Why Prompt Writing Is Not the Most Important AI Skill

Geoffrey Chen

After AI Enters the Workflow · Season One: “From Answering Questions to Participating in Work” · Article 2

Imagine two people asking AI to draft a project risk report.

The first spends a long time polishing a prompt: “You are a senior risk consultant with twenty years of experience. Analyse this project rigorously and professionally, identify the principal risks and make clear recommendations.” The instruction sounds sophisticated, but it includes no budget, schedule, contract, technical dependencies, record of existing problems or organisational risk criteria.

The second person writes only a few lines. She attaches the current project plan, risk register, latest meeting minutes and reporting template. She identifies which documents are authoritative, marks figures that remain unconfirmed, instructs the AI to derive risks only from the supplied material and requires every assessment to point to evidence. Gaps must be labelled for investigation rather than filled by inference.

The second approach is far more likely to produce usable work. The reason is straightforward: the first prompt mostly assigns a tone and a role; the second creates the conditions in which the task can be performed.

Prompts matter. An ambiguous instruction produces avoidable ambiguity, and goals, formats and constraints should be expressed clearly. But once AI enters real work, the decisive ability is no longer the ability to compose an ingenious sentence. It is the ability to organise tasks, materials, rules, tools and completion criteria into an environment in which a model can participate responsibly.

Prompts solve an instruction problem

A prompt begins by answering a limited question: what do we want the model to do now?

“Summarise this document” is less useful than “Summarise this security report for directors without technical backgrounds in five points; preserve every risk rating and date, and do not add conclusions not present in the report.” The second instruction narrows the audience, output, protected details and evidentiary boundary.

Those distinctions produce real benefits. Clear language can reduce uncertainty, focus attention and make an output easier to inspect. For a one-off translation, reformatting task or brainstorming exercise, the prompt may be the dominant variable.

But most work is not a single input followed by a single answer. A customer response depends on account history, the current contract, refund policy and delegated authority. A code change depends on architecture, dependencies, tests, deployment conditions and repository conventions. A research article depends on verifiable sources, dates, jurisdictions, evidence levels and citation rules. No rhetorical refinement can supply facts that the system has not received.

When a fluent model fills those absences with plausible language, blaming the user’s “bad prompt” conceals the more important failure: the work system did not provide sufficient, correct, current and traceable context.

From prompt engineering to context engineering

The growing interest in “context engineering” reflects more than a change in vocabulary. It expands attention from the immediate instruction to everything a model can use during a task:

  • system rules and user instructions;
  • the current conversation and task state;
  • files, database records and retrieved passages;
  • available tools and the values they return;
  • examples, templates, coding conventions and acceptance criteria;
  • plans, summaries and error records from earlier steps.

Anthropic’s guidance on context engineering for agents treats context as a scarce resource. A working system has to retrieve material when needed, compress completed history and prevent excessive or stale information from overwhelming what matters now. Anthropic, “Effective context engineering for AI agents”

This is why “put everything into the prompt” is not a solution. Research on long-context language models found that relevant material can be used less effectively when buried in the middle of a long input than when placed near the beginning or end—a pattern described as being “lost in the middle”. TACL, “Lost in the Middle: How Language Models Use Long Contexts”

Context engineering is therefore an editorial and architectural task. It asks what must remain in the immediate view, what should be retrieved on demand, what may be summarised, what must remain verbatim, what has expired and which source prevails when records conflict.

Models need levels of authority, not just background information

Work materials do not all carry the same status. An idea raised in a meeting, an unconfirmed figure, a signed contract and an approved policy cannot be treated as equivalent simply because they sit in the same folder.

People familiar with an organisation often recognise these distinctions implicitly. “Draft” means unapproved. An old date triggers a search for a newer version. A colleague’s speculation is not a formal decision. A model may not share that tacit knowledge. If a system supplies text without author, date, version and status, AI can easily present an obsolete instruction as current or turn a proposal into an accomplished fact.

High-quality context therefore requires metadata and relationships:

  • who produced the record and under what authority;
  • which period, project or person it governs;
  • whether it is evidence, an inference, a draft or a formal decision;
  • whether a later record supersedes it;
  • which parts may be disclosed and which are restricted;
  • which source controls when two documents disagree.

Retrieval-augmented generation, or RAG, was developed to combine a model’s internal parameters with external documents that can be retrieved and updated. The original work showed why access to an external corpus can improve knowledge-intensive generation. NeurIPS, “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”

Retrieval does not, however, manufacture authority. It can select an obsolete policy, miss an exception or return a passage with similar vocabulary but a different meaning. The organisation must still curate its sources, preserve their status, assess retrieval quality and make final claims traceable to evidence.

A task also requires a definition of “done”

Even accurate context is insufficient if nobody defines what completion means.

“Analyse these documents” does not say whether the result should identify every contractual deviation, produce a management summary, reconcile numerical differences or recommend a decision. Without acceptance criteria, neither the model nor the reviewer has a stable basis for distinguishing a completed task from an eloquent attempt.

Completion criteria may include required fields, prohibited omissions, numerical tolerances, tests, citations, approval gates and explicit treatment of uncertainty. For an agent, they should also cover process: which tools it may call, how many unsuccessful attempts are acceptable, when it must ask for help and what evidence it must return with the result.

This is closely related to evaluation. Anthropic’s discussion of agent evaluations emphasises defining tasks and graders that expose whether an agent has actually reached the desired state, not merely produced an attractive final message. Anthropic, “Demystifying evals for AI agents”

A good prompt may describe an ideal answer. A good work design establishes observable conditions under which the answer can be adopted.

Why prompts remain so popular

Prompt advice is popular partly because it is visible and individually controllable. A user can rewrite a sentence immediately and compare two outputs. Knowledge governance, access control, source maintenance and evaluation require cooperation across teams. They expose incomplete records and unclear responsibilities that an organisation may have tolerated long before AI arrived.

Prompt culture also fits the conversational interface. The system appears to respond to words, so success seems to belong to the person who knows the right incantation. Yet the interface hides the infrastructure: model configuration, system instructions, retrieval, permissions, tool definitions, stored state and post-processing.

This does not make prompting irrelevant. It places prompting within a larger competence. A professional user should know how to state an objective and constrain an output, but should also recognise when no prompt can compensate for missing evidence, confused authority or an untestable task.

Five abilities that matter more

When AI participates in work, five broader abilities become decisive.

First, task decomposition: separating a broad request into units with different evidence, risk and review requirements.

Second, context selection: providing enough information for the task without exposing unrelated data or burying the governing facts.

Third, source governance: marking origin, date, version and authority, and resolving conflicts among records.

Fourth, evaluation design: specifying what can be tested automatically, what needs expert judgement and what must cause the work to stop.

Fifth, responsibility design: identifying who may approve, reject, correct and account for the result.

NIST’s Generative AI Profile places these concerns within ongoing governance, mapping, measurement and management. The framework’s usefulness lies precisely in moving attention beyond a single interaction to the complete context of use. NIST, “Generative Artificial Intelligence Profile”

Conclusion: AI competence is first an ability to organise work

The most capable AI user is not necessarily the person with the most elaborate prompt. It is the person who can turn a vague objective into a bounded task, assemble the right evidence, distinguish official facts from provisional claims, limit tools and permissions, define completion and create a credible review path.

Prompt writing remains one part of that practice. It expresses an instruction at a particular moment. But the quality of real work depends on the environment in which that instruction is interpreted and acted upon.

The central lesson is therefore:

When AI participates in work, the decisive skill is not finding the sentence that makes the model sound intelligent. It is constructing a work system in which useful output can be distinguished from plausible improvisation.

That skill belongs as much to editors, domain experts, records managers, engineers and accountable decision-makers as it does to “prompt engineers”. AI does not remove the need to organise knowledge and responsibility. It makes the quality of that organisation immediately visible.

Primary sources and further reading

Continue reading: Explore the After AI Enters the Workflow series.

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.