Contact
← Back to blog

Practical Cheatsheet to Maximize Coding with GPT-5

Published: 2025年08月24日

Core Points

  • No ambiguity or conflicting instructions
  • Declare reasoning level (low / medium / high)
  • Segment rules in XML-like blocks
  • Avoid terms like “exhaustively” or “thoroughly”
  • Add a self-evaluation prompt before coding
  • Control tool budget and parallelism

10-Second Quick Summary

  • Zero ambiguity
  • Specify reasoning level
  • Fix the output contract in XML
  • Avoid strong imperatives
  • Self-reflection first
  • Limit tool calls and concurrency

Ready-to-Paste Prompt Template

<code_editing_rules>
  <guiding_principles>
    - Keep changes minimal. List side effects.
    - Prefer modularity and reuse.
    - Prioritize type safety (TypeScript strict).
  </guiding_principles>
  <stack_defaults>
    - Styling: Tailwind CSS
    - Framework: Next.js App Router
    - Tests: Vitest + Testing Library
  </stack_defaults>
  <output_contract>
    - Changed files list
    - Patch diff and summary
    - Test code and run command
  </output_contract>
</code_editing_rules>
<reasoning_effort level="medium"/> <!-- trivial=low, hard=high -->
<self_reflection>
  - Create a 5-item checklist before coding and score it.
  - If not perfect, re-propose the design before implementing.
</self_reflection>
<persistence>
  - Proceed with assumptions, then list “assumptions and alternatives” at the end.
  - Tool budget: up to 6 calls. Parallel search up to 2.
  - Ban unrelated exploration and dependency installs.
</persistence>

Task-Specific Mini Snippets

Bug Fix (small)

<task intent="fix-bug" severity="minor">
  - Write repro steps first.
  - List 3 cause hypotheses → test the most likely one.
  - Change at function scope. Keep public API stable.
</task>
<reasoning_effort level="low"/>

Refactor (medium)

<task intent="refactor" scope="module">
  - Remove cycles. Raise cohesion, lower coupling, report numbers.
  - Preserve existing tests. Explain any snapshot updates.
</task>
<reasoning_effort level="medium"/>

Greenfield (large)

<task intent="greenfield">
  - List use cases, boundaries, and error-time UX.
  - Fix API contract first. Show I/O types.
</task>
<reasoning_effort level="high"/>

Anti-Patterns → Improvements

  • Bad: “Gather all info then answer,” “exhaustively”
    Better: “Max 4 tool calls. Try in priority order. Summarize mid-results before coding.”
  • Bad: “Use the optimal approach.”
    Better: “Prefer latency <300 ms, memory <200 MB, ≤3 files changed.”

IDE/Agent Integration Rules

  • Store the XML blocks in .cursor/rules or AGENTS.md
  • Write repo-specific prohibitions (no new deps, no file deletions, no public API breaks)
  • Set default tool budget and parallelism
  • Add self-check to the PR template

Evaluation Checklist

  • Instruction unambiguity: removal rate of vague terms
  • Reasoning-level fit vs task difficulty
  • Minimal-change score: touched lines/files
  • Regression safety: added tests and coverage delta
  • Observability: consistent logging and error handling
  • Reproducibility: clear setup and run steps

Example Uses

  • Draft App Router migration → self-check → PR with diffs
  • Gradual refactor of weakly typed utilities
  • Add a cache layer while preserving API compatibility

Recommended Thinking Process

  1. Classify task difficulty → set reasoning level
  2. Freeze the output contract first (diffs, tests, steps)
  3. Constrain exploration via tool budget and parallelism
  4. Self-check before coding. Ensure rework criteria are met
  5. Implement with minimal changes → regression tests → record assumptions and alternatives