Closing the Sonnet-to-Frontier Gap in Claude Code (research report, 2026-07-04)

Priority order: effort xhigh/max + ultracode (1) > verification loops/hooks (3) > single-turn spec (2) > review subagent (4) > plan mode (5) > explicitness (6) > context hygiene (7) > parallel review (8).

1. Raise effort — the single biggest lever

Sonnet 5 defaults to high; Anthropic recommends xhigh for the hardest coding/agentic work, max removes constraints. Effort shifts roughly a model-generation in intelligence (Sonnet 5 @ medium ~= Sonnet 4.6 @ high). In Claude Code: /effort xhigh (persists), /effort max (session), "effortLevel": "xhigh" in settings.json, or /effort ultracode (xhigh + dynamic workflow orchestration). One hard turn: include ultrathink in the prompt. If reasoning is shallow, raise effort rather than prompting around it. (platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-sonnet-5, code.claude.com/docs/en/model-config)

2. Well-specified single-turn specs, not drip-fed multi-turn

Anthropic: underspecified prompts conveyed progressively over turns reduce token efficiency and sometimes performance. Front-load task, intent, constraints, out-of-scope, verification step. Best pattern: "Interview me in detail using AskUserQuestion about implementation, edge cases, tradeoffs. Then write a complete spec to SPEC.md" — execute in a fresh session. (code.claude.com/docs/en/best-practices)

3. Give Sonnet a check it can run

Frontier models self-verify more; Sonnet needs the loop built in. Escalating gates: (a) "run the tests and iterate until they pass, show the output"; (b) /goal re-checked every turn; (c) Stop hook that blocks turn-end until a script passes (deterministic); (d) fresh-context verifier subagent. Always demand evidence, not assertions.

4. Fresh-context adversarial review subagent

"Use a subagent to review the diff against PLAN.md. Check every requirement is implemented, listed edge cases have tests, and nothing outside scope changed. Report gaps, not style preferences." Define .claude/agents/reviewer.md with a stronger model / higher effort — spend deep-reasoning budget on review. For code review, Sonnet over-filters: "Report every issue including uncertain/low-severity ones; goal is coverage; include confidence and severity for downstream filtering."

5. Plan mode before touching code

Explore → plan (plan mode) → implement against plan → verify. Matters more for Sonnet, which commits earlier. Skip only for one-sentence diffs.

6. Be literal and explicit

Sonnet-tier follows instructions literally, does not silently generalize, and needs "above and beyond" behavior explicitly requested. State scope ("apply to every section, not just the first"), point at exemplar files, paste actual errors, explicitly request edge cases/validation/tests.

7. Aggressive context hygiene

Performance degrades as context fills — hits smaller models hardest. /clear between tasks; after two failed corrections, /clear and rewrite the prompt with what you learned; delegate exploration to subagents so file dumps never enter main context; keep CLAUDE.md ruthlessly short.

8. Writer/Reviewer parallel sessions

One session implements, a second fresh-context session reviews (edge cases, race conditions, pattern consistency), findings fed back. Two Sonnets in this pattern often beat one unreviewed pass.