How prompt engineering changed in 2026

Is prompt engineering dead in 2026?

Prompt engineering in 2026 is not dead; it split in two. Casual prompting (personas, magic phrases, few-shot rituals) became a commodity the models absorbed. What remains is a narrower, sharper skill: writing the minimum correct specification for a specific model, and knowing where each frontier model's defaults end and your instructions need to begin. That is the short answer; the rest of this post lays out the evidence, and it points in one direction: the craft moved from addition to subtraction.

The public debate is loud (IEEE Spectrum says dead, Forbes says evolved), but both camps are describing the same event from different vantage points. What died is the 2023 grimoire of prompt hacks; what replaced it is per-model spec discipline. Ethan Mollick's framing is the cleanest: stop treating LLMs "like a spellbook" and start treating them as capable, literal executors who need management.

The graveyard: what stopped working

"Take a deep breath and think step by step" once bought ~8 points on GSM8K, in one 2023 paper, for one PaLM 2 configuration, found by automated search (Yang et al., OPRO). A 2026 re-evaluation with a reproducible harness found the delta on modern models is "small, zero, or even slightly negative." The trick got trained into the distribution; the incantation became furniture.

Few-shot examples, the most canonical technique of all, now measurably hurt on three task shapes: extraction (the model anchors on your example's item count), creative work (it photocopies the example's voice), and strict formatting (it copies incidental choices like quote style), per a May 2026 ablation study whose sharpest finding is that example length matters more than count: one 800-token example can quietly tank accuracy. The giant system prompt did not survive either: OpenAI's measured 10-15% accuracy gain from deleting redundant instructions undercuts the assumption that more prompt is better.

Sean Goedecke made the same argument before this round of results: "Prompts are a worse form of technical debt than code… Prompts will decay silently… every single model upgrade could turn a functional prompt into a non-functional one." Theo made a whole video from that essay. His conclusion is one this blog has been circling since post 004: a prompt tuned for one model version can become the reason the next version underperforms.

How do you prompt Opus 4.8, Fable 5, and GPT-5.6 differently?

The three models most people are routing between right now have divergent prompting profiles, according to their own vendor documentation. The same instruction produces different outcomes depending on which model receives it.

Sources: Anthropic's Opus 4.8 and Fable 5 prompting docs; OpenAI's GPT-5.6 guidance. All vendor-published; the cross-model routing consequences are ours.

The table shows a pattern: Opus 4.8 wants scope made explicit because it stopped guessing. Fable 5 wants goals, because a carefully enumerated procedure is now a ceiling on its output: Anthropic states that skills written for prior models "are often too prescriptive for Claude Fable 5 and can degrade output quality." GPT-5.6 wants silence, because every redundant sentence is measured drag. Each model's correction points the same direction, less prompt rather than more; what differs is which words are being cut, and why.

Two more vendor details are worth noting. Anthropic's fix for Fable's long-run fabrication (the audit-each-claim line in the table) reportedly "nearly eliminated fabricated status reports even on tasks designed to elicit them," covering with one sentence ground a monitoring harness used to cover. Opus 4.8 also has a house aesthetic (cream backgrounds, serif display, terracotta) that generic negative prompting does not remove: the options are writing a fully concrete spec or asking it to "propose 4 directions first." That kind of vendor-specific detail is what prompt engineering has become: knowing what a specific model needs, not memorizing incantations.

Shorten your skill descriptions

This section is the practical recommendation the rest of the post has been building toward: your agent's skill and tool descriptions are almost certainly too long, and that costs you twice.

Cost one is selection accuracy. Tool-selection studies show accuracy collapsing as tool count and description bulk grow: one cited measurement has selection accuracy dropping from 43% to under 14% as tools pile up, and Anthropic engineering has put it near 49% past 50 tools, because "the attention cost of parsing many schemas degrades the model's ability to reason about any of them well." Meanwhile the "MCP Tool Descriptions Are Smelly!" paper found the ecosystem's median description is already ~10 tokens - the verbose ones are outliers, and not the good kind. Cost two is that every description rides in the context you are paying for, every turn, whether the skill fires or not.

The fix that has worked on this blog's own pipeline, offered as a labeled anecdote, is to rewrite every skill description into two clauses, "Use if…" and "Don't use if…," and delete everything else. The negative clause matters more than the positive one. Modern frontier models are so eager to be helpful that an unfenced skill fires constantly. Skills on this pipeline that seemed valuable ended up firing on everything, and the fix was a sharper fence: narrowing the trigger condition rather than lengthening the description, or in some cases switching the skill back to explicitly user-invoked. Anthropic quietly did the same thing to itself, shrinking its own frontend-design skill prompt for Opus 4.8. Progressive disclosure is the same principle at the architecture level: a short description stays always loaded, and the full instructions load only on invocation.

What to actually do

• Run the deletion pass. Take your longest system prompt, remove every repeated or "obvious" instruction, and eval before and after. OpenAI's numbers say you should gain accuracy while cutting cost by a third to two-thirds.
• Keep a per-model prompt profile, not a universal prompt. Three short files: Opus (explicit scope, positive examples, effort=xhigh for agentic work), Fable (goals not procedures, the anti-fabrication audit line, hardest version of the task), GPT-5.6 (deduplicated, one reasoning tier lower than you think).
• Rewrite skill descriptions as "Use if… / Don't use if…" and demote over-eager skills to user-invoked. A precise negative clause that limits when a skill fires does more than additional descriptive text meant to make it sound useful.
• Treat surviving prompts as debt. Re-run your prompt evals on every model upgrade; Goedecke's argument is that skipping this lets prompts decay silently until an upgrade breaks one without warning.
• Then look up the stack. The prompt layer is shrinking; the next post, Part 3: loop engineering, covers where that engineering effort moved.