On July 7, 2026, Ethan Mollick (@emollick) posted a line that landed with 22K+ views in hours:
"Even before the agentic revolution, prompting tricks stopped being very valuable, as our research has shown. The best approach to AI right now is to clearly specify your goals, your output, what 'good' & bad look like, how to test the results... (yes, this is just management)"
A follow-up linked Wharton Generative AI Labs — four Prompting Science technical reports that systematically stress-test the incantations prompt-engineering culture sold for two years: politeness formulas, chain-of-thought, tipping, threatening, and expert personas.
The thread's best reply distilled it: "The unlock was just writing a real spec. Always was — prompting tricks just hid it."
This guide maps Mollick's tweet to Reports 1–4, what the benchmarks actually measured, what still works in the agentic era, and how this connects to loop engineering and Thariq's Fable field guide — where the bottleneck is the map (your spec), not magic words.
TL;DR — What people are asking
Question
Wharton / Mollick answer
Are prompting tricks dead?
Mostly for hard factual tasks — tricks don't reliably move GPQA / MMLU-Pro
What works instead?
Goals, output shape, good/bad rubric, tests — i.e. management
Less on reasoning models; costly on all models (Report 2)
Do expert personas help?
No on accuracy (Report 4) — tone maybe, facts no
Tip or threaten the model?
No significant benchmark gain (Report 3)
Is politeness universal?
No — sometimes helps, sometimes hurts (Report 1)
What format for specs?
Whatever you know — RFP, PRD, SOP; content beats template
What about agents?
Specs + verification loops — prompts are one layer in the stack
Mollick's argument: management, not magic
Mollick's July 7 post makes two claims:
Prompting tricks decayed in value before agents — this is not "Fable made prompts obsolete overnight." Wharton's controlled studies on 2024–2026 frontier models already showed weak, contingent effects on hard benchmarks.
The replacement is specification discipline — define:
Goal — what outcome you need
Output — format, audience, constraints
Good vs bad — acceptance criteria, anti-patterns
Tests — how you verify before shipping
He explicitly says format matters less than content. If your organization lives in PRDs, write a PRD. If SOPs are native, use SOPs. The model does not reward prompt aesthetics — it rewards legible requirements, the same way a strong engineer or contractor would.
That aligns with what Claude Code builders already rediscovered: the map is not the territory — your prompt is a map; the codebase and constraints are territory. Tricks decorate the map; specs align it with reality.
Prompting Science Report 1: Complicated and Contingent
Title:Prompt Engineering is Complicated and Contingent Authors: Lennart Meincke, Ethan Mollick, Lilach Mollick, Dan Shapiro Source:Wharton GAIL
Core finding
The report is about variability, not hero models. The same LLM produces dramatically different scores from small changes in:
How you evaluate — there is no single universal benchmark pass/fail standard; metric choice changes apparent model quality
How you prompt — including politeness and answer constraints
What breaks the "one true prompt" myth
Belief
Wharton result
"Always be polite to the model"
Sometimes helps, sometimes hurts — question-dependent
"Always constrain the answer format"
Sometimes helps, sometimes hurts
"This prompt formula works everywhere"
False for hard, objective tasks
Implication: Prompt engineering as recipe collection fails. You need task-specific instructions and evaluation aligned to your goal — exactly Mollick's management framing.
Report 2: The Decreasing Value of Chain of Thought
Chain-of-Thought (CoT) — instructing models to "think step by step" — on reasoning vs non-reasoning model families.
Results
Model type
CoT effect
Non-reasoning / older / smaller
Modest average gain — but more answer variability, can break questions answered correctly without CoT, much higher token cost
Reasoning-native models
Marginal accuracy gain at best; 20–80% longer responses
Many 2026 frontier models reason internally by default — explicit CoT in the user prompt duplicates work and burns latency and money. See also explainx.ai's zero-shot vs few-shot vs CoT guide — CoT remains a tool, not a religion.
Agent angle: In Claude Code loops, verification beats verbose scratchpad prompts. Put effort in stop conditions and tests, not "think extra hard" boilerplate.
Report 3: I'll Pay You or I'll Kill You — But Will You Care?
Title:I'll pay you or I'll kill you — but will you care? Benchmarks:GPQA, MMLU-Pro Source:Wharton GAIL
What they tested
Two viral social prompting tactics:
Tipping — offering payment for better answers
Threatening — e.g. job-loss or harm framing (publicly endorsed by Google co-founder Sergey Brin in May 2025 per Wharton's report context)
Results
No significant overall improvement on challenging benchmarks
Per-question swings can look dramatic — a prompt helps question A and hurts question B
No reliable way to predict which questions benefit from which emotional framing
July 7 replies joked about unsubscribing from Claude or firing the model — Wharton's data says treat that as performance theater, not engineering.
What might still matter (outside this study): Tone and urgency can affect creative writing or roleplay — these reports focus on hard multiple-choice accuracy, not long-form product work. Do not confuse "no GPQA lift" with "specs don't matter."
Report 4: Playing Pretend — Expert Personas Don't Improve Factual Accuracy
"You are a world-class physics expert" on physics items
No significant accuracy gain (exception: possible effect on Gemini 2.0 Flash only)
Off-domain expert
Physics expert on law questions
Marginal differences — not a reliable boost
Low-knowledge
Layperson, young child, toddler
Generally harmful to accuracy
Practitioner takeaway
Wharton: focus on task-specific instructions, not cosplay. "You are a senior staff engineer" does not substitute for file paths, invariants, test commands, and rollback criteria.
Personas may still help voice and audience — executive summary vs developer doc — but not GPQA-grade factuality. That matches system prompt design in 2026: context and constraints, not fictional CVs.
How the four reports fit together
text
Report 1 → No universal prompt formula; evaluation method matters
Report 2 → "Think step by step" declining ROI on reasoning models
Report 3 → Tipping/threatening ≠ benchmark gains
Report 4 → Expert personas ≠ accuracy gains
Mollick → Replace tricks with goals, rubrics, tests (management)
Agents → Specs feed loops; harness runs verification
The through-line: prompting tricks optimized weak models on easy-to-measure swings. As models and agent harnesses matured, returns moved upstack — to spec quality, context assembly, tool access, and evaluation loops.
What to write instead of tricks — a spec checklist
Mollick's tweet translates into concrete artifacts:
1. Goal statement (one paragraph)
What decision or deliverable this session must produce. Not "help me with code" — "migrate auth middleware to OAuth2 PKCE without breaking mobile clients."
Commands, eval scripts, human review gates — the evaluating prompts mindset applied to agent runs, not single replies.
5. Format you already use
PRD, RFP, SOP, DESIGN.md, CLAUDE.md — Mollick says pick what your team already reads. DESIGN.md templates for UI; loop goals for long-running agents.
Agentic era: why specs matter more, not less
Mollick says tricks faded before the agentic revolution. Agents amplify the stakes:
Tricks era
Agent era
One-shot Q&A
Multi-step execution across tools
Human catches errors immediately
Errors compound across files and minutes
Prompt = entire interface
Prompt + MCP tools + memory + payment rails
An under-specified agent does not fail politely — it ships the wrong thing confidently. Wharton's contingency results predict exactly that: small prompt changes → large outcome swings on hard tasks unless you anchor with tests.
Benchmarks ≠ all work — GPQA/MMLU-Pro are hard multiple-choice; creative, strategic, and organizational tasks may still respond to framing
Personas for tone — Report 4 targets factual accuracy, not brand voice
CoT not forbidden — Report 2 says marginal gains on reasoning models, not "never use CoT"
Persuasion ≠ performance — Wharton's separate 126,000-conversation persuasion study found models still comply more with objectionable requests under influence tactics — social hacks affect safety, not GPQA scores
Do not strawman Mollick into "prompts don't matter." He says tricks don't matter — clear specs and evaluation do.
Bottom line
Ethan Mollick's July 7, 2026 thesis: stop treating LLMs like a spellbook; start treating them like capable, literal executors who need management — goals, output definitions, quality bars, and tests.
Wharton's Prompting Science Reports 1–4 supply the evidence: contingent prompts, declining CoT ROI, useless tips/threats, and non-helpful expert personas on GPQA and MMLU-Pro.
The practical upgrade for 2026 is not a new magic phrase. It is the spec you should have written anyway — now enforced by agent loops, harness code, and verification before merge.
Findings summarize Wharton's Prompting Science reports and Mollick's July 7, 2026 public posts. Benchmarks, models, and best practices evolve — read the full technical reports before changing production agent policies.
