Ideagents
Private testing · Not submitted or approved

Turn your ideas into products

Explore possibilities, challenge assumptions, and plan your next move with AI.

Ideagents is a ChatGPT plugin with a record-checking service. ChatGPT, as the host, performs all reasoning and generates every viewpoint, and you make each decision. Six changeable profiles or custom roles set the perspectives. Guided sessions end each turn so you can respond. The record and any diagram are updated after decisions, and a stateless Ideation Lab service checks the current record when the plugin's tools are called. The output is a decision record with a scoped next step; Ideagents does not build or deploy products.

View examplesSet up your sessionHow it works

Set up your session

This builds a prompt to paste; it does not save preferences. The controls generate text on this page. They do not run Ideagents, create an account, save settings or send anything. The settings take effect only when you paste the prompt into a ChatGPT account where the private-testing plugin is installed, and they apply within that conversation.

Sets the skill and opening line of the prompt.

Shape an app or product around a small usable journey.

Separate roles with commas, for example: security, accessibility reviewer. Roles are analytical perspectives, not real people.

Mode

Guided gives one perspective per turn. Quick gives one complete response.

Outputs
Update cadence

Applies while documentation or diagrams are on.

Starting prompt

Use the Ideagents develop-idea skill to develop this idea: [describe your idea and its real constraints]

Use the App Builder profile.
Work in guided mode: give one short perspective, then end your turn so I can respond.
Keep a working record, and update it after each material decision or change.
Update a diagram after each material decision or change when a diagram would clarify it.
I may change these settings at any turn.

Replace the bracketed text before sending. Copy setup places the text on your clipboard only.

Change settings in ordinary chat

Send an ordinary message at any turn; no command syntax is required. The descriptions below state the behavior the skills instruct. Responses are generated by the host and can vary.

“Add security”
Adds the perspective, states the resulting setting and preserves your other settings.
“Pause”
Identifies the checkpoint and ends the assistant's turn. “Continue” resumes from it.
“Skip to handoff”
Produces that stage and marks material checks not performed as unreviewed.
“Turn diagrams off”
Stops automatic diagrams. A prior diagram is marked stale if a later change affects it.

Example sessions for three workflows

Illustrative example, not a live AI response. These samples were written in advance to show the format. No model runs on this website, and real sessions will differ.

Samples condense a session. In guided mode, each perspective, finding, brief section or unresolved choice arrives in a separate turn, with your reply between turns. A direct request for a deliverable may run as one response.

In real use, ChatGPT generates every perspective. Generated user viewpoints are illustrative: they are not interviews, research findings or customer statements. Verify them with real users.

Copy prompt places the text on your clipboard only. The prompt invokes Ideagents only in a ChatGPT account where the private-testing plugin is installed. This website does not install or run Ideagents.

Develop an idea

  1. 1 · Your idea and constraints

    Use Ideagents to develop an idea: a shared grocery list app for roommates that splits costs automatically. Constraints: I'm building alone in two weeks, mobile web only, and I won't handle payments.

  2. 2 · Relevant perspectives

    Product
    The underlying need may be settling costs rather than maintaining a list. Determine which need comes first.
    Design
    Cost discussions between roommates can be sensitive. Show who added an item by default, not who owes whom.
    Engineering
    Record balances only. Moving money adds scope that two weeks cannot cover.
    User (illustrative, not an interview)
    A roommate might use this if adding an item takes one tap and no cost notices arrive late at night. Verify with real roommates.

  3. 3 · Challenge and trade-offs

    Automatic per-item splitting could add the friction the app is intended to remove.

    • Split every itemAlways current, but each purchase requires agreement.
    • Monthly summarySimpler to build, but balances lag by up to a month.

  4. 4 · Your direction

    “Monthly summary. No payments.”

  5. 5 · Decision record and next-step brief

    Decided
    Monthly summary. No payments.
    Not chosen
    Per-item splitting.
    Proposed, not decided
    Item attribution on the shared list (design suggestion). Revisit the summary interval if testers ask to settle up mid-month.
    Proposed next step
    Build one shared-list prototype in the two-week build period, then test it with three households and review the monthly summary after the first month. This timeline extends beyond the build and is proposed, not decided. Success measure: all three households still adding items when the first monthly summary is reviewed.

Challenge a proposal

  1. 1 · Your proposal and constraints

    Use Ideagents to challenge this proposal: replace our onboarding tour with an AI chat assistant. Context: 41% of new users finish onboarding. We have one quarter and four engineers.

  2. 2 · Relevant perspectives

    Product
    Identify where users leave first. A chat assistant helps only if confusion causes the drop-off.
    Design
    An empty chat box requires new users to know what to ask when they know the least.
    Engineering
    Adds ongoing inference cost, evaluation work and failure modes that are difficult to test.
    User (illustrative, not an interview)
    A new user who skips tours might also skip a chatbot and prefer immediate useful output. Verify with real new users.

  3. 3 · Challenge and trade-offs

    The proposal assumes the tour format is the problem. The supplied context does not establish that.

    • Replace the tourA larger commitment that is slower to evaluate and costs more to operate.
    • Fix the drop-off pointsNarrow and measurable, and likely to ship sooner.

  4. 4 · Your direction

    “Fix the drop-off first. Keep the assistant as a later experiment.”

  5. 5 · Decision record and next-step brief

    Decided
    Address onboarding drop-off first.
    Deferred
    AI chat assistant, retained as a later experiment.
    Proposed, not decided
    Revisit the assistant if completion stays flat after the redesign.
    Proposed next step
    Collect one week of step-level data, redesign the two steps with the highest loss, then compare for three weeks. One PM, one engineer.

Prepare a handoff

  1. 1 · The decision and constraints

    Use Ideagents to prepare a handoff: we've agreed to add offline mode to our field inspection app, and engineering needs a brief. Constraints: Android tablets, shifts of up to eight hours without signal, existing REST API.

  2. 2 · Draft brief sections

    Scope
    Proposed: full offline, the broadest reading of the agreed offline mode. Whether version one should be narrower is unresolved.
    Sync status
    Proposed: show sync status on every record, not one global spinner.
    Conflict policy
    Proposed for decision before build: last write wins per field, with append-only photos.

  3. 3 · Unresolved choice

    Which offline scope should version one support? The conflict policy and per-device storage limit remain separate open items.

    • Full offlineEverything works, but sync and storage risk increase.
    • Own inspections onlyCovers the daily job and is likely to ship sooner.

  4. 4 · Your direction

    “Own inspections only for version one.”

  5. 5 · Decision record and next-step brief

    Decided
    Version one supports offline create and edit of your own inspections.
    Out of scope
    Offline browsing or search of other inspectors' records.
    Proposed, not decided
    Conflict policy. The draft suggests last write wins per field with append-only photos. It remains unresolved until you decide it. Sync status. The draft suggests showing status on every record, not one global spinner. This is not a decision.
    Proposed acceptance
    Not decided. An inspection with 40 photos completes in airplane mode, syncs after reconnecting, survives an app restart, and loses no photos on sync conflicts.
    Open questions
    Final conflict policy; storage limit per device.
    Proposed next step
    Decide the conflict policy and the per-device storage limit, then issue this brief to engineering.

Workflow from idea to decision record

  1. State the idea

    Describe it with its real constraints: time, team, budget and exclusions. Paste your settings, or accept a suggested profile you can change.

  2. Review perspectives

    While you develop an idea, ChatGPT generates short viewpoints from the active profile or your custom roles. In guided mode, it gives one and ends its turn. Any user viewpoint is illustrative.

  3. Review the challenge

    One concise challenge and the trade-offs that matter most.

  4. Choose a direction

    You decide. Ideagents does not vote or choose for you. A correction updates the affected conclusions before the session continues.

  5. Receive the record

    A decision record that separates decisions from proposals, with a scoped brief for the next step. Any diagram carries the same statuses.

  • Turns and stopping

    In guided mode, the assistant gives one short perspective when you develop an idea, one perspective or finding when you challenge it, or one brief section or unresolved choice during a handoff. It then ends its message, which leaves the next turn to you. It continues only after you reply, and “Pause” ends at an identified checkpoint.

    To stop a response while it is being generated, use the Stop control your host provides. Ideagents cannot cancel text in progress or receive input during a response. It adds no voice feature and makes no claim about host voice controls.

  • Working record and diagram

    The same assistant keeps one working record and, when useful, a diagram. With on-decision updates, it revises them in its reply after a material choice, correction or scope change. No update runs in the background or between turns. An output that is off or awaiting your request is marked stale when a change affects it.

    When the assistant calls the Ideagents tools, the record-checking service checks the revision and derives matching Markdown, safe Mermaid source and a portable JSON capsule with a content hash. The current record stays in the host conversation or in a capsule you copy.

    Native documents or canvases are used only where the host provides and authorizes them. Otherwise, in a guided chat, record revisions appear as short change notes against the last complete text, and the complete text appears at first issue, synthesis, export or on your request. Guidance for each diagram revision calls for copyable Mermaid source with a readable outline or table alongside it, even if the host renders Mermaid, because some diagram labels can appear as numeric codes instead of words.

Illustrative reconciliation after your direction. Static example written in advance from the Prepare a handoff sample. It is not generated output or a product screen, and its statuses do not indicate that any system was built or verified. The next-step path in this figure is a host-drawn plan draft, not current service output; the backend generates a separate checked decision map, which does not draw this sequence.

Your direction: “Own inspections only for version one.” Result: R1 and the earlier plan draft listed full offline as proposed. R2 and the revised plan draft apply the same change in the same reply.

Working record · R2

  • D1 Offline scope: own inspections only User-selected
  • D2 Conflict policy: last write wins per field Proposed
  • Next Decide the D2 conflict policy and the per-device storage limit, then issue the engineering brief Proposed
  • Alternative Full offline Deferred

Illustrative host-drawn plan draft · reflects R2

  1. D1 Own inspections offline User-selected
  2. D2 Conflict policy Proposed
  3. Decide per-device storage limit Proposed
  4. Engineering brief Proposed

Outside version one: full offline Deferred

Scope and limitations

  • Reasoning runs in ChatGPT

    Structured instructions guide the ChatGPT assistant through three workflows. ChatGPT performs all reasoning on the host model, and Ideagents uses no separate model API. Host settings govern data retention, limits, and model, effort and voice availability. Ideagents cannot change the host model or effort.

  • Record-checking service

    When the assistant calls the Ideagents tools, a stateless Ideation Lab service receives a bounded current decision record and your explicit changes. It returns the defaults for the six profiles, checks that each revision applies as one complete change, checks statuses, superseded items and freshness, and derives matching Markdown, safe Mermaid source and a portable JSON capsule with a content hash.

    The service has no database and does not log application content or intentionally retain it. It makes no outbound network requests or model calls and runs no background jobs. Full transcripts should not be sent to it. See Privacy.

  • Current record and capsule

    The current record stays in the host conversation or in a capsule you copy. To copy a capsule, use the Copy control on its code block if the host provides one, then paste the complete code block intact. A content hash does not prove your approval, facts, history or identity. Outputs remain drafts that require verification.

  • Perspectives are generated

    ChatGPT generates every viewpoint. They are not independent experts, real users, interviews or professional advice. Outputs are drafts that require verification.

  • No separate account or key

    No separate model key, model account or Ideagents account. Your ChatGPT plan's usage limits apply.

  • This website runs no model

    The examples were written in advance. This static site has no analytics or model calls. The session setup controls only generate text on this page; they submit and store nothing. Its hosting may process ordinary request logs.

  • Availability

    Private testing only. Ideagents has not been submitted for review, is not approved and cannot be installed publicly. There is no waitlist. Questions: see Support.