcing it.
• ·Runs on a loop: check for new messages/emails, then update contacts/to-dos/events.
• ·Presented as an example of augmenting Claude's generic memory with opinionated structure via skills.

A skill that reads Slack and builds a weekly status report to send to a manager or post in a personal channel.

• ·One of several business-workflow skills Thariq maintains.
• ·Other examples: a Twitter DM/notification reader that correlates bug reports with Slack, and a demo-video skill that reads new feature code and builds a Remotion demo.

Claude can read its own conversation history stored in .claude/projects to audit workflows and propose skills.

• ·Use it to ask Claude for tips on how you could work better — but prompt very specifically.
• ·Can be asked to propose new skills, though output tends to be noisy with low-value suggestions.
• ·Also usable to count how often each skill triggers when auditing a setup.

A skill can install a hook that injects reminders every turn or every few turns for the remainder of a session.

• ·Example given: a /safe skill that, once invoked, reminds Claude each turn it's doing something dangerous and must double-check.
• ·Referenced in Thariq's skills article as a way to insert reminders dynamically rather than statically in claude.md.

Opus 4.7 takes instructions more literally, so over-triggering usually means claude.md needs relaxed wording.

• ·Caps-locked 'never do this' style rules can cause over-indexing on Opus 4.7.
• ·Anthropic has published an Opus 4.7 prompt guide and a migration skill to point at existing skills.
• ·Audit pattern: ask Claude to read .claude/projects and check how often each skill triggers.

Prefer composable skills (debugging, code quality, feature design) over role-prompted sub-agents like 'front-end agent'.

• ·Role prompts like 'you're an expert designer' are largely obsolete — Claude doesn't need to be told to write good code.
• ·Verbose agent instructions over-constrain Claude, especially on Opus 4.7.
• ·Preferred pattern: main Claude Code spawns sub-agents that assemble skills at the right moment for the task.

Claude Design is an opinionated wrapper around the model with strong design defaults; Claude Code has more code-base context.

• ·Claude Design connects a GitHub repo and ingests a design system similarly to the HTML-porting pattern shown in the demo.
• ·Defaults tend to look better out of the box thanks to Nate and the design team.
• ·Claude Code is preferred when you want to point at a specific code base; Claude Design is harder to steer at arbitrary repos.
• ·Video work in Claude Design was inspired by Thariq showing Nate how Anthropic's launch videos are built.

Anthropic's launch videos are generated by Claude Code using the Remotion library — pairs well with a pre-built design-system HTML.

• ·The live build set up a Remotion project with a custom player, terminal component, cursor, and button showcase.
• ·Jason suggests builders use this approach for hackathon demo videos at the end of the week.
• ·Combining design-system HTML + Remotion is what makes Claude-generated videos match Anthropic's polish.

Michael Cohen is running a dedicated Managed Agents session the morning after, West Coast time.

• ·Bring Managed Agents questions there — including the clinic-style continuous-state use case Thariq deferred.
• ·Session video will be posted in the Discord resources channel for other time zones.