createrole

10 min read

Writing prompts for agents in 2026. Where Anthropic's and OpenAI's official guidance converge

Read side by side, the 2025 to 2026 official agent prompting guides from Anthropic and OpenAI point the same way. Delete instructions instead of adding them, restrain the model instead of pushing it, and write goal contracts instead of step lists. This post collects the most important shared positions and the most actionable rules.

  • prompting
  • agents
  • method

Between 2025 and 2026, Anthropic and OpenAI each published a body of official material on how to write prompts for AI agents: engineering blog posts, platform documentation, cookbooks and guide PDFs. Read side by side, the two sets of documents move in the same direction in 2026. This post collects the most important shared positions and the most actionable rules. Every point comes from the official texts, cited by number with links at the end.

Seven shared shifts

  1. Less is more is now the official position. Both companies ask you to delete prompt text rather than add it. OpenAI's GPT-5.6 guidance reports measured results: trimming the system prompt raised eval scores by roughly 10 to 15 percent and cut tokens by 41 to 66 percent [O8]. The GPT-5.5 guidance says to rebuild from a new minimal baseline when migrating, not to carry the old prompt stack over, and explicitly classifies procedural step-by-step instructions as noise [O7]. Anthropic's guidance for Claude Fable 5 says the same: prompts written for older models are often over-specified and drag down the quality of newer models, and the aggressive phrasing once used to fight laziness ("CRITICAL: You MUST") now causes over-triggering and should be toned down to plain language [A1][A2].
  2. The direction has flipped from pushing to restraining. The classic pattern of early 2025 was nagging the model not to be lazy. The 2026 models follow instructions closely and lean toward action, so official sample sections are now mostly about preventing excess: unrequested actions, over-verification, over-planning, over-use of tools [A1][A2][O6][O8].
  3. From step instructions to goal contracts. OpenAI's recommended template evolved from Reasoning Steps to Goal / Success criteria / Constraints / Stop rules: do not write the process, write what counts as done and how to verify it [O7]. Anthropic's phrase is "right altitude": specific enough to guide behavior, but heuristics rather than choreography of every step [A3].
  4. Both publish copy-ready contract blocks. Anthropic ships <default_to_action>, an anti-fabricated-progress section and a checkpoint section [A1][A2]. OpenAI ships <tool_persistence_rules>, <completeness_contract>, <verification_loop> and <instruction_priority> [O6].
  5. Tool context has become a platform feature. Slimming the tool list moved from prompt trick to platform capability. Both offer on-demand tool search (Anthropic defer_loading [A15], native in OpenAI from GPT-5.4 [O6]) and let the model write code to orchestrate tool calls (Anthropic Programmatic Tool Calling [A15], OpenAI PTC [O8]).
  6. Eval culture is institutionalized. "Eval, read the transcript, make a targeted fix" is the only sanctioned path for prompt iteration. Anthropic: no eval score is accepted until someone has read the transcripts closely [A13]. OpenAI gives a two-step metaprompting method [O3].
  7. A model upgrade is the moment to remove scaffolding. After Opus 4.6 eliminated "context anxiety", Anthropic removed its own sprint-splitting harness [A11]. Guardrails written for old models are often a liability for new ones.

How to write the system prompt

Sections with semantic labels are common ground. Anthropic uses XML tags or Markdown headings to divide blocks, with tag names consistent across prompts and nestable [A1][A3]. OpenAI prefers Markdown headings, finds XML works well, and reports that wrapping long context in JSON performs particularly badly [O1].

Give reasons, not just commands. Anthropic suggests treating the model as a smart new hire: "NEVER use ellipses" works less well than explaining that the output will be read aloud by TTS, because the model generalizes from the reason [A1].

Instruction conflicts are the top-priority bug. GPT-5 follows instructions with surgical precision, so contradictory instructions make the model burn reasoning tokens trying to reconcile them rather than picking a side at random [O2]. GPT-5.4 provides an explicit <instruction_priority> block: user instructions override style defaults, safety and privacy do not yield, newer instructions override older ones [O6].

Demote absolutes. Reserve ALWAYS and NEVER for real invariants such as safety rules and required fields, not for judgment calls [O6]. Anthropic points the same way [A1].

Write personality as behavior rules. Do not write "friendly"; write "state the answer directly; when the user reports a problem, confirm the specific problem before giving next steps" [O8]. Keep personality short. It cannot replace goals, success criteria and tool rules [O7].

When to act and when to ask first

Both companies turned this into a standard section in 2026, and they draw the line the same way: by reversibility of side effects, not by task type.

  • Anthropic: go ahead with reversible local operations, confirm before hard-to-reverse, shared or destructive ones. The sample lists rm -rf, force push and messaging external parties, and forbids destructive shortcuts around obstacles [A1]. The Fable 5 checkpoint section: pause only when the user is genuinely needed, meaning destructive or irreversible actions, real scope changes, or input only the user can provide [A2]. A further boundary: when the user is describing a problem, asking a question or thinking aloud, the deliverable is your assessment. Report findings and stop; do not start fixing [A2].
  • OpenAI: GPT-5.6 grants authority by request type. For answer, explain and review requests, inspect and report without implementing changes. For change and build requests, make local changes and run non-destructive verification without asking. External writes and destructive operations require confirmation [O8].

Dedicated sections for long autonomous runs

Anthropic's Fable 5 guide [A2] is the most systematic official material on this topic:

  • Anti-fabricated progress: before reporting progress, audit every claim against the tool results in this session; report only work with evidence to point to; if tests failed, say so with the output; if a step was skipped, say it was skipped. Anthropic says this section nearly eliminated fabricated progress reports in testing.
  • Anti-early-stop: before ending the turn, check the last paragraph; if it is a plan, a question or a promise, do it now with a tool call.
  • Anti-context-anxiety: avoid showing the model a remaining-token countdown; if you must, add a line that it has ample context left and should not wrap up early because of context limits.

OpenAI's counterpart is a quantified progress-update rule. GPT-5.1 asks for one or two sentences every few tool calls and at least every six steps [O3]; GPT-5.2 tightens this to updates only at a new phase or a plan change, with no narration of routine tool calls [O4]. One counterexample is worth noting: for gpt-5.x-codex, remove every prompt that asks for an upfront plan or a preamble, or the model will stop mid-task, because the behavior is already built in [O5].

Tool descriptions are the highest-leverage text

  • Descriptions are prompt engineering. Anthropic: write a tool description as you would documentation for a new colleague, with usage examples, edge cases, input format and boundaries with other tools [A4]. OpenAI's intern test: given only the information the model gets, could an intern use the function correctly [O9]?
  • Split responsibilities. The tool definition says what the tool is and its parameter format; the system prompt says when to use it and when not to [O9][O1]. In 2026 OpenAI pushes timing, side effects, retry safety and error modes down into the description as well [O7].
  • Design out mistakes. Change the parameter structure so misuse is impossible: make a relative-path parameter require absolute paths [A6]; replace toggle_light(on: bool, off: bool) with an enum [O9].
  • Make errors actionable. No bare error codes; give corrective guidance such as "Invalid format. Expected: user_id (integer)" [A4]. On an empty query result, try one or two fallback strategies before reporting not found [O6].
  • Tool count is about overlap, not number. More than 15 distinct tools is fine; fewer than 10 overlapping ones breaks down [O9][O12].

The platform-level counterparts:

MechanismAnthropicOpenAI
On-demand tool loadingTool Search Tool: in a 58-tool scenario, upfront cost fell from 77K to 8.7K tokens and tool selection accuracy rose from 79.5% to 88.1% [A15]Native tool search from GPT-5.4; allowed_tools restricts the current subset [O6][O9]
Model writes code to orchestrate toolsProgrammatic Tool Calling: intermediate results stay out of context, 37% fewer tokens [A15]PTC, with a call for restraint: prefer direct calls when one call suffices or intermediate outputs are small [O8]
Usage knowledge beyond the schema1 to 5 input_examples in the tool definition; accuracy on complex parameters rose from 72% to 90% [A15]Usage examples in the prompt's Examples section [O1]

Multi-agent: get the single agent right first

Both companies are restrained here in 2026. Anthropic warns that multi-agent systems consume 3 to 10 times the tokens of a single agent, and says to improve the single-agent prompt first and add complexity only with evidence [A14]. OpenAI likewise says to maximize single-agent capability first [O12].

The pattern both endorse as most successful is the verification sub-agent: the main agent does the work, and an independent agent with a clean context verifies it, unpolluted by the implementer's reasoning [A14][A9][A2]. The failure mode to watch is a verifier that runs one or two tests and declares success; the fix is to hard-code "must run the full test suite" [A14]. Split by context boundary, not by job role: the agent that builds a feature should also test it [A14].

Evaluation and iteration

  • Start with 20 to 50 tasks drawn from real failures; the later you start, the harder it gets. A task should be one where two domain experts independently give the same pass or fail [A13].
  • Balance the set: test both when to act and when not to act [A13].
  • Do not over-specify steps in the grader, or an agent that finds a valid alternative path gets marked as failing [A13].
  • OpenAI's two-step metaprompting: first diagnose, giving the model the current prompt and failure logs and asking it to identify the failure mode and cite the prompt lines most likely causing it; then patch with small explicit edits, no restructuring, keeping structure and length roughly the same [O3].
  • Ablation: remove one group of instructions, examples or tools at a time and re-run the eval. This is where the "less is more" numbers come from [O8].

What this means for building digital employees

For any team running agents in production, the shared message compresses to a few sentences. Audit existing system prompts and remove anti-laziness phrasing and over-specified step lists written for weaker models. State when to use and when not to use each tool, and give corrective guidance in error messages. Add anti-fabricated-progress and anti-early-stop sections to the reminders for autonomous tasks. For multi-agent work, prefer a verification sub-agent with a clean context. Build 20 to 50 eval tasks from real failures, and start early.

Sources

Anthropic

OpenAI

Most cookbook and platform pages carry no explicit date; the dates above were cross-checked against the accompanying model release dates, GitHub commit history and third-party citations.