Personas
Sources
Docs: Ponder Primer, Powers. Web App: Open Persona below the Ponder input. Workspace managers can edit a Persona’s Prompt and Profiles.Story
A Persona is a reusable approach to work. Its prompt tells Ponder how to respond, and its profiles choose the providers and text settings. A Persona carries no tools or tool permissions: tools arrive only through Powers, and the Workspace defaults decide which ones need approval. The same Persona can be used across many Problem threads without changing anyone’s Workspace access.Scope
Each Persona has exactly Simple, Standard, and Strong profiles. These are saved configurations, not separate prompts or hidden effort settings. New Personas start with the Workspace’s default chat model for each profile, which is Parable Portal unless the Workspace sets its own; change their rankings or settings when you want them to behave differently. The composer starts with text. Each ranked entry opens one Configuration editor for its complete saved configuration, including structured output and media settings. Provider availability depends on configured adapters and the selected model. Saving a valid configuration does not make an unavailable operation executable. Speech and transcription settings can be saved, but the default OpenRouter route refuses execution: its audio endpoints do not establish the privacy and provider restrictions Ponder requires. An available SDK adapter does not by itself make an approved conversation route available.Structure
Choose an approach
Persona: Automatic uses the Coordinator. Open the picker and choose another Persona when its responsibility better matches the next request; the chosen row shows a selected radio button on its left, and a row’s Edit action appears when you hover or focus it:
Changing Persona affects the next Message. It does not reassign the Problem or
rewrite earlier responses. Selected Problems, Personas, and Powers display their
names; category prefixes remain only on the Automatic choices.
A shared method, different responsibilities
Coordinator carries the core Parable model in its prompt so Automatic can reason from the same vocabulary before retrieving narrower Powers. It defines Workspace and Parable context; Problems, Messages, Personas, profiles, Powers, Processes, Proofs, and Pings; Pieces and the Page/Package/Pane hierarchy; Perceptions, Plots, Palettes, Policies, Preferences, scopes, and handles; participant responsibilities and active Participation Modes; and the Personal, Changes, Version, publication, Deployment, and Report lifecycle. That default context keeps distinctions that change the result explicit: definition Versions are not Report data readings, saved settings are not app definitions, temporary Participation Mode does not grant responsibility, and Workspace membership is not a Parable RACI seat. The Coordinator still acts directly, discovers relevant Powers and tools, preserves request context, and does the smallest useful work. It reports the completed outcome and why it acted, without surrounding progress commentary. The other eight built-in approaches share a method. They start with the people and their work, including configuration, administration, and support. They shape the Page/Pane experience and its native Perception, settings, and Policies before writing SQL. They retrieve the relevant Powers as needed rather than adding every instruction to each response. When an unresolved decision would change the work, a Persona can put focused questions in the conversation, with possibilities and their consequences. It waits for an answer before continuing the dependent work. Routine reversible choices do not need a new approval conversation. For visual work, the method calls for inspecting actual Pane previews and, where useful, comparing authored options. That includes small defects such as one-pixel seams, doubled borders, misalignment, clipped controls, wasted space, and unintended gutters. A captured picture proves what was rendered; it does not prove keyboard use, a successful save, or live data execution. Those need separate interaction and data checks. The same care applies to the contracts behind the screen: names, types, errors, and operations should be understandable to the people and tools that use them. Behavior checks should expose a real missing contract or defect and then verify its correction. Compilation, screenshots, and tests each provide different evidence; none substitutes for the others. Council considers several perspectives within one response. Its name does not mean separate agents or experts were spawned. Critic stays focused on the particular claim under review; Council combines the important findings into a recommendation for the whole application. The shared guidance is included in those eight Personas’ saved Prompts; Coordinator uses its concise prompt on its own. Platform updates can refresh an unchanged built-in prompt. Workspace edits to that prompt, custom Personas, provider profiles and rankings, and completed response history are preserved.Write one shared prompt
Open the Persona’s Edit action. The editor holds the Persona’s Name, its Handle (fixed once saved) and its icon, then two sections: Prompt and Profiles. Choose Prompt and write the Markdown instructions that should apply to all three profiles. The prompt is supplied as instructions when Ponder runs; it is not printed as an introductory chat message. Persona and Power editors share one joined icon, name, and handle control. Markdown inputs omit line numbers. Name the audience, desired outcome, evidence, and boundaries. Use a Power for a repeatable task rather than putting every possible method into one Persona.Rank providers
In Profiles, choose Simple, Standard, or Strong. Each profile keeps its own ordered model list:- drag any ranked row by its handle to change the order, or pick the handle up with Space and move it with the arrow keys;
- switch a non-Portal entry off without losing its configuration, or remove it;
- search the catalog below the list and choose Add on a row (the action shows when you hover or focus the row); and
- open a row to edit that entry’s complete Configuration document.
Inspect and edit the complete configuration
Open a ranked model to edit its Configuration. The same editor is used for Parable Portal and other models. It shows every supported operation block and its declared fields, with completion, hover help, and validation. A field you saved shows your value, an authored null included. A field the SDK declares a default for shows that default marked# default; leave it and
nothing is saved, change it and your value is. Every other field is null:
unset, so the runtime or provider default applies. For a nullable setting
without a default, an empty value marked # unset means no override. Opening
the editor never saves defaults or modifies the Persona. Provider-specific
options and output schemas retain their authored nulls.
When you leave text.maxOutputTokens unset, Ponder uses the selected model’s
known maximum response allowance, reduced when necessary to leave room for the
estimated conversation context. Set a smaller value when you want a shorter
allowance. If the catalog does not report a maximum, including Portal’s automatic
model selection, the provider’s default applies. The model’s available context
and provider limits still apply; the allowance does not force a long answer.
This is the saved configuration, not a resolved provider request. Runtime and
provider defaults still apply; credentials, instructions, authorized tools, and
server policy remain under Ponder’s control. Portal chooses the underlying
model automatically. Image, audio, video, and other operations need a configured
provider that supports them.
The editor is available in shared Persona Profiles and private Profile
settings. It preserves neighboring operation blocks and comments. Invalid edits
remain available for correction while the dialog is open and block saving or
updating.
Make private changes in chat
Choose the profile in the input footer, then open Profile settings (shown when theponder_profile_settings feature flag is on). This
is the same ranked-model editor for the selected Persona and profile. A badge
beside the title says whether you are looking at the Persona or your
Private copy. Edits stay in the dialog until you choose what to do with them:
- Save for Me keeps them as a private copy for you in this Workspace, stored in this browser. It does not edit the shared Persona or appear for another person.
- Update Persona saves that profile to the shared Persona, if your current access permits it.
- Cancel discards the edits.
- Reset, shown while a private copy exists, discards that copy.
Schema
Native configuration
The YAML document belongs to one ranked entry. Text options live in itstext
block; leave an option unset to use the runtime or provider default:
providerOptions.openrouter.reasoning. The generic top-level reasoning option
is not available with the current adapter; it is rejected rather than silently
ignored.
The editor provides completion, hover information, and validation. Unknown
native fields, invalid values, or runtime-owned fields block saving. Model
identity for Text comes from the ranking. Other operation blocks select a
configured model using a provider:model-id reference. Credentials, messages,
instructions, and tool definitions remain controlled by Ponder.
The same document contains every configured operation. A structured output
(text.output) shapes the final answer in chat and in a Process run alike;
Process runs are not available yet. Ponder gives the answer as a value that
must validate against the output’s schema, and the reply shows that value as a
JSON code block. One difference: in chat, text the model wrote before it
answered stays in the reply, ahead of the code block, while a finished run’s
reply holds only the code block. When the model gives its answer together with
an operation that needs your approval, the reply waits for your decision and
then finishes with that same answer, whether you approve the operation or not.
An answer that does not validate, or an
answer written as text, goes back to the model with what to fix. After three
failed tries a chat reply fails with PA-CH-057, which asks you to try again
or to simplify the output setting, and a run fails with
PONDER_RUN_STRUCTURED_OUTPUT_INVALID. output: {type: text} asks for plain
text. A saved structured output looks like this:
text, image, video, speech,
transcription, streamTranscription, realtime, translation, embedding,
and reranking. Their availability in a conversation depends on the configured
provider and runtime. In particular, a saved video configuration is not proof
that the provider supports the required zero-data-retention policy. Ponder must
reject an unavailable or incompatible route without weakening that policy.
A configuration can contain at most 16 KiB. Provider-specific options still need
to be supported by the selected model; a valid YAML shape is not a guarantee
that every upstream endpoint supports it.
Security
Tools and approvals come from Powers
A Persona has no settings for Parable tools. Ponder can use a generated tool only while one of the Workspace’s Powers uses it, and the Workspace default for that tool, or else the operation’s own default, decides whether Ponder asks first. Deleting something asks every time by default. Image and video generation are the exception: they come from the Persona’s native configuration, not from Powers. For an Ask request, review the title, consequence, and input summary, then choose Deny once or Allow once. Managers can remember Always or Never as the Workspace default. The policy must save before the pending call continues.Follow the Workspace AI policy
A Workspace’s AI policy decides which AI processors may receive its data. Ponder sends each conversation to OpenRouter with zero data retention and provider data collection denied, and it checks the policy separately for Simple, Standard, and Strong:- A ranked model, or the Parable Portal fallback, answers only when the policy lets its processor serve that profile. Ponder never falls back to a model the policy does not allow.
- When the policy allows no processor for the selected profile, the input says Chat is unavailable under this Workspace’s AI policy and sends nothing. A message already on its way ends with the same explanation, and a Run accepted before the change fails with it instead of running.
- When the policy refuses only some of a profile’s ranked models, a conversation and a Process run both leave those out and use the others in rank order. A model refused after a Process run was accepted no longer fails the run; the run fails only when the policy refuses all of them.
- New Personas rank the Workspace’s default chat models. Changing those defaults leaves existing Personas as they are.
Stability
Portal remains available
Parable Portal is the default text route. Disabling or removing the last usable text entry visibly restores Portal, and server-side resolution enforces the same invariant for saved and private configurations. The one exception is the Workspace’s AI policy: when it switches OpenRouter off for a profile, Portal is unavailable too, and Ponder says so instead of falling back. In a conversation, ranks after the first are OpenRouter’s fallback list for rank 1’s request, not independent configurations. A later rank answers only when rank 1 fails upstream and its provider restrictions match rank 1’s, and it runs under rank 1’s settings: text settings, reasoning, output budget, native media, and context and output limits. The reply’s cost is what OpenRouter charged for the model that answered. A missing eligible zero-data-retention endpoint does not remove the Portal fallback. Every attempt retains the same privacy requirements; fallback never permits data collection or weaker access. In a Process run, each ranked entry is tried in turn under its own settings, never as another entry’s fallback. The run starts with the first entry. When an entry fails upstream, after its own retries, the run continues from where it stopped under the next entry, and keeps that entry for the rest of the run. When every entry has failed, the run tries again from the first entry, and after its last retry it fails withPONDER_RUN_MODEL_UNAVAILABLE. Each entry’s
calls are billed at that entry’s own prices. A structured output is the active
entry’s, and an entry that takes over starts with all three tries.
A provider outage can still prevent a response. Once output or a tool effect has
committed, Ponder does not switch models and replay the work. Approval
continuation retains the accepted model and configuration; changed instructions
or revoked authority require a new Message instead of silently changing the
response being approved.