AI Catchup

Prompt Engineering in 2026: The Playbook That Works Across Claude and GPT

By 6 min read

Prompt engineering in 2026 has two layers: cross-model fundamentals (clear direct instructions, motivated rules, 3-5 structured examples, XML boundaries, long documents before the question) and model-specific overrides that matter more than ever, because the newest models need instructions deleted as often as added. Start here, then apply the guide for your model.

Prompt engineering in 2026 is two skills, not one. The first is the stable cross-model craft: clear instructions, good examples, unambiguous structure. The second is newer: knowing what your specific model no longer needs, because on the latest frontier models the most common prompting mistake is instructions left over from weaker predecessors. This guide covers the fundamentals both Anthropic and OpenAI publish, then hands off to the model-specific layer.

The model-specific guides: prompting Claude Fable 5 (effort, long runs, memory, safeguards), prompting Claude Opus 5 (verbosity, deleting verification, subagent caps), and prompting GPT-5.6 (roles, structure, effort, agentic patterns). Read the fundamentals here once; read the model guide for whatever you run in production.

Key Takeaways

  • Be clear, direct, and specific. Both vendors' first rule. Anthropic's golden test: if a colleague with minimal context would be confused by your prompt, the model will be too.
  • Explain the why. A rule with its motivation ("never use ellipses, because a text-to-speech engine reads this") outperforms the bare rule; models generalize from the reason.
  • Use 3-5 diverse examples wrapped in XML-style tags. Examples are the most reliable way to steer format, tone, and structure on every current model.
  • Structure with XML and Markdown. Separate instructions, context, examples, and input with tags; order developer messages identity, instructions, examples, context.
  • Long documents go first, query last: up to 30% better responses on multidocument tasks per Anthropic, and cache-friendly ordering saves money on both platforms.
  • Control cost with effort, not thinking bans. Adaptive thinking plus effort settings replaced manual budgets; low effort with thinking on usually beats thinking off.
  • For agents, prompt for state: structured test files, progress notes, git checkpoints, and grounded progress claims carry long tasks across context windows.

The Fundamentals: Clarity, Motivation, Examples

Modern models follow instructions well enough that the old bag of tricks ("act as an expert," "you are the world's best programmer") matters far less than the quality of information you provide. What replaced it:

Be explicit about the output you want. "Create an analytics dashboard. Include as many relevant features and interactions as possible. Go beyond the basics" reliably outperforms "create an analytics dashboard." If you want above-and-beyond behavior, ask for it; if you want minimal scope, say that instead.

Attach the reason to the rule. Anthropic's canonical example: "never use ellipses" is weaker than "your response will be read aloud by a text-to-speech engine, so never use ellipses, since the engine cannot pronounce them." The model generalizes from the explanation to cases your rule never anticipated. This is also why context about who the work is for and what it enables improves agent output.

Show, don't describe. Include 3-5 examples that mirror your real use case, vary enough to avoid teaching unintended patterns, and sit inside example tags so the model can tell demonstration from instruction. Both vendors rank this among the most reliable techniques they have.

Structure: Roles, Tags, and Document Placement

Give every prompt an unambiguous skeleton. On OpenAI's Responses API that means the instructions parameter or a developer message for rules (prioritized ahead of user input), with the developer message ordered identity, then instructions, then examples, then context. On Claude, a one-line role in the system prompt focuses tone and behavior, and XML tags (<instructions>, <context>, <input>) keep mixed content unambiguous on both platforms.

For long-context work (20k+ tokens), placement matters more than wording. Put documents at the top, wrapped in per-document tags with source metadata, and your question at the end; Anthropic reports up to 30% better response quality from this ordering on complex multidocument inputs. For extraction-heavy tasks, ask the model to quote the relevant passages first, then answer from the quotes. And keep reusable content at the beginning of the prompt: both platforms' prompt caching rewards stable prefixes with real cost and latency savings.

Thinking, Effort, and the New Cost Controls

The cost dial moved. Anthropic's current models use adaptive thinking, where the model decides when and how much to think, calibrated by an effort parameter; manual thinking budgets are deprecated, and on the newest models setting one returns an error. OpenAI exposes reasoning.effort from none through max on GPT-5.6. The shared guidance: prefer lowering effort over disabling reasoning, because low effort with thinking enabled usually outperforms no thinking at similar cost.

Two prompting consequences follow. First, general instructions beat prescriptive steps: "think thoroughly" often produces better reasoning than a hand-written plan the model must follow. Second, tune anti-laziness prompting down when migrating: instructions written to fight undertriggering on older models ("CRITICAL: you MUST use this tool") cause overtriggering on current ones. Normal language ("use this tool when...") is enough.

One deprecation worth calendaring on each side: Anthropic no longer supports prefilled assistant responses on current models (migrate to structured outputs or direct instructions), and OpenAI shuts down reusable prompt objects (v1/prompts) on November 30, 2026 in favor of prompts versioned in code.

Agentic Prompts: State, Scope, and Honesty

Long-running agent work has its own fundamentals, consistent across vendors:

ProblemThe prompt-level fix
Losing track across context windowsStructured state files (a tests.json, a progress notes file), git checkpoints, and a first-window prompt that sets up tests and setup scripts
Scope creep and over-engineering"Only make changes directly requested or clearly necessary"; no abstractions for one-time operations, no defensive code for impossible scenarios
Gaming the tests"Tests verify correctness, not define the solution"; require general implementations and have the model flag incorrect tests rather than working around them
Hallucinated claims about code"Never speculate about code you have not opened"; require reading referenced files before answering
Fabricated progress reportsRequire every status claim to trace to a tool result from the session, with unverified work flagged as such
Parallel tool usageExplicitly ask for independent tool calls in parallel; both platforms' current models sustain it near 100% when prompted

Constraints remain the guardrails they always were: output format ("return only the function body"), scope ("only modify handleSubmit"), pattern ("use the error handling from lib/errors.ts"), and negative constraints ("do not add new dependencies"). Layer them, and keep persistent, project-level context in instruction files like CLAUDE.md so every session starts oriented; see our Claude Code context management guide for that discipline.

Iterate, Then Delete

Iterative refinement still beats one-shot prompting: structure first, complexity in layers, targeted corrections, and a restart with better initial context once you have made three corrections without converging. But 2026 adds a final step the old loop lacked: after your prompt works, delete lines from it and re-test. On the newest models, instructions the model no longer needs are not merely dead weight; verification instructions cause over-verification on Claude Opus 5, prescriptive skills degrade Claude Fable 5 output, and aggressive tool-triggering language causes overtriggering everywhere. The best prompt is the shortest one that still passes your evals.

Sources

Keep building the workspace playbook

Frequently Asked Questions

Is prompt engineering still relevant with 2026 frontier models?

Yes, but it inverted in places. The fundamentals (clarity, examples, structure) still lift quality on every model, while over-prompting now actively hurts: Anthropic warns that verification instructions cause over-verification on Opus 5 and that prescriptive skills degrade Fable 5 output. Knowing what to delete is the new skill.

What is the single highest-leverage prompting technique?

Structured examples. Both vendors put well-crafted few-shot examples near the top: Anthropic recommends 3-5 diverse examples wrapped in example tags, and OpenAI recommends diverse input/output pairs in the developer message. Examples steer format and tone more reliably than any written description of them.

Where should long documents go in a prompt?

At the top, above the query and instructions. Anthropic reports that placing queries at the end after longform inputs can improve response quality by up to 30% in tests, and recommends wrapping each document in XML tags and asking the model to quote relevant passages before answering.

How do I control cost on reasoning models now?

Through effort settings rather than thinking budgets. Anthropic deprecated manual thinking budgets in favor of adaptive thinking plus an effort parameter, and OpenAI exposes reasoning effort from none to max on GPT-5.6. Lower effort with thinking enabled usually beats disabling reasoning outright.

Get the weekly AI Catchup

Tools, practices, and what matters, in your inbox every week.