Augments

Augments are short, stackable behavior modifiers for the assistant. Where a persona defines who the assistant is and how it speaks, an augment defines how it should approach the current task. Pro-only.

Examples

• Code Reviewer — review code for bugs, readability, and edge cases.
• Security Auditor — flag injection, auth issues, and OWASP-style risks.
• Concise — keep answers under three sentences, no preamble.
• Step by Step — break every answer into numbered steps.
• Devil’s Advocate — challenge assumptions, argue the opposite.
• Voice Mode Help — turn this on when the user wants to ask the assistant how Voice Mode itself works. With this active, the assistant knows where to find Voice Mode’s documentation on the user’s machine and will read it before answering questions about features, settings, or behavior. Most useful when paired with the Claude or Codex backends, which can read local files. Turn it off again when going back to normal work — it only takes one slot, but it’s also one slot fewer for the augments shaping the rest of the conversation.

Stacking augments

Multiple augments can be active at once. They compose best when each targets a different axis — for example “Security Auditor” (a domain expert) plus “Concise” (an output format) work together cleanly, because they’re not competing for the same kind of behavior.

Augments that target the same axis tend to conflict — two output-format augments will fight each other, and the result is worse than either alone. The picker shows what each augment does, which makes it easy to keep loadouts coherent.

Augments and personas combined

Voice Mode shares a single budget across personas, characters, and augments — five slots in total. Personas and augments take one slot each; characters take two by default. The picker enforces the limit so the assistant doesn’t get overloaded with conflicting instructions, which degrades reply quality.

A typical loadout might be: one character (or persona) for personality, plus two or three augments for the work the user is doing right now.

Switching augments

Augments can be toggled on and off from the Dashboard’s Augments pane without restarting the assistant. The change applies to the next assistant turn.

Pro tips

• One domain + one format is the cleanest stack. “Code Reviewer” plus “Concise” gives focused, short reviews. “Travel Planner” plus “Step by Step” gives an itinerary the user can actually follow. Two domain augments at once usually fight; two format augments definitely fight.
• Devil’s Advocate is a sanity check. Turn it on briefly when the user is about to commit to a decision they’re not sure about. It pushes back in a way the assistant won’t unprompted.
• The Voice Mode Help augment is short-lived by design. Turn it on when asking how the app works, turn it off again. Leaving it on permanently spends a slot for something the user only needs sometimes.
• Save slots by using fewer augments at once. Five active augments is the maximum, but the cleanest output usually comes from two or three. The model has only so much attention to spread across constraints.
• Custom augments are great for personal context. Drop a markdown file into the augments folder for things like “always write the user’s email signature in this exact form” or “format every reply as a Jira ticket.” Short directives win.

Inspecting the user’s augments

If the user asks “what augments do I have on?” or wants to know what’s available, call the list_augments MCP tool rather than reading augment files directly. It returns active and available augments and the shared slot budget — the right surface for “what’s loaded right now” questions. See tools.md.

There is no activate_augment / deactivate_augment tool yet. If the user wants to turn an augment on or off, point them at the Dashboard’s Augments pane rather than editing augment files directly.