Agent prompts are selectable workflow profiles for CodeAlta sessions. They tell the agent how to work: implement directly, plan first, review only, coordinate child sessions, keep visible notes, or follow a project-specific release workflow.
CodeAlta separates prompt-related instructions into a few layers:
| Layer | Files | Use it for |
|---|---|---|
| Agent prompt | prompts/agents/<id>.prompt.md |
Session workflow, mode behavior, delegation style, review rules, and when to ask the user. |
| System prompt | prompts/system/<id>.system-prompt.md |
Stable host/agent invariants shared by one or more agent prompts. |
Most customization belongs in an agent prompt. Change a system prompt only when you really need to change stable host-level instructions across workflows.
The active agent prompt is included when a session starts or resumes. Its optional system property chooses which system prompt id to use; when omitted, CodeAlta uses default. Optional composition fields in the agent prompt frontmatter can also turn generated context sections on or off for that workflow.
CodeAlta ships two built-in agent prompts:
| Mode | Prompt id | Best for | What it does |
|---|---|---|---|
| Default | default |
Normal implementation/build work | Inspects relevant files, edits when implementation is feasible, verifies, self-reviews diffs, reports concrete outcomes, and executes approved plan files. |
| Plan | plan |
Larger or riskier tasks where you want a plan before edits | Researches read-only, writes an implementation-ready Markdown plan under .alta/plans/, asks for review, and can hand off to Default when approved. |
Plan mode is prompt-enforced behavior, not a hard filesystem sandbox. The built-in Plan prompt tells the agent not to mutate project files except for the plan under .alta/plans/, but normal host permissions still exist. Use normal review practices for sensitive work.
Typical prompts:
Use Plan mode for this task. Inspect the relevant docs and code,
then write an implementation-ready plan before any source edits.
Execute the approved plan at .alta/plans/2026-06-04-workflow-docs.md.
If git is active and .alta/plans/ is not ignored, the built-in prompts treat plan files as repository artifacts: Plan keeps the saved plan current, and Default keeps it synchronized with implementation progress.
Prompt resources are layered in this order:
~/.alta/prompts/.<project>/.alta/prompts/.Each root has the same layout:
prompts/
system/
default.system-prompt.md
my-custom-system.system-prompt.md
agents/
default.prompt.md
plan.prompt.md
reviewer.prompt.md
If multiple roots contain the same prompt or system id, the later source overrides the earlier one: project overrides global, and global overrides built-in. This lets you keep a global reviewer prompt while giving one repository a stricter project-local reviewer.prompt.md.
Prompt ids come from file names. For example, reviewer.prompt.md creates or overrides the agent prompt id reviewer.
Use a custom prompt when you repeat the same workflow often enough that it deserves a named mode.
Ctrl+G Ctrl+H or /prompt.reviewer, triage, or release.Ctrl+T / /next_prompt.A minimal prompt file looks like this:
---
name: Reviewer
description: Review-only workflow that reports risks before code changes.
---
You are the active CodeAlta review agent for this session.
Review the requested change without editing files. Inspect relevant code,
tests, and docs. Report findings first, ordered by severity, with file and
line references when possible. If no issues are found, say so and mention any
verification gaps.
description is optional but should be decision-useful: CodeAlta surfaces effective agent prompt descriptions in generated model context so agents can discover available workflows and switch or delegate appropriately.
Keep descriptions short and specific, for example “Plan-first release workflow that writes a checklist and asks before publishing.” Avoid long policy text in descriptions; put detailed behavior in the body.
Agent prompt frontmatter is the user-facing entry point for prompt composition. The file name selects the agent prompt id, the prompt's system field selects the system prompt id, and optional boolean fields override which generated context sections CodeAlta includes.
| Field | Required? | Use it for |
|---|---|---|
name |
Yes | Display name in the prompt selector. |
description |
No | Short decision-useful summary shown to agents and users. |
system |
No | System prompt id from prompts/system/<id>.system-prompt.md; defaults to default when omitted. |
skills |
No | Include available/active skill guidance. |
project_context |
No | Include repository instruction files such as AGENTS.md. |
runtime_context |
No | Include current date, platform, working directory, project root, and session kind. |
tool_guidance |
No | Include generated host-tool guidance and available agent-prompt discovery. |
Do not restate defaults in every prompt. When the boolean fields are omitted, CodeAlta uses its normal defaults, which currently include all generated sections when content is available. Add a boolean only when a workflow intentionally needs a different composition.
For example, a narrow review prompt can keep the normal system prompt while disabling skill and tool-discovery guidance:
---
name: Minimal Reviewer
description: Review-only workflow with a smaller generated context.
skills: false
tool_guidance: false
---
Review the requested change without editing files. Report findings first,
ordered by severity, with file and line references when possible.
Good agent prompt bodies are concrete enough to shape behavior but small enough to stay understandable:
For advanced workflows, the body can mention CodeAlta live-tool capabilities the model may use, such as session delegation, sticky notes, reminders, or structured asks. Users still prompt for the outcome; the model invokes live-tool commands internally when the selected provider/session supports them. See Advanced Agent Workflows for recipes and command-group coverage.
Use the Agent: selector below the prompt editor to choose the prompt for the current draft/session. Built-in prompts appear first, followed by global prompts and project prompts.
The prompt manager lists built-in, global, and project prompts, shows shadowed overrides, and lets you create, edit, save, or delete global/project prompt files. Built-in prompt and system prompt files are visible for inspection but read-only; create a global or project file with the same id to override one.
Agents can also use the prompt and session live-tool command groups internally when you ask for prompt automation. For example, you can ask:
List the available agent prompts for this project and recommend which
one fits a read-only security review.
Create a project agent prompt named release-checklist that plans release
steps, keeps notes visible, and asks before publishing.
Switch this session to the reviewer prompt for the next turn.
Those are user prompts, not terminal commands. CodeAlta-managed agents translate them into the appropriate prompt/session operations when available.
A project prompt can specialize a global workflow without changing other repositories. Common examples:
plan.prompt.md that adds repository-specific plan-file sections;reviewer.prompt.md that includes local test commands and review gates;release.prompt.md that knows the repository's packaging and publishing checklist;triage.prompt.md that asks child sessions to inspect logs, issues, and recent session history.Project-local prompts are part of the repository workspace. Review prompt changes like code changes: a prompt can change how agents decide, ask, delegate, and mutate files.
System prompt files carry host-level behavior and should be short, stable, and explicit. Agent prompts are better for workflow-specific session behavior.
Use the agent prompt system frontmatter field when a workflow needs a custom system prompt:
---
name: Release Coordinator
description: Release workflow using the team release system prompt.
system: team-release
---
Coordinate the release checklist, verify packaging steps, and ask before
publishing.
Keep composition overrides close to the agent prompt that needs them. Disabling generated context such as tool guidance can make advanced prompt workflows less discoverable to the agent, so prefer the default generated sections unless a specific workflow has a clear reason to remove them.