Prompt Chain Vault — Draft, Critique & Finalize Sequences for 8 Real Business Deliverables

Prompt Chain Vault is a library of eight original, three-step prompt chains — Draft, Critique, Finalize — for the business documents that are hardest to get right in a single pass. No agent framework, no API key, no plugin required.

Why this exists

In September 2026, the fastest-growing prompt repository on GitHub isn't a coding tool — it's prompts.chat (formerly Awesome ChatGPT Prompts), a community-maintained library that's crossed 169,000 stars and bills itself as the world's largest open-source prompt collection, complete with a CLI, an MCP server, and a Claude Code plugin for teams that want to self-host it. It's a genuinely useful reference: thousands of single-turn "act as X" prompts spanning every role imaginable, MIT/CC0 licensed, free to fork and remix.

But scale is exactly its limitation for real work. Every entry in it is one isolated prompt — paste it, get a single response, and you're on your own for the two or three follow-up turns it actually takes to get a board memo, a PRD, or a case study into shape you'd send to someone else. Quality also varies prompt to prompt, since anyone can submit one, so there's no guarantee the prompt you paste today produces something usable on the first try.

Prompt Chain Vault takes the opposite approach — not any of that project's code or prompt text, just the observation that one-shot prompting fails on anything with real stakes. Instead of a single prompt per document, each of the eight entries here is a fixed three-step chain: a Draft prompt that produces a complete first pass, a Critique prompt that turns the model into an editor on its own output, and a Finalize prompt that revises against that critique into something ready to send. Every chain is original writing, built and tuned by hand.

What's inside

• 8 complete three-step prompt chains (24 prompts total) for high-stakes business documents
• A Universal Chain Builder — turn any one-shot prompt you already rely on into your own Draft → Critique → Finalize chain
• A Chain Integrity Checklist — 6 checks to run before a chain's final output goes anywhere real

Best for

• Founders, managers, and consultants who need a document that's actually ready to send, not a competent-but-generic first draft
• Anyone who's pasted one giant "write the whole thing" prompt into a chat window and gotten back prose that reads fine but doesn't hold up under scrutiny
• People without an editor, cofounder, or second pair of eyes to catch what a first draft misses before it goes out

What you'll need

• Any chat LLM — ChatGPT, Claude, Gemini, or similar — no plugins, agents, or API access required
• The raw facts for whatever you're writing (names, numbers, dates, context) — each Draft prompt asks for these upfront
• Three consecutive turns per document, run in the same conversation — each prompt in a chain depends on the model's prior reply

Chain discipline (read this first)

Every chain below follows the same three-step shape. Run all three, in order, in one conversation — don't skip Critique even when the Draft looks fine, since catching what a first pass missed is the entire point:

1. Draft — produces a complete first version fast, from the facts you provide. Not precious; its job is to get everything on the page once.
2. Critique — the model re-reads its own draft against an explicit, document-specific checklist and lists every weakness it finds. It does not fix anything at this step, only diagnoses.
3. Finalize — revises the draft against the critique and outputs the final, formatted version. This is the one you actually send.

The 8 prompt chains (copy/paste)

1. Board / Investor Update Memo

Draft: "Write a board update memo covering this period. Inputs: company name [NAME], period [DATES], key metrics [LIST METRICS WITH NUMBERS], wins [LIST], challenges [LIST], asks of the board [LIST IF ANY]. Structure: TL;DR (3 sentences max), Metrics, Wins, Challenges, Asks, Next period focus. Keep it under 600 words and state numbers plainly — no spin."

Critique: "Re-read the memo you just wrote as a skeptical board member seeing these numbers for the first time. List every place where: a claim isn't backed by a number, a challenge is downplayed relative to its actual severity, the TL;DR oversells the period, or a metric is presented without the context needed to judge if it's good or bad. Don't rewrite anything yet — just list the problems."

Finalize: "Rewrite the memo addressing every issue from your critique. Every challenge must be stated as plainly as every win. Every metric needs enough context (trend, target, or comparison) for a board member to judge it without asking a follow-up question. Output the final memo only, no commentary."

2. Crisis / Incident Communication Statement

Draft: "Write a public or customer-facing statement about this incident. Inputs: what happened [FACTS ONLY], who is affected [SCOPE], current status [RESOLVED/ONGOING/MITIGATING], what we're doing about it [ACTIONS], and what we're NOT yet sure of [UNKNOWNS]. Tone: direct, accountable, no corporate hedging language. Under 250 words."

Critique: "Read this statement as an affected customer who is already frustrated. Flag every sentence that reads as deflecting blame, minimizing impact, or using vague language where a specific fact was available in the inputs. Flag any claim of resolution that isn't fully supported by the stated status. List problems only, don't rewrite."

Finalize: "Revise the statement to fix every issue from the critique. Replace any hedging phrase with the direct, specific version. If status is not fully resolved, say so plainly rather than implying otherwise. Output the final statement only."

3. Product Requirements Document (PRD)

Draft: "Write a PRD for this feature. Inputs: feature name [NAME], problem it solves [PROBLEM], target user [USER], proposed solution [DESCRIPTION], out of scope [LIST], success metric [METRIC]. Structure: Problem, Goals, Non-Goals, User Stories, Requirements (numbered, testable), Success Metrics, Open Questions."

Critique: "Review this PRD as a skeptical engineer who has to estimate and build it. For each numbered requirement, flag whether it's actually testable (someone could verify pass/fail) or vague enough to be interpreted three different ways. Flag any user story with no corresponding requirement, and any requirement that contradicts a stated non-goal. List issues only."

Finalize: "Rewrite the PRD fixing every flagged requirement so each one is independently testable. Resolve the flagged contradictions explicitly rather than deleting one side silently — note in Open Questions if a real tradeoff decision is still needed. Output the final PRD only."

4. Customer Case Study

Draft: "Write a customer case study. Inputs: customer name/industry [DETAILS], problem before using us [PROBLEM], what we implemented [SOLUTION], results with real numbers [METRICS], a quote if available [QUOTE]. Structure: Challenge, Solution, Results (numbers first), Quote. Under 500 words."

Critique: "Read this as a skeptical prospect who has been burned by vague case studies before. Flag every result stated without a number, timeframe, or baseline to compare against. Flag any claim that isn't directly supported by the inputs given — don't let inferred or invented specifics slide through. List issues only."

Finalize: "Rewrite the case study fixing every flagged claim — either add the missing number/timeframe/baseline from the original inputs, or soften the claim to only what's actually supported. Never invent a statistic that wasn't in the inputs. Output the final case study only."

5. Post-Mortem / Incident Report (Internal)

Draft: "Write an internal post-mortem. Inputs: what happened and when [TIMELINE], impact [SCOPE AND DURATION], root cause if known [CAUSE], contributing factors [LIST], what was done to resolve it [ACTIONS], follow-up items [LIST]. Structure: Summary, Timeline, Root Cause, Contributing Factors, Action Items (owner + due date each), Lessons. Blameless tone — describe what the system allowed, not who erred."

Critique: "Re-read this post-mortem for two failure modes: (1) any sentence that names or implies blame on a specific person rather than a system or process gap, and (2) any action item missing an owner or a due date. List every instance of each, don't rewrite yet."

Finalize: "Rewrite the post-mortem fixing every flagged sentence to describe the system/process gap instead of the person, and adding a placeholder [OWNER] / [DATE] to any action item missing one so it's visibly incomplete rather than silently dropped. Output the final report only."

6. Performance Review Narrative

Draft: "Write a performance review narrative for this employee. Inputs: role [ROLE], review period [DATES], specific accomplishments with impact [LIST], areas for growth [LIST], overall rating context [RATING/LEVEL IF APPLICABLE]. Structure: Summary, Strengths (with concrete examples, not adjectives alone), Growth Areas (framed as specific and actionable), Overall. Keep language behavior-based, not personality-based."

Critique: "Read this review as an employee relations specialist checking for two things: (1) any strength or growth area stated as a personality trait ('is a great communicator') rather than a specific, observable behavior or outcome, and (2) any growth area that's vague enough the employee wouldn't know what to actually do differently. List every instance, don't rewrite yet."

Finalize: "Rewrite the review converting every flagged trait-based statement into a specific behavior or outcome from the inputs, and every vague growth area into one concrete, actionable next step. Output the final review only."

7. Executive Summary for a Long Report

Draft: "Here is a full report: [PASTE FULL REPORT TEXT]. Write a one-page executive summary: the single most important finding first, then 3-5 supporting points each with the number or evidence behind it, then the recommended action. No new claims beyond what's in the report."

Critique: "Compare your summary against the full report line by line. Flag anything in the summary that isn't directly traceable to a specific passage in the report, and flag any major finding in the report that the summary left out entirely. List issues only, don't rewrite yet."

Finalize: "Rewrite the summary removing or correcting any untraceable claim, and adding back any major finding the critique flagged as missing. Every point must trace to something actually in the report. Output the final summary only."

8. Renewal / Upsell Proposal Email

Draft: "Write a renewal/upsell email to this customer. Inputs: customer name and current plan [DETAILS], usage/results since they started [DATA], the specific upgrade or renewal being proposed [OFFER], pricing [PRICE], any time-limited terms [TERMS IF ANY]. Under 200 words, one clear ask, no generic filler."

Critique: "Read this email as the customer, already skeptical of upsell emails. Flag any sentence that's generic filler rather than specific to this customer's actual usage/results, any place urgency is implied without a real time-limited term to back it up, and anywhere the ask isn't singular and clear. List issues only."

Finalize: "Rewrite the email removing every flagged filler sentence, removing any implied urgency not backed by a real term, and sharpening the ask to one clear next step. Output the final email only."

Supporting framework: Universal Chain Builder

Use this to turn any single prompt you already rely on — from this library, prompts.chat, or your own notes — into your own three-step chain.

"I currently use this single prompt for [TASK]: [PASTE YOUR EXISTING PROMPT]. Turn it into a three-step chain: (1) a Draft prompt that produces a complete first pass from the same inputs, (2) a Critique prompt that reviews that draft against 3-5 specific, document-appropriate failure modes and lists them without fixing anything, and (3) a Finalize prompt that revises against the critique into a final version. Output all three as separate, ready-to-paste prompts."

Supporting framework: Chain Integrity Checklist

Run any chain's final output — from this library or one you built yourself — through these 6 checks before it goes anywhere real:

1. Critique was actually run: Did you paste the Critique prompt and read its output before Finalize, or skip straight from Draft to "good enough"? Skipping it defeats the point.
2. Every critique point addressed: Compare the Finalize output against the Critique list — is every flagged issue actually fixed, or did one quietly survive?
3. No invented facts: Does the final version contain any number, quote, or claim that wasn't in your original inputs? Flag and remove anything invented.
4. Tone consistency: Does the Finalize step's tone match what the document actually needs (blameless, direct, accountable — whichever the chain called for), or did it drift back to generic corporate voice?
5. Length discipline: Does the final version respect the length constraint the Draft prompt set, or did revision quietly bloat it?
6. Read-aloud test: Read the final output aloud once. Anything that sounds like it was written by a committee instead of a person gets flagged for one more pass.

How to use

1. Pick the chain matching the document you need to produce.
2. Fill in the bracketed inputs on the Draft prompt and send it.
3. Paste the Critique prompt next, in the same conversation — don't skip it.
4. Paste the Finalize prompt and take its output as your final version.
5. Run the Chain Integrity Checklist before sending the result anywhere real.
6. For any document not covered here, run the Universal Chain Builder against a prompt you already use to build your own chain.

FAQ

Is this the prompts.chat / Awesome ChatGPT Prompts repository? No, and it doesn't contain any of that project's prompt text or code. This library takes the observation that single-shot prompting fails on high-stakes documents and builds eight original three-step chains around it, for people who want a finished document, not a library of one-off ideas to browse.

Do I need an agent, API access, or a specific model to use this? No. Every prompt is plain language designed to be pasted into any chat LLM's normal chat window, one turn at a time.

Is this affiliated with OpenAI, Anthropic, Google, or the prompts.chat project? No. This is an independent, model-agnostic prompt library and isn't affiliated with or endorsed by any AI lab or the maintainers of the project that inspired it.

Can I use these chains with my own documents, not the examples given? Yes — every Draft prompt is written with bracketed inputs specifically so you can substitute your own facts, numbers, and context.

What if Finalize still misses something from the Critique? Reply with "you didn't address [SPECIFIC POINT] from your critique — fix that and re-send the final version" rather than accepting a partial fix.

Notes: Digital product, delivered as the copy/paste content on this page — no separate file download required. All 24 prompts are original writing.