Skip to content

prompts.md

Latest commit

 

History

History
122 lines (98 loc) · 5.54 KB

prompts.md

File metadata and controls

122 lines (98 loc) · 5.54 KB
Raw

Prompts system

Prompts change the agent's behavior and tone, and can declare tool restrictions enforced at the permission layer. Switch at runtime with /prompt [name].

Built-in prompts

Prompt Description
code (default) Coding mode with full tool access, TDD workflow
plan Planning-only mode — edit/write/apply_patch/bash/webfetch are denied at the permission layer (via deny_tools frontmatter). Plan is delivered as the chat reply; the user saves it to disk if desired.
review Code review mode — edit/write/apply_patch/webfetch denied, but bash stays so the reviewer can inspect read-only (git diff, git log, grep); effectful commands still prompt. Findings delivered in chat
debug Debug mode — finds root cause before proposing fixes
ask Read-only mode — edit/write/apply_patch/bash/webfetch denied via deny_tools
brainstorm Design-only mode — explores ideas and presents designs without code
frontend-design Frontend design mode — distinctive, production-grade UI
review-security Security review mode — same deny list as review (bash kept for read-only inspection); finds exploitable vulnerabilities
simplify Code simplification mode — refines for clarity without changing behavior
write-prompt Prompt writing mode — creates and optimizes agent prompts
default Default system prompt — the base built-in prompt

Per-prompt tool restrictions

Each prompt is a markdown file with optional YAML frontmatter declaring its tool restrictions:

---
deny_tools: [edit, write, apply_patch, bash, webfetch]
description: Read-only planning mode
---
You are dirge in plan mode. …

The permission checker refuses any denied tool BEFORE the call leaves dirge — even under --yolo mode. A denied tool's definition is also withheld from the request, so the model isn't offered a call it can only have rejected (matched by bare name or the qualified mcp_tool:<server>:<name> form; the mcp_tool / plugin_tool umbrellas stay enforcement-only, since a tool definition carries no marker of which server exported it).

Applies symmetrically to MCP tools: an entry in deny_tools matches an MCP-exported tool when the entry equals any of the following:

  • the bare tool name as the MCP server registered it (e.g. edit matches an MCP edit tool from any server — convenient blanket deny, but be aware that deny_tools: [edit] intended for the built-in editor will also block an MCP server's edit tool)
  • the qualified mcp_tool:<server>:<name> form (for narrowly denying a specific server's tool)
  • the umbrella mcp_tool (denies every MCP tool from every server)

For surgical control over one MCP tool without affecting the built-in, use the qualified form.

Critic control (frontmatter)

Two optional keys steer the F6 in-loop critic (see config.md, critic_provider) per prompt, without touching config:

  • critic: false — suppress the critic and the diff-aware code reviewer for this prompt only (both share the critic_provider judge, so the flag gates both; read-only/exploratory modes leave no diff to review anyway). The goal gate (--goal) is unaffected: it has its own judge under its own fixed preamble. critic: true, or omitting the key, inherits the global behavior.

  • critic_preamble: — override the critic's system preamble for this prompt. Wins over critic_preamble in config and the built-in. Inline string or a YAML block scalar (|) for multi-line:

    ---
critic: false
---
    ---
critic_preamble: |
  You are a security-focused reviewer.
  Block only on concrete, in-scope gaps.
---

An empty critic_preamble is treated as unset (inherits). Block-scalar indentation is stripped; folding (>) is not supported.

Custom prompts

Custom prompts can be placed in $XDG_CONFIG_HOME/dirge/prompts/ as .md files (available to every project), and/or in <project>/.dirge/prompts/ for project-local prompts. A project-local prompt with the same stem as a global or built-in prompt overrides it.

Model-aware steering

Separate from the selectable prompt modes above, the harness appends a model-specific guidance fragment to the preamble based on the active model family (resolved from the provider + model id). This is automatic and has no config key.

Today only DeepSeek chat models (v3/v4) receive a fragment, prompts/steering/deepseek.md, embedded in the binary via include_str!. It encodes a Plan-Execute-Verify working method, structural-constraint framing, an explicit success/never contract, and an anti tool-call-repetition rule — the dominant failure modes for DeepSeek in agentic loops. It is appended last in the preamble so it sits closest to the conversation / action boundary, where rules best resist "prompt-distance drift" in long tool-calling sequences.

The DeepSeek reasoner (R1) is deliberately excluded — it ignores the system prompt — and every non-DeepSeek model is unaffected. Steering fragments live under prompts/steering/ rather than prompts/, so they are not selectable /prompt modes and never appear in the /prompt list. To edit the guidance, change the markdown file and rebuild.

Context files

The agent automatically loads AGENTS.md or CLAUDE.md from the project root, ancestor directories, and ~/.config/dirge/agent/AGENTS.md as a global fallback. Use -n / --no-context-files to disable.