How Prompting Claude Changed: The 2024 Playbook vs. What Works Now

Every Era's Advice Was Right, Then Wrong

If you learned to prompt AI from a 2024 CLE, a LinkedIn carousel, or a prompt pack for lawyers, you learned real techniques that produced real improvements. Role-play framing genuinely helped 2023-era models hold a professional register. Step-by-step instructions genuinely stopped 2024-era models from skipping steps. "Think step by step" genuinely deepened reasoning on models that would not reason deeply on their own.

The uncomfortable part: every one of those techniques compensated for a specific model weakness, the weaknesses got engineered away, and the compensations did not become harmless leftovers. They became costs. A prompt built for the model you learned on can actively degrade the model you use now. This post is the story of how that happened and what replaced it, because understanding why the old advice existed is the fastest way to stop paying for it.

The Incantation Era (2023 to Mid-2024)

Early chat models were brilliant and unreliable in equal measure, and users discovered that certain phrases shifted the odds. "You are an expert attorney" nudged the register upward. "Think step by step" (a technique from an actual research paper) elicited visible reasoning that improved answers. "Take a deep breath" measurably helped one leading model on math problems, which tells you how strange this era was. Prompting looked like spellcasting because, functionally, it was: nobody fully knew why the phrases worked, only that they did.

The legal profession translated this into templates. Prompt packs sold role preambles, output scaffolds, and magic closers ("be thorough, cite your sources, do not hallucinate"), and they were worth using at the time.

The Framework Era (2024 to 2025)

As models steadied, the advice matured into acronym frameworks: role, context, task, format, and their many cousins. We taught one ourselves, and our CRAFT framework guide is still one of the most-read pages on this site. Frameworks were a real upgrade over incantations because most of their content was not magic at all: they forced you to state the facts, the audience, and the deliverable, which is simply briefing the work properly.

But the frameworks smuggled the incantation era forward inside them. The R in every acronym was role-play. The step-by-step choreography stayed. The verification commands multiplied ("double-check every citation before responding"). Templates grew because adding a rule always felt safer than deleting one, and by mid-2025 a typical legal prompt template was two pages of accumulated compensations wrapped around four lines of actual substance.

The Turn (Late 2025 to 2026)

Then the model builders started deleting their own instructions. Anthropic cut more than 80 percent of Claude Code's system prompt for the current model generation, and quality did not drop. Boris Cherny, the creator of Claude Code, began giving users advice that would have been malpractice two years earlier: periodically delete your instruction files entirely and watch what the model does on its own. The community summarized the philosophy in three words: let the model cook.

The reasoning is not mystical. Current models plan multi-step work unprompted, verify their own output unprompted, and hold a professional register unprompted. Instructions that force last year's workarounds onto this year's model do not add safety; they override better judgment. Your twelve-step contract-review script makes the model follow your twelve steps instead of the better plan it would have made itself. Your "double-check everything" line buys hedging, not accuracy. Your role preamble buys nothing at all. We covered the deletion side of this shift in depth in Stop Over-Prompting Claude; what follows is the constructive side.

What Works Now: Context Over Incantations

The current method fits in one sentence: state the goal and the constraints that matter, supply the context only you have, and let the model find the path. In practice, a strong legal prompt today has five ingredients, none of them magic:

• The facts of the matter. Generosity here has always helped and still does. Two pages of relevant facts with one sentence of instruction beats one paragraph of facts wrapped in two pages of rules.
• Jurisdiction and governing law. "Under Florida law" changes the answer. No model generation has made this optional.
• Audience. A memo for a partner, a letter to opposing counsel, and an email to an anxious client are three different documents from the same facts. Say who reads it, not who the model should pretend to be. This is what the old role-play was groping toward.
• Your actual constraints. Length limits, the client's authorized range, positions your firm will not take, citation style. Constraints are the load-bearing part of a professional prompt; they survived every purge.
• Output format. Table, numbered issues, letterhead-ready letter. Cheap to specify, expensive to retrofit.

Compare the eras on the same task. The 2024 version: "You are a senior insurance-defense partner with 20 years of experience. Think step by step. First, review the demand letter. Second, identify each claim..." and ten more steps. The 2026 version: "Draft a response to this demand letter that positions us for the mediation. Facts and the letter attached; Ohio law; audience is opposing counsel who overreached and knows it; two pages; firm but door-open." The second prompt is shorter, and on current models it produces the better letter.

What Survived Every Era

Three things never stopped being true, and they separate minimal prompting from lazy prompting:

• Exact requirements stay exact. Court-mandated formats, citation form, filing procedures: where only one sequence is acceptable, spell it out precisely, every time. Guidelines loosen; requirements never do.
• Verification is a workflow, not a prompt line. "Cite your sources" never made citations real, and more than a thousand court filings with fabricated citations have been logged publicly to prove it. Every citation gets confirmed in a real database before it enters a filing, in every era, under every prompting philosophy.
• Iteration is the actual skill. The first prompt is a draft. "Tighten this," "more formal," "now a client version" have always been where the quality comes from, and conversation is cheaper than the perfect mega-prompt in every model generation.

If You Have a Template Folder, Do This

Do not delete it; retune it. Save a copy, strip each template down to the five ingredients above (cut the role-play, the step choreography, the thinking commands, the verification theater), run old and new on the same closed matter, and keep whatever the comparison proves. Most lawyers find the minimal version wins and the deleted lines were costing quality the whole time; occasionally a line turns out to be load-bearing context in disguise, and it earns its way back in as one plain sentence. Put the exercise on a six-month cycle, because the model under your templates will change again.

The deeper lesson of the whole arc: prompting advice has a shelf life, and the only durable skills are the ones that were never model-specific: knowing your facts, your audience, your constraints, and your duty to verify. Those were lawyering skills before they were prompting skills. They will outlast whatever comes after this era, too. For the working library that applies all of this, our Claude prompts for lawyers collection and the prompt library are maintained against the current models.