Verdict: Prompt engineering as a job title lost 79% of its postings between 2023 and early 2026 (Neural Digest, April 2026). The replacement is not "better prompting" — it is context engineering (designing the whole information environment around the model, not the wording of one message) and skill-by-demonstration (teaching an agent by recording yourself do the task once, not by writing instructions). Anthropic stripped over 80% of Claude Code's system prompt for Claude 5 models with no measurable quality loss (Anthropic, July 24, 2026), and shipped a "Record a Skill" feature that learns a workflow from a single narrated screen recording (The Decoder, July 2026). Both moves land the same point: writing long, brittle instructions for a smart model is now the expensive way to get mediocre results.
Last verified: 2026-07-27 Three things that actually changed the calculus:
- Anthropic cut 80% of Claude Code's system prompt and the model performed the same (Confirmed).
- Claude's "Record a Skill" turns one screen+voice recording into a reusable, rerunnable skill — no prompt-writing required (Confirmed).
- Prompt-engineer job postings fell 79% from 2023 peak, replaced by workflow/agent design roles (Reported).
Pricing on the new models is volatile — last checked 2026-07-27. Re-verify before any production commitment.
What is context engineering and why does it replace prompt engineering?
Context engineering is the practice of assembling the right information around the model across many requests, instead of optimising the wording of a single request. A prompt is one message. Context is the whole stack: the system prompt, the attached skills, the CLAUDE.md / AGENTS.md files the agent loads on startup, the linked references it can pull in, the memory it carries across turns, and the tools exposed to it. Anthropic draws that line explicitly: "the prompt is only a small part of the context it gets. Much of your context is assembled from your system prompt, Skills, CLAUDE.md files, memory, and other sources" (Anthropic, July 24, 2026).
Why that distinction matters now: every Claude worked example in the public field guide is "an interaction with Claude iteratively to discover what you want to build" Anthropic docs. You no longer send a single perfect prompt — you build an environment that lets the agent iterate. The skill shifted from writing prose to engineering a workspace.
Why is prompt engineering as a standalone skill dead?
Because the models got good enough that they no longer need a human polishing instructions to extract the right behaviour. Three forces converged in 2025-2026:
- The job market collapsed. IEEE Spectrum ran a piece titled "AI Prompt Engineering Is Dead" in 2025. Industry-tracker job postings for "prompt engineer" roles fell ~40% from 2024 to 2025 (Future Factors, April 2026) and ~79% from the 2023 peak to early 2026 (Neural Digest).
- The premium job is now workflow design. Bernard Marr's January 2026 Forbes piece frames the shift plainly: "AI has evolved past the stage of simple, chat-based interactions and carrying out instructions one by one... Today, enterprise AI is agentic" (Forbes, Jan 20 2026). The new skills it names are tool selection, context management, output verification, and deciding where human sign-off still matters — none of which are prompt-writing skills.
- The labs themselves document prompt removal. Anthropic did the experiment for everyone: they removed eighty percent of their carefully-tuned system prompt for Claude Code and the model's coding-eval score didn't move (Anthropic, July 2026). If removing instructions is free, those instructions were never load-bearing. They were overhead the model had been spending reasoning tokens to resolve.
What replaced a prompt-engineer's job is an agent architect's job — designing whole workflows, choosing which tool the agent can call, wiring in verification, and managing context across multi-hour runs. If you want one frame for the new role, see our decision framework on when single-agent vs multi-agent AI wins.
How does Anthropic's context engineering post change how you write instructions?
The lessons from their 80% prompt cut are concrete, and they apply to anyone writing CLAUDE.md, AGENTS.md, agent system prompts, or custom skills. The moves Anthropic made — and recommends you make — are:
1. Convert your inline rules into references the agent can pull on-demand
Claude Code's old system prompt contained detailed sections on code review and verification, always loaded even when irrelevant. Anthropic moved those into separate skills the agent calls only when it needs them. Same information, attached on demand instead of always-on. Less context = less reasoning spent on figuring out which instruction wins. For your own agents, anything that isn't needed at first turn should be a file the agent can @-mention, an MCP server it can call, or a skill file it can load.
2. Express your taste in tool shapes, not paragraphs
Anthropic's own example: instead of long prose telling Claude how to use a todo list, they made the status parameter an enumeration (pending | in_progress | completed) and constrained the count. The shape of the tool is the instruction. If you find yourself writing "make sure to only have one task in_progress at a time," that sentence belongs in the tool's parameter description, not the chat.
3. Stop dual-railing — pick one instruction per behaviour
Their post names the failure mode directly: "one line said to leave documentation as appropriate while another said DO NOT add comments, and the model spent reasoning deciding which instruction won". That's not one prompt's problem; it's a system-prompt problem. Read your CLAUDE.md and AGENTS.md as if you were the model. Any pair of sentences that contradict each other is tax you pay on every turn.
4. Use progressive disclosure, not upfront dumping
The old instinct was to put everything in the system prompt because the model "might need it later." Claude 5-model agents are now good at loading context when they need it. Anthropic moved verification and review into skills the agent selectively calls. Treat your context stack like Progressive Disclosure: anchor the agent with a small set of high-leverage rules, then let it pull depth from files, references, and @-mentions when the task demands them. (If you are setting this up from scratch, the free Claude Code setup guide covers how to structure CLAUDE.md for exactly this rhythm.)
Anthropic shipped a command to do this automatically: run /doctor inside Claude Code v2.1.206 or later, and it will report what it wants to cut from your CLAUDE.md before changing anything (charlesjones.dev summary, July 2026). That is the lowest-risk way to see how much of your context stack is actually carrying load.
What is "Record a Skill" and how does it fit the new model?
While context engineering is the design layer, skill-by-demonstration is the capture layer — the part where you stop writing instructions at all.
On July 22, 2026, Anthropic shipped Record a Skill inside Claude Cowork. The workflow is three steps (The Decoder; byteiota):
- Record. Open Claude Cowork on the desktop app, click the
+menu, select Record a skill. Perform the task you want automated while narrating your reasoning out loud. Claude logs every click, every typed character, and your voice together. - Refine. Claude converts that recording into a written skill file. It asks clarifying questions about anything ambiguous — your narration provides the why, which pure click-replay cannot.
- Run. The next time the task comes up, the skill runs itself. No re-prompting.
The voice thread is what makes this meaningfully better than the click-replay OpenAI shipped for Codex a month earlier. If your workflow has a conditional step — "I add a note here because the client name doesn't match the invoice format" — a click-replay silently drops the judgement. A voice-narrated skill captures it. One stunning example, recorded in early coverage: someone taught Claude to set up an Instagram DM automation in ManyChat — a tool with no API and no MCP. You just can't write that integration in a prompt, and Record a Skill does it anyway (byteiota).
Availability as of Last Verified date: Pro, Max, and Team plans only (Android Authority). The feature runs through Claude-in-Chrome, which makes it slower than a scripted automation — but it works on software that has literally no integration path. For a hands-on look at how computer-use agents navigate this reliability gap, see our 2026 guide to ChatGPT and Claude computer-use reliability.
How do you actually run the post-prompting workflow in 2026?
Here's the trimmed-down workflow which replaces "open Claude and write a better prompt." Five steps:
- Set the environment, not the wording. A two-paragraph
CLAUDE.mdthat names your tools and stack beats a 400-word inline specification every single time. The preloadedAGENTS.mdanywhere serious AI gets done (including our own piece on the editorial pipeline) already follows this discipline. - Record the workflow, don't transcribe it. Open Cowork, click
+, Record a skill. Do the work, talk through the why of each move. The output is a reusable asset, not a one-shot prompt. - Run
/doctoronce a quarter. It tells you what to delete and only changes files with your approval. This is the replacement for "iterate on your prompt." - Use references and skills for the long tail. Anything that's only needed sometimes goes into a file or skill the agent pulls on demand. Anything needed at first turn stays short.
- Verify with tools, not instructions. Spend your effort building verification steps the agent can run itself (Playwright MCP, a test suite, a rubric) instead of writing "please double-check your work." Models that can self-check loop back until the task passes; that's worth more than any paragraph you wrote.
If you want a worked example of how this drills down into agent architecture (single-agent vs multi-agent, one-model vs orchestration), our 2026 orchestration playbook walks through the trade-offs end-to-end.
Are the new Claude 5 models meaningfully better at this workflow?
Record a Skill and /doctor arrive alongside a genuinely-strong Claude release, so the gains compound. Claude Opus 5 launched July 24, 2026 at $5 per million input / $25 per million output tokens — the same price as its predecessor Opus 4.8, and exactly half the price of Claude Fable 5 ($10 / $50) (Anthropic, July 24 2026). On Anthropic's internal coding benchmarks the model comes within 0.5 percentage points of Fable 5's peak score on CursorBench 3.2 at max effort, at half the per-task cost (Anthropic). On Frontier-Bench v0.1 it tops every other model they shipped. The point for this conversation is that Opus 5 needs even less guidance than Opus 4.8 did — the 80% prompt cut was done for Opus 5 and Fable 5 specifically.
| Model | Input $/M tok | Output $/M tok | Context | Best for | Source |
|---|---|---|---|---|---|
| Claude Opus 5 | $5 | $25 | 1M tokens | Default everyday agentic + coding; needs minimal system prompt | Anthropic |
| Claude Fable 5 | $10 | $50 | 1M tokens | Long-horizon, multi-day autonomous agents | Anthropic |
| Claude Opus 4.8 | $5 | $25 | 1M tokens | Same cost tier; Opus 5 supersedes it on benchmarks | Anthropic |
Pricing is the volatile bit — re-verify on Anthropic's pricing page before any production work. For a deep dive on the model itself, see our full Claude Opus 5 review.
What this means for you
For builders and developers. The skill you are now hiring for (or training for yourself) is not "writes the perfect prompt." It's "designs the agent's context stack and lets it iterate." Stop optimising prose. Start optimising the system prompt size, the tool shapes, the file structure, the verification loop. Run /doctor once a quarter and delete what it flags. Keep a real backlog of skills written and recorded — they're your leverage, and Record-a-Skill means non-coders on the team can contribute them.
For small business operators. This is unambiguously good news. The hardest part of getting value out of AI — translating your tacit process knowledge into written form — has a new replacement: just record yourself doing the task with the mic on. Tools that never had APIs or MCP integrations are now in reach of the agent. The lift on your side is small; the lift on the agent's side is real. To start mapping these workflows end-to-end, see how to run a small business on AI agents — the delegation framework and stack.
For hiring managers. A "prompt engineer" line item is mis-priced on both sides. The work that needs doing — workflow architecture, tool design, verification, output quality engineering — is harder than prompt crafting and pays accordingly. Reframe the role before the role reframes itself.
FAQ
Q: Is prompt engineering actually dead, or just renamed? A: The standalone craft of optimizing single prompts is dead — Anthropic's 80% prompt cut, the 79% drop in "prompt engineer" job postings, and Forbes' January 2026 reframing all confirm that. The underlying skill of communicating clearly with a model is more valuable than ever; it just lives inside a broader role now (workflow/architecture design) where context engineering and skill-by-demonstration carry most of the work.
Q: What is the difference between a prompt and context engineering?
A: A prompt is the message you send for a single interaction. Context engineering is the design of everything the model sees across many interactions — system prompt, Skills, CLAUDE.md files, memory, references, and tool surfaces. Anthropic uses context engineering to refer to all the persistent guidance a model uses, distinct from any one user's prompt (Anthropic, July 2026).
Q: What is Claude's Record a Skill feature and how do I get it?
A: Record a Skill is a Cowork feature (in the desktop Claude app, + menu → Record a skill) that captures a single screen recording with voice narration and turns it into a reusable, re-runnable skill Claude can execute on its own. As of Last-verified date it is available to Pro, Max, and Team plan subscribers (The Decoder, July 2026).
Q: How do I apply this without breaking my current setup?
A: Start with /doctor inside Claude Code (v2.1.206 or later). It reports what it wants to cut from your CLAUDE.md before changing anything — a low-risk audit. Convert inline rules to skills and references where possible, pick one rule per behaviour, and move conditional logic into tool parameters rather than prose. Test on a copy of the skill before deleting anything load-bearing.
Q: Does this mean I should never write another prompt? A: No. You still ask good questions; that's basic literacy. What's dead is complex prompt engineering — multi-section system prompts, countless few-shot examples, hard enough to need a full-time "prompt engineer" role. Per the labs' own field guides, plain conversational requests now produce benchmark-level results. Time previously spent prompt-tuning is better spent on tool design and verification (Claude docs; Future Factors).
Q: Do the new Claude 5 models really need fewer instructions? A: Anthropic reports yes — specifically for Claude Opus 5 and Fable 5 they stripped over 80% of Claude Code's system prompt with no measured loss on coding evaluations (Anthropic, July 24 2026). The new models have improved reasoning and tool-use judgement that older models needed the prompt scaffolding to supply.
Discussion
0 comments