---
url: (synthesis — every claim traced to a vendor URL inline)
fetched: 2026-08-12
---

# The generation shift: prompt + context engineering, 2023 instruct models vs 2026 frontier reasoning models

Every row cites a live vendor page fetched 2026-08-12.

## The shift in one table

| Dimension | 2023-era instruct doctrine | Current frontier-reasoning doctrine | Vendor source |
|---|---|---|---|
| Chain-of-thought | "Think step by step" was THE technique; OpenAI GPT-4.1 guide still prescribes induced CoT ("thinking out loud") for +4% agentic pass rate | "Since these models perform reasoning internally, prompting them to 'think step by step' or 'explain your reasoning' is unnecessary." (OpenAI). Anthropic demotes manual CoT to "a fallback" for when thinking is off; "A prompt like 'think thoroughly' often produces better reasoning than a hand-written step-by-step plan." | developers.openai.com/api/docs/guides/reasoning-best-practices; developers.openai.com/cookbook/examples/gpt4-1_prompting_guide; platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices |
| Few-shot examples | Few-shot walls as default steering; Anthropic docs still say "Include 3–5 examples for best results" for general use | OpenAI reasoning models: "Try zero shot first, then few shot if needed." Anthropic Claude Code team: "giving examples actually constrains them to a certain exploration space"; "removing examples was extremely helpful, because it was just more creative than the examples we gave it." (Thariq, fireside chat) | developers.openai.com/api/docs/guides/reasoning-best-practices; claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models; simonwillison.net/2026/Jul/21/cat-and-thariq/ |
| "You MUST" instruction stacks | GPT-4.1 agentic reminders: "You MUST plan extensively before each function call..." (+20% SWE-bench); anti-laziness and tool-forcing boilerplate everywhere | Anthropic: "Where you might have said 'CRITICAL: You MUST use this tool when...', you can use more normal prompting like 'Use this tool when...'." and "Instructions like 'If in doubt, use [tool]' will cause overtriggering." | developers.openai.com/cookbook/examples/gpt4-1_prompting_guide; platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices |
| Rigid rule lists | Enumerate every behavior and edge case; prohibition lists | "Rules become judgment": replace "never write multi-paragraph docstrings" with "Write code that reads like the surrounding code." Fable 5: "you can steer most behaviors with a brief instruction rather than enumerating each behavior by name." | claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models; platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5 |
| Persistence scaffolding | "You are an agent - please keep going until the user's query is completely resolved..." required verbatim (GPT-4.1); still recommended through GPT-5.1 | Anthropic migration guide: "If you've added scaffolding to force interim status messages ('After every 3 tool calls, summarize progress'), try removing it." Opus 5 needs damping, not spurring | developers.openai.com/cookbook/examples/gpt4-1_prompting_guide; platform.claude.com/docs/en/about-claude/models/migration-guide |
| Verification / self-check prompts | "Before you finish, verify your answer against [test criteria]" — recommended for coding/math | "Claude Opus 5 verifies its own work without being told to... remove them: instructions like these cause over-verification... removing them reduces wasted tokens with no loss in quality." | platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5 |
| Prefill tricks | Prefill assistant turn to force JSON, skip preambles, dodge refusals | Hard-deprecated: "Requests with prefilled assistant messages to these models return a 400 error. Model intelligence and instruction following have advanced such that most use cases of prefill no longer require it." | platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices |
| Thinking budgets | Manual extended thinking with `budget_tokens` (2025) | Adaptive thinking: "Claude Fable 5 and Claude Mythos 5... thinking is always on"; `budget_tokens` returns 400 on Claude 4.7+; control via `effort` | platform.claude.com/docs/en/about-claude/models/migration-guide |
| Context management | Hand-rolled truncation/summarization in app code | Vendor infrastructure: server-side compaction ("Server-side compaction is the recommended strategy"), context editing, memory tool; OpenAI `/responses/compact` with "loss-aware compression" | platform.claude.com/docs/en/build-with-claude/compaction; developers.openai.com/cookbook/examples/gpt-5/gpt-5-2_prompting_guide |
| Upfront context stuffing | Paste everything into the prompt; RAG-everything | Just-in-time retrieval: "lightweight identifiers... dynamically load data into context at runtime"; "Presenting tools as code on a filesystem allows models to read tool definitions on-demand" (98.7% token cut) | anthropic.com/engineering/effective-context-engineering-for-ai-agents; anthropic.com/engineering/code-execution-with-mcp |
| Discipline name | "Prompt engineering" — one-shot craft of instructions | "Context engineering" — "curating and maintaining the optimal set of tokens... during LLM inference"; Anthropic's separate technique doc pages have literally been consolidated into one best-practices page with migration guidance | anthropic.com/engineering/effective-context-engineering-for-ai-agents; platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview |

## Explicit vendor statements that old techniques now HURT

1. **Forced CoT unnecessary (OpenAI):** "Since these models perform reasoning internally, prompting them to 'think step by step' or 'explain your reasoning' is unnecessary." — developers.openai.com/api/docs/guides/reasoning-best-practices
2. **Contradiction-heavy mega-prompts actively damage reasoning models (OpenAI):** "poorly-constructed prompts containing contradictory or vague instructions can be more damaging to GPT-5 than to other models, as it expends reasoning tokens searching for a way to reconcile the contradictions." — developers.openai.com/cookbook/examples/gpt-5/gpt-5_prompting_guide
3. **Aggressive tool-forcing now overtriggers (Anthropic):** "If your prompts were designed to reduce undertriggering on tools or skills, these models may now overtrigger. The fix is to dial back any aggressive language." — platform.claude.com/docs/.../claude-prompting-best-practices
4. **Verification scaffolding now wasteful (Anthropic):** "leaving them in causes over-verification" / "removing them reduces wasted tokens with no loss in quality." — platform.claude.com/docs/.../prompting-claude-opus-5
5. **Prescriptive skills degrade the frontier model (Anthropic):** "Skills developed for prior models are often too prescriptive for Claude Fable 5 and can degrade output quality." — platform.claude.com/docs/.../prompting-claude-fable-5
6. **Examples constrain exploration (Anthropic):** "giving examples actually constrains them to a certain exploration space" — claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models
7. **Old context-maximization prompts backfire (OpenAI/Cursor):** the prompt "often caused the model to overuse tools by calling search repetitively, when internal knowledge would have been sufficient." — developers.openai.com/cookbook/examples/gpt-5/gpt-5_prompting_guide
8. **The general law (Anthropic):** "smarter models require less prescriptive engineering, allowing agents to operate with more autonomy." — anthropic.com/engineering/effective-context-engineering-for-ai-agents

## What did NOT change (honesty slide)
- Clarity still rules: Anthropic's golden rule (colleague test) and "be clear and direct" survive every generation. OpenAI GPT-4.1: "AI engineering is inherently an empirical discipline" — evals still gate everything.
- Structure still helps: XML/delimiters recommended by both vendors in 2026.
- Context is still finite: "treating context as a precious, finite resource will remain central to building reliable, effective agents." (Anthropic)
- New model quirks create NEW steering needs: Opus 5 verbosity, subagent over-spawning, over-verification — the prompt count goes down, but calibration work doesn't disappear; it moves from "make the model act" to "keep the model in scope."