Create Profile
Guide the user through creating a Talon persona/profile interactively. This is a conversational wizard. Ask one question at a time, infer sensible defaults, show the draft, refine it, then scaffold it.
Core principles
- One question at a time. Keep the interaction tight.
- Use
talonctl add-persona. Do not edittalond.yamldirectly. - Confirm before mutating files. Show the inferred summary first.
- Write/send defaults are gated. Put write/send capabilities in
requireApprovalunless the user explicitly wants broader access. - Preview before reload. The operator can inspect and edit generated files
before running
talonctl reload.
Architectural note: this workflow relies on shell execution of
talonctl, which bypasses Talon's host-tool capability model. Treat it as an operator action.
Phase 1: Gather intent
Ask: "What should this persona do?"
If needed, follow up one question at a time to learn:
- Main purpose
- Whether it needs file read/write, messaging, HTTP/API access, memory, subagent delegation, scheduling, or database access
- Whether quality or speed/cost matters more
Phase 2: Infer defaults
Build a first draft from the user's description.
Name
Slugify the purpose into a short persona name.
Examples:
code review->code-reviewerPR security audit->security-auditor
Provider and model
- Default provider: inherit from the current operator context if one is already implied; otherwise ask.
- Infer a tier first, then map it to a provider-specific model.
- If unclear, default to
balanced.
| Tier | Use when intent suggests | Claude provider | Gemini provider |
|---|---|---|---|
strong |
review, audit, analyze, plan, architect | claude-opus-4-6 |
gemini-2.5-pro |
balanced |
summarize, draft, write, general assistance | claude-sonnet-4-6 |
gemini-2.5-flash |
fast |
classify, tag, triage, quick, simple | claude-haiku-4-5 |
gemini-2.5-flash |
backgroundProvider (optional) — provider used when this persona spawns a background agent. Resolution order: explicit tool arg → backgroundProvider → persona's foreground provider (only if it is also enabled under backgroundAgent.providers) → backgroundAgent.defaultProvider. Set when the foreground provider is unsuitable for background work (e.g. local Ollama running short-context models). Must be enabled under backgroundAgent.providers — the daemon refuses to start otherwise.
backgroundModel (optional) — model passed to the background provider. Paired with backgroundProvider; the daemon refuses to start if backgroundModel is set without backgroundProvider. Useful for sending background work to a more capable model than the foreground.
Capabilities
Use these exact labels.
Host tools
| Intent hints | allow |
requireApproval |
|---|---|---|
| memory, remember, context, knowledge | memory.access:* |
|
| message, notify, send, communicate | channel.send:* |
|
| search, fetch, http, api | net.http |
|
| queue async work, background jobs | subagent.background |
|
| ask another agent inline, specialist consultation | subagent.invoke |
|
| delegate to / hand work to another persona | persona.send:<persona-name> (one label per allowed target; persona.send:* for any) |
|
| image, photo, screenshot, what does this show | image.read:thread |
|
| schedule, cron, recurring tasks | schedule.manage |
|
| query, database, SQL | db.query |
image.read:thread exposes image_lookup: text-only discovery of the thread's
Images and saved Findings. It never returns pixels, never calls a model, and needs
no credentials, so it is safe to grant on its own.
If the persona should actually answer what an image shows, capability labels are
not enough — it also needs subagent.invoke and vision listed under the
persona's subagents key in talond.yaml. There is no talonctl add-persona
flag for sub-agent assignment, so edit the config (or note it in the summary for
the user to add afterwards). Pixels go only to the vision sub-agent; the persona
itself remains text-only. Warn that uploaded GPS/EXIF and Findings persist after
the image file expires (default 24h) and can appear in replies.
Provider-native
| Intent hints | allow |
requireApproval |
|---|---|---|
| code, review, file, read, analyze source | fs.read:* |
|
| write, edit, fix, refactor, implement | fs.read:* |
fs.write:workspace |
Rules:
- Write/send capabilities default to
requireApproval. - Users can later promote items to
allow. - If no capability is clearly needed, leave the lists empty.
- Note:
requireApprovalcurrently records configuration intent only. Runtime approval enforcement is not yet implemented — tools listed there are still accessible. Mention this when presenting the summary.
Skills
Default to no skills unless the user explicitly asks for one.
Description
Generate a 1-2 sentence description of the persona's purpose and strengths.
This is written as YAML frontmatter in system.md and exposed via the
background_agent profiles action so that agents can discover available
profiles at runtime. It is also what other personas see when choosing a
persona_send delegation target (roster in their system prompt and
persona_list), so state what tasks the persona should receive and what it
returns.
Good descriptions are specific enough to differentiate profiles:
- "Senior code reviewer — focused security and performance analysis with PR review expertise"
- "Deep web research — multi-source search, cross-referencing, synthesized answers with citations"
- "Personal butler — formal, anticipatory, handles daily tasks with understated charm"
System prompt
The system prompt file (system.md) uses YAML frontmatter for metadata:
---
description: "<the description from above>"
---
# <name> — System Prompt
<prompt body>
Generate a concise prompt body covering:
- Role and goals
- Working style
- Boundaries
- Any important constraints implied by the capability choices
Phase 3: Present summary
Show a readable summary before doing anything:
Name: code-reviewer
Description: Senior code reviewer — focused security and performance analysis
Provider: claude-code
Model: claude-opus-4-6
Allow: fs.read:*, memory.access:*
Require approval: channel.send:*, fs.write:workspace
Skills: (none)
System prompt: [preview first 3 lines]
Then ask what should change, or whether to create it.
Phase 4: Refinement loop
Handle edits such as:
- "rename it to pr-checker"
- "use the fast one"
- "switch to gemini"
- "add fs.write"
- "move channel.send to allow"
- "show me the full system prompt"
- "create it"
After each change, re-show the summary and wait for approval.
Phase 5: Create and preview
After approval:
- Write the generated prompt to a temporary file.
- Run:
talonctl add-persona --name <name> \
--description "<description>" \
--model <model> \
--provider <provider> \
--capabilities "<comma-separated allow>" \
--require-approval "<comma-separated requireApproval>" \
--skills "<comma-separated skills>" \
--system-prompt-file /tmp/<name>-system.md
- Show the generated
personas/<name>/system.md. - Offer to edit
personas/<name>/system.mdbefore reload. - If the user wants richer tone/personality files, suggest running the
create-personalityskill next and add those files before reload. - When the user confirms the generated files look right, run:
talonctl reload
If talonctl is not on PATH, use the repo-local equivalent the environment
already uses.
After creation
Give short next-step suggestions:
- Use it as a background profile:
background_agent spawn profile="<name>" prompt="..." - Bind it to a chat channel if desired:
talonctl bind --persona <name> --channel <channel> - For richer voice/tone files, run the
create-personalityskill