--- slug: prd-workshop title: "The New PRD Workshop: Refine Any Spec for AI Coding Tools" excerpt: "The PRD Workshop is a new Draftlytic feature for refining an existing spec through AI chat tuned to your target coding tool. Here's how it works." primaryKeyword: "PRD Workshop" publishedAt: 2026-04-18 readingTimeMin: 7 author: "Robert Boylan" tags: - prd-workshop - prd - ai-coding-tools - cursor - lovable - refine-prd --- You wrote a PRD last week. It got the app generated, the prototype built, the screenshots into the group chat. It worked. But now you're looking at it again, about to paste it into Cursor for a second pass, and you can already see the parts that are going to make the AI guess. The auth section is one sentence. There's no mention of mobile. The feature list is in the order you thought of them, not the order they should be built. It's not a bad spec. It's just not quite ready for the next prompt. That gap is what the new **PRD Workshop** is for. It's a new Draftlytic feature, live now on every paid plan, that takes a PRD you already have, in any shape, from anywhere, and lets you refine it through an AI conversation tuned to the AI coding tool you're actually going to paste it into. This post walks through why it exists, how it works, and the two different ways you can use it. ## Why most PRDs aren't quite ready for the model A PRD (a product requirements document, basically a plan that says what the app should do before any code gets written) is the single best piece of context you can give an AI coding tool. We've written before about [why AI code drifts the longer you work](/blog/why-ai-code-gets-worse-over-time), and the short version is: the model only knows what you tell it. The PRD is the telling. The catch is that the same PRD doesn't land the same way in every tool. A spec written for Claude can be long, structured, and prose-heavy. Paste that same spec into Lovable and it gets condensed, half-read, or reinterpreted in ways you didn't ask for. Cursor wants short bullets with file names. v0 wants a description of the screen, not the architecture. Bolt.new will happily start building before it understands you. On top of that, a PRD ages. You add a feature. You change auth providers. You decide it's actually mobile-first. The doc you wrote last week is now a few small lies away from being accurate, and those lies are exactly the bits the AI will run with. Most people deal with this by re-prompting until the output looks right. The Workshop is what happens when you fix the spec instead. ## What the PRD Workshop does Open the Workshop from the user menu, the dashboard, or directly from the export modal of an existing project. You paste a PRD (up to 50,000 characters, which is a lot), pick the AI coding tool you're aiming for (Claude, Cursor, Lovable, Bolt.new, v0, Gemini, ChatGPT, Windsurf, Replit Agent, or "not sure"), and you're in. The screen is split in two. On the left, your PRD, rendered as plain text so what you see is exactly what the model sees. On the right, a chat. You ask for a change. The AI rewrites the PRD to match. The new version replaces the old one in the left pane, and the change shows up as a new revision in the header so you can step backward, compare, or revert if it went sideways. It's the smallest possible loop: - **You ask.** "Make the auth section explicit about social login." "Add a non-goals section." "The features list is in the wrong order, prioritise by what's needed for the MVP." "Cut the marketing copy, this is going into Cursor." - **The AI rewrites.** Not just appends. Not just comments. The PRD itself changes, in place, and you can see the new version immediately. - **You keep, revert, or push further.** Each turn is a revision you can roll back to. If the AI overcorrected, hit revert and try again with a tighter prompt. Sessions live in your browser, not the database. That means a Workshop session doesn't sync across devices, and clearing your local storage ends it. It also means there's no surprise data trail and no project record being created in the background. Download the markdown when you're done and the artifact is yours, free of credits. ## Start with a Scan, or just start chatting The first thing you see after pasting is a fork in the road: do you want to **scan** the PRD first, or jump straight into chatting? The Scan is a one-shot diagnostic. It costs 50 credits and runs exactly once, against the PRD you just pasted, with the target tool you picked. It comes back with a verdict (_weak_, _good_, or _excellent_) and a tailored fix list of three to eight concrete things to address. Not vague feedback. Specific lines like "the auth section names Supabase but doesn't say which providers" or "there's no error-state guidance for the upload screen, which Lovable will guess at." Once the Scan finishes, that score and fix list become the first message of your Workshop chat. Which means your follow-up prompts can be tiny. "Apply fix 3." "Tackle the auth one but keep the existing wording for everything else." The AI already knows what you're talking about because it's looking at the same list. Skipping the Scan is fine too. If you already know what you want to change, just open the chat and ask. Use the Scan when you have a PRD you didn't write yourself, or when you've been staring at it long enough that you can't see the gaps anymore. Skip it when you have a specific edit in mind and don't want a structured critique first. ## Why the target tool matters Picking a target tool isn't cosmetic. It changes the AI's editing philosophy. Tell the Workshop you're targeting Claude, ChatGPT, or Gemini, and it leans toward preserving the structure you have, fleshing out thin sections, and respecting the prose. Tell it you're targeting Lovable, v0, or Bolt.new, and it does the opposite, condensing aggressively toward the four to eight thousand token range that those tools handle well, cutting marketing language, and rewriting features as concrete "build this screen" instructions. Cursor and Windsurf get bullet-heavy reference docs with file and component names where it makes sense. Replit Agent gets short, declarative "build me X" framing. "Not sure" is a real choice that gives you tool-agnostic editing without nagging you to pick one. Picking the wrong tool isn't a one-way door. There's a target-tool switcher in the chat header, so if you change your mind halfway through a session, the next turn picks up the new philosophy. ## What it costs (and what's free) The Workshop runs on credits, the same way the rest of Draftlytic does. But the cost shifts with the size of the PRD it's editing, because a bigger doc needs more model work: - **10 credits** per chat turn when the PRD is under 5,000 characters. - **15 credits** when it's between 5,000 and 15,000 characters. - **25 credits** when it's over 15,000 characters. - **50 credits** for a one-shot Scan, regardless of length. The Send button shows the live cost before you click, so there's no guessing. Downloads, copies, reverts, and switching the target tool cost nothing. So does pasting in a fresh PRD and starting a new session. The credits go where the AI work happens, and nowhere else. One last protection: vague prompts get rejected before any credits are spent. "Improve the PRD," "fix it," "clean this up" all bounce with a nudge to either be specific or run a Scan. It's there because vague prompts produce mushy edits, and mushy edits cost credits without helping. See the [full pricing breakdown](/pricing) for how Workshop credits sit alongside the rest of the plan. ## The takeaway A PRD that's almost right is the most expensive kind to ignore. Every prompt you run on top of it inherits its gaps, and after enough rounds you've built something the spec doesn't actually describe. The cheap fix is to fix the spec. That's what the PRD Workshop is for. Paste what you have, pick the tool you're aiming at, ask for the changes you can already see, and walk out with a markdown file that's ready to drop into your next coding session. The thinking is the hard part. The Workshop is the frictionless way to do it.