Video summary
How Anthropic Builds And How Engineering Will Change Soon | Thariq Shihipar
Main summary
Key takeaways
Summary of technological concepts & product/engineering takeaways (Anthropic / Claude Code)
Treat Claude as a “thought partner,” not just an autopilot
- Best practice for onboarding/usage:
- Provide the right context.
- Ask whether Claude can do the task.
- Iterate on system-level steps.
- Key shift: move from “build the product” to “build the system that builds the system” (higher abstraction).
Onboarding changes: less technical ramp-up, more cultural/social
- Claude can cover many “trivial setup” gaps via Q&A.
- Onboarding buddies still matter for:
- Team context
- How work gets approved
- Social/cultural alignment
Why internal experience can differ from external AI perception
- The speaker argues enterprise teams often don’t directly “type code all day” anymore; work is higher-level.
- They emphasize:
- Automation investment has risk/cost.
- Model capability improvements increase the likelihood ROI over time.
Knowledge work is increasingly “reducible to code”
Recommended examples:
- Accounting via Python scripts instead of Excel
- Video editing via FFmpeg and libraries
General idea:
- If tasks can be expressed as code steps, code agents (e.g., Claude Code) can do substantial work.
Harness engineering is load-bearing even as models improve
- The harness (tools, workflows, safety, permissions, sandboxing, verification) becomes more complex as models run more autonomously.
- Example: Auto Mode / Auto Code
- Requires a more sophisticated permission model and safety.
- Runs can last hours, not minutes.
- Artifacts matter because they’re more than output:
- They act as a form of structured prompting that helps Claude represent long-running work.
Autonomy and “fully autonomous changes” depend on what counts
- Different functions vary:
- Designers may contribute assets (e.g., Figma), which Claude Code can leverage via MCP.
- Anthropic’s goal:
- Claude handles more glue work.
- Humans focus on unique daily work and decision-making.
Limits of “one sentence ticket → don’t interrupt until done”
- Works only when the ticket/spec is already well-shaped.
- Real GitHub issues often have:
- Ambiguity
- Multiple concerns
- Autonomy fails when users don’t provide enough structure/unknowns.
New work replacing old “write React line by line”
- Intern-style tasks shift from implementation toward:
- Research and measurement
- New processes such as evals across large user bases
- Trade-offs and verification
Computer/browser use: better, but still edge-case heavy and latency-limited
- Reported issues:
- Stuck flows (e.g., inability to type passwords due to integrations like 1Password)
- Latency framing:
- Computer interaction is harder than code because it resembles a state machine
- Less control and harder reversibility
Loop engineering / workflows
- “Loop engineering” means:
- Instead of prompting each time manually, build a system that repeatedly prompts/monitors (often via Claude Tag and scheduled automations).
- Example workflow concepts:
- Daily triage/feedback
- Morning loops that check fired events and suggest ideas
- Monitoring funnels and improving them
- Trade-off:
- Setup can be expensive.
- Verification + value-focused loops can increase velocity.
Where the industry is headed: loops across the full SDLC
- Claude Code ≈ implementation
- Claude Tag / Claude-as-agent ≈ broader SDLC:
- Feedback
- Code review
- CI/CD incidents
- Incident “babysitting,” etc.
- Expect more generative interfaces using artifacts:
- Interactive reports, plans, diagrams, code snippets
Artifacts explained as interactive web app-like outputs
Claude artifacts can:
- Display plans (diagrams, file snippets, schemas)
- Present reports readable on devices (e.g., phone)
- Call MCPs to fetch data and render views
- Example: inbox reading/tagging
- Core emphasis:
- Move beyond text-in/text-out
- Stay in the loop
Model usage philosophy: planning/implementation split
Rather than “always pick the smartest model,” use models by role:
- Fable (smaller/planning) for:
- Planning/brainstorming/spec discovery (finding unknowns)
- Opus 5 (larger/execution) for:
- Implementation
- Harness/verification reduces the need to obsess over model switching.
Prompting skill is broader than “write a prompt”
What matters is the accumulated setup:
- Skills
- Verification harness
- Added context/data
Example improvements:
- Remove ~80% of system prompt in Claude Code
- Remove examples from tool descriptions (examples were often negative and counterproductive)
How to “nudge” compute usage
Instead of generic advice (“believe in yourself”), explicitly grant permission to:
- Use workflows/agents
- Brainstorm first, then execute
- Run overnight (with goals)
Note:
- Default model behavior tends to be efficient/fast; deeper work requires steering.
Verification and evals for product changes
They use:
- User metrics (e.g., how much people like outputs)
- Internal/external evals across tasks
Hard problem:
- Some undesirable behaviors may not be caught by typical evals
- Example: telling users to “go to sleep”
Style/taste control for subjective domains (e.g., UI/design)
Recommended approach:
- Stay in the loop and supply references in machine-readable formats:
- Figma > screenshots for extracting attributes (e.g., border radius)
- HTML mockups > images for iteration
The hard part:
- Knowing when output is “good enough” requires domain expertise (“becoming more tasteful”).
Code ownership/maintenance under high code generation
Maintenance emphasis shifts:
- Less focus on superficial style/naming
- More focus on:
- Verification harnesses
- Scaffolding
- Vision for what the codebase is good at
Testing recommendations:
- About ~100x more testing code than before
- Use:
- Fixtures/mocks
- Storybooks
- Production-data-derived fixtures
Human element still matters:
- Direction
- Replay/undo in multiplayer
- Decisions about code evolution
Caution on tech debt waiting for “next model”
- Deferring some refactoring/tech debt can be reasonable if the timeline allows it (startup-style: deliver value in months vs. years).
- But it depends on circumstances—deliver value on short cycles.
Reliability with higher velocity
- Increased AI-driven code volume correlates with more incidents/sevs in large orgs.
- Countermeasure theme:
- Build the “dream testing + deployment environment,” including:
- Replay requests
- Chaos testing
- Fixtures and mock databases
- Use Claude to improve uptime.
- Build the “dream testing + deployment environment,” including:
Internal tooling culture
- Many engineers maintain custom scripts/tools/harnesses.
- Claude Tag scheduling examples mentioned:
- Personal scheduling
- Email workflows
- Multi-tenant/display setups for viewing multiple Claude Code sessions
- “Hackable” tooling is encouraged so people create variants.
External visibility and career advice
- Strong recommendation:
- Share work externally (Twitter/X, Reddit, etc.)
- Write about shipped projects
- Treat it as expanding your “luck surface area.”
- Authentic posting/engagement beats “ask an AI to generate everything” shortcuts.
- Still learn fundamentals:
- Computer systems, constraints, memory allocation
- Backend services, etc.
Sponsorship note (mentioned in subtitles)
- Work OS described as an API layer to quickly implement enterprise SaaS requirements such as SSO, SCIM, audit logs, rather than building them from scratch.
Main speakers / sources
- Thariq Shihipar — interviewer/host (asks questions)
- Dark — engineer at Anthropic (cloud code team; primary interviewee)