Claude Code’s Next Era — Thariq Shihipar, Anthropic

What it claims

This is a practitioner interview, not a release note. Its usable content is a set of concrete working practices and a mechanism-level description of Claude Mods that the summary only names:

• Effort settings scale with task type. Thariq ran the roughly 70 Terminal Bench 3 problems across effort levels and read the transcripts. High versus low effort "changes the evals a lot" for security. It changes software engineering much less, because extra effort mostly goes into verification and edge-case testing.
• The rough split: code review and security at high or max, UI work at low or medium, and API work higher, so edge cases get covered. No per-level numbers are given.
• Most failures at high and max effort are not knowledge gaps. In the eval transcripts, the model usually considered the correct solution and then rejected it. Asking for implementation or decision notes exposes those rejected options so you can redirect.
• Frontier models may soon dominate smaller ones on tokens too. A smarter model skips verification work (for example, spinning up Chromium for screenshots), so it can do simple tasks in fewer tokens. Thariq says this is "not quite true yet."

How Claude Mods work

• Origin. Mods were internally called "function hooks." They run inside Claude Code's TypeScript runtime instead of as external scripts triggered by events.
• What a mod can do:
• read turn count, tokens used and messages in scope
• register on many more events than hooks support
• spawn subagents and parse their structured output
• modify the UI, which hooks cannot do
• Where they run. Mods work in the CLI and desktop. Claude Tag support is "maybe."
• Who built it. It was a collaboration between someone on the Bun team and someone on the Claude Code team.
• Forked agents. These reuse the parent's prompt cache, so a small side request is cheap. The result does not stay in the main context, which lets a mod act as a supervisor.
• Composition. Mods can hook into each other. For example, a mode-selector mod lets other mods register as modes (auto-router, artifact mode, plan mode).

Worked example: a quiz-after-task mod.

flowchart TD
  A[End of assistant turn] --> B[Forked agent: task completed?]
  B -->|false| C[No action]
  B -->|true| D[Generate quiz as JSON]
  D --> E[Parse in mod]
  E --> F[Render above prompt input]

The quiz mod runs after every turn, so it is "slightly token-intensive." Other mods mentioned:

• a register-assumption tool that collects assumptions and shows them at the end
• a model router, not shipped by default because routing is "a hard problem"
• a next-steps supervisor
• a dashboard mod that maintains an artifact

Artifacts now have an attached database and an artifact MCP. Multiple Claude instances can read and write shared state through it, for example a kanban board, and the artifact can talk back to those instances.

How to use it

• Spend effort on the first prompt. Iterating with "undo and redo this" is what burns rate limits. State whether the work is a prototype or production and how much compute it merits. Information density matters more than format, so rambling voice prompts are fine.
• Start new projects without a Claude.md. Add only failure modes that keep recurring, and re-check them when the model changes, since Fable 5 and Fable 5.1 differ. Stale failure-mode lists over-constrain the model.
• Use the new tooling. Evals plugins for skills were just added, and Agents.md support is coming.
• Try /eli5, installable as a plugin. Its key instruction is "big pictures, few words." It came from untangling complex internal incidents and counters text-heavy artifacts.
• Treat prompting like executive communication. Swyx's suggestion is the SCQA memo structure: situation, complication, question, answer.

Limits

• Truncated source. Only the first ~60k of ~113k characters were fetched, ending around the 45-minute mark. The security half is not covered here: Exploit-Bench, the Hugging Face scorer incident, Pacing the Frontier, probes and Auto Mode.
• Unpublished numbers. The effort findings come from a blog post that is still in progress. The interview gives no figures.
• Unshipped features. Several mods are personal works in progress. Local hands and Projects multiplayer are described as future work.