Murder Mystery Architect AI Kit — 4 Ready-To-Use AI Agents to Build a Mystery Party You Can Play In Too

From AGENTORY — ready-to-use AI Task Kits. Drop in. Power up.

Works with the AI you already have. Upload one file, tell it "Start", answer a few questions — then build one original mystery on a single Case File that nothing may contradict, keep the solution sealed so you, the host, can play blind, and produce the full event package: character booklets, clue cards, invitations, and a sealed reveal.

Four specialist agents that take your mystery party from "we should host one" to a complete, checked, sealed event package — running inside the AI you already use.

No new app. No subscription. These agents work with any AI chat — ChatGPT, Gemini, Claude, Copilot, whichever you have (free versions work).

How it works (about 5 minutes per agent, once):

1. Add the agent file to your AI chat — upload the .md file, or on a phone, open the .txt version, copy it, and paste it in.

2. Type "Start".

3. Answer the agent's questions, one at a time — your guest count, your setting and theme, your tone. That's the calibration, and it's what makes the mystery fit YOUR party.

4. Save the complete updated Case File and Party Profile each agent hands back to you — the agent keeps it current in the chat and returns it whole, never as pieces for you to assemble — and the next agent starts from your real party.

Use the agents in order:

1. Plot Architect — builds the Case File: the victim, the suspects, the motive web, the true timeline, and the alibi matrix — and writes the solution only into the Case File's sealed zone, so everything it shows you is structure, never the answer.

2. Character Writer — writes a full player guide for every guest from the Case File, then runs the balance check: every guest holds at least one unique, needed piece of the puzzle, and no guide contains another character's private facts.

3. Clue Master — builds the clue ledger and the evidence plan from ordinary household items and printables, sets the release schedule across the acts, registers the red herrings, and runs the checks that gate production — above all the solvability check: the solution must be deducible from the clues actually released to players, alone.

4. Host Director (Event Producer) — invitations, casting from your own notes about your guests, the run-of-show with dinner pacing, the voting and awards round, production of the event package — gated on every check passing, including a completeness pass for every player guide — and the dropout protocol if a guest cancels.

Every guest gets a full player guide — a leaflet, not a summary card. Each guide carries a detailed story briefing told from inside the character (their world, their evening, what is at stake for them), what they know and what they do not know, their secret and objectives, and playing material for every act: dialogue propositions — lines to open with, press with, deflect with, reveal with — and actions to perform, with timing cues for when to raise what they noticed. A complete costume guide (garments, colours, hair and make-up, signature props, where to source them from your own wardrobe) comes with a costume image where your platform genuinely generates images in that session, or a detailed visual spec and a ready-to-paste image prompt where it does not — labelled as a prompt, never passed off as a picture. And the guides are held to it: a named completeness check counts the sections, the briefing's depth, and the propositions in every act, and a guide that falls short is finished before it counts as built. "Just a few lines each" is a request this kit declines — a guest handed a stub cannot play their part.

One Case File, no contradictions. All four agents work from the same shared Case File — victim, suspects, timeline, alibi matrix, clue ledger, and solution exist once, in one place. No booklet, card, invitation, or guide page may contradict it. If you change an established fact, the change forces a full consistency re-check of everything built from that fact before anything depending on it is final — nothing is quietly patched, because a quiet patch is how a solvable mystery stops being solvable.

You can play: the host-blind design. In most home-built mysteries the host reads the solution to run the night, and spends their own party outside the game. Not here. Every artifact is classified HOST-SAFE or SEALED before it is produced: the solution, the true timeline, and the murderer's identity live only in sealed artifacts — the Case File's sealed zone and the sealed reveal pack — and the reveal is opened blind, in front of everyone, at the accusation moment. A spoiler sweep checks every host-safe page for solution-bearing facts and for formatting tells before the package is final.

Two ways to work — your choice, per session. Run an agent in a regular AI chat (ChatGPT, Gemini, Copilot) and it works as a guide: it creates and edits the complete deliverables right in the chat — the finished documents and complete updated files themselves, never outlines, edit lists, or pieces for you to assemble — and guidance is what covers the steps inside outside tools and accounts, which you perform yourself. Run the same file on an agentic AI — Muse, Claude Code, Devin, and similar — and it can do more: if you want, it walks you through connecting your documents, design tools, or email and then carries out the steps itself, showing you each consequential action for your approval first, verifying as it goes. Same agents, same rules — what changes is what your platform lets them do. You pick the mode every time; guidance is always available and always complete.

What "produced" means — stated plainly, because it differs by mode. In Connected mode, where a tool like Google Docs or Canva is genuinely connected, the Host Director assembles the actual event package files — per-character booklets, the host Event Guide, printable clue and evidence cards, invitations, and the sealed reveal pack — and verifies each file exists before reporting it produced. In Guide mode, you receive the same package as print-ready content plus layout specifications and step-by-step assembly instructions — and the images themselves, generated right in the chat wherever your session can generate images; an image prompt appears only where it genuinely cannot, labelled as that session's limit, never passed off as a picture. A plain chat does not produce a finished illustrated PDF, and no agent in this kit will ever describe Guide-mode output as one.

Not sure what your AI can actually do? Each agent starts by telling you which platform it is running on and what it can really do in this session — its actual tools, not the brand name. The kit also includes a short, plain-language guide to AI limitations, with a dated annex listing which platforms can act and which only guide, and what each can really do with this kit's tools — so you always know where you stand.

Audio, honestly. The reveal always ships as a read-aloud script you perform cold, in front of everyone, without having read it beforehand. A generated audio version — of the reveal, or of a prop message — is included only where your platform genuinely has audio capability in that session; where it doesn't, you get the script plus simple phone-recording instructions instead. No document in this kit states that an audio file exists when only a script was produced.

A realistic first session: with the Plot Architect you answer its interview one question at a time and reach the first milestone in about 10 minutes: a calibrated Case File, a host-safe summary of your case's structure (setting, cast size, act structure, clue counts per act, difficulty, recorded tone — no solution fact), your invitations as the first produced artifact in the form your mode honestly delivers, and the seal demonstrated. That is the first session — not a finished party. The characters, clues, checks, and the rest of production come step by step after that, and this listing won't pretend building a coherent mystery is a 10-minute job.

The worked example: the materials walk one complete example end to end — "The Last Tonic", a 1920s jazz-club dinner mystery for 8 suspects — so you can see every format, check, and artifact before you build your own. Your party gets a new, original mystery of its own; the example is there to show the shapes, in both modes, host-safe and sealed sides side by side.

What's inside:

- 4 agent files, each in .md AND .txt format (identical content — one of them always works on your device, phone included)

- My Case File and Party Profile — fill it in once (agents help you), paste it into any new chat, skip the interviews; it is the kit's single source of truth: your party facts and calibration, the cast table and public timeline, the check and change records, and the sealed zone you do not read

- Prompt Sheet — the exact phrases to say to each agent, for every job, including the approval phrases and the seal phrases

- AI Limitations Disclaimer (.md + .txt) — what AI can and cannot do, in plain language, with this kit's tools as the examples

- A step-by-step README that gets you running in minutes

- Troubleshooting & FAQ — an accidental peek, a dropped-out guest, a chat that cannot make PDFs: the known rough edges and their fixes

- Changelog — what changed in this version

Who it's for:

- Hosts running a mystery party at home — typically 6–14 guests, often over dinner — who want an original mystery built for their actual guest count, setting, and tone

- Hosts who want to play in their own party: guessing, suspecting, and being surprised alongside their guests instead of running a game whose ending they already know

- Anyone who has been burned by a generic AI mystery that contradicted itself by the second act — a character in two rooms at once, a clue that proved something a later clue un-proved, a solution that needed a fact nobody was ever given

Who it's NOT for:

- Anyone looking for a boxed party kit shipped in the post, or software that runs the night unattended — this kit is instruction files for your AI, plus what you print and seal yourself

- Anyone who wants to know the murderer in advance and host from inside the secret — the design's whole point is that you don't (a host who insists is walked through what that changes, in plain terms, before anything proceeds)

- Anyone who wants characters based on real people beyond the traits they themselves supply — every character in this kit is fictional

What you need:

- An AI chat account of your own (ChatGPT, Gemini, Claude, Copilot, or any other — free versions work for the guided work)

- Your party's facts — a guest count, a date, the rooms in play, and your own notes about your guests for casting; the agents use only what you supply about a guest, and infer nothing

- A way to print (your own printer or a print shop), ordinary envelopes or folders for the packs and the sealed reveal, and ordinary household items for props — every clue-bearing prop also exists as a printable paper version

Honest limits:

Each agent is a text instruction file for your existing AI — it is not a trained AI model, not a separate app, and not a boxed game. What it can do depends on where you run it: in a chat-only AI it guides and drafts — creating and editing the complete deliverables right in the chat — and you perform only the actions inside your outside tools and accounts; on an agentic AI it can act through the tools you choose to connect — always showing you the exact action first, always with your approval before anything is sent, and never unattended. Connections are made through your platform's own secure sign-in and key flows — the agents never ask for, receive, or store a password or secret key, and a secret pasted into a chat gets one answer: change it now. The seal's strength differs by mode, and the kit says so rather than blurring it: in Connected mode the seal is structural — sealed content is written directly to your folder and never displayed in the chat; in Guide mode the seal is a stated discipline — sealed content arrives inside a clearly fenced block with save-without-reading instructions, and the design never requires you to read it, but a plain chat cannot lock you out of its own conversation. Mode honesty holds throughout: Connected mode assembles actual PDFs and images and verifies them; Guide mode delivers print-ready content, layout specs, and assembly steps — plus the images themselves wherever the session can generate them, with an image prompt only as the labelled fallback where it cannot — and calls them exactly that. The reveal script is always text; audio exists only where the session's platform genuinely produced and verified it. Tone is calibration, not drift: you record Family (no gore, no adult themes) or Classic (tasteful, still non-graphic) at the start, and every artifact respects it — anything beyond it is proposed and re-confirmed, never introduced silently. All characters are fictional; real-guest traits enter a character only as you supply them, and only as casting and colour. The agents' worked example, "The Last Tonic", is an illustration created for this kit to show the formats end to end — not a customer story. You check every output before you rely on it.

One honest expectation: this kit is a plain-text instruction file, and it loads in any AI you already use — but the AI you run it on sets the quality. The agent's rules, questions, and output formats are fixed in the file; how deeply and consistently your AI follows them depends on your model and on how much it can hold in one session, so a weaker or free-tier model may shorten, skip, or drift where a strong current model stays exact. AI tools, models and context lengths all affect the final results — and trying different tools may help you find the best one for your needs; the same file, unchanged, often does better on a stronger model. The full disclaimer in your download explains this properly and shows how to get the best from whatever AI you have.

Price: $19 — Standalone: all 4 agents, everything included.

Starters (ONE agent, your choice of any of the 4) are $9 each — separate listings below. Starter owners complete the set with the Upgrade, $10 — the difference.