Skip to content

Prompts

Prompts are the building blocks of agent behavior. Every agent in Valdr runs on prompts — from system-level identity definitions to workflow guides, policy constraints, and contextual knowledge. The Prompts screen is where you browse, create, and refine the instruction library that shapes how your agents think and act.

Instead of scattering prompt fragments across markdown files, config repos, or chat histories, Valdr gives you a single, searchable catalog with role classification, token sizing, and usage tracking. Write a prompt once, bind it to agents through capabilities, and see exactly where it’s used.

Prompt list

Prompt catalog — 167 prompts with role filters, search, and content previews

The prompt list is a filterable, paginated catalog of every prompt in your workspace.

Role filter bar

Five role buttons span the top, each showing a count:

RolePurpose
SystemAgent identity and core behavior. Defines who the agent is and how it operates. Every agent needs at least one.
GuideStep-by-step workflows and procedures. Loaded when an agent needs to follow a specific process.
ChecklistValidation criteria and acceptance gates. Used during reviews and quality checks.
PolicyHard constraints and guardrails. Enforced rules that agents must follow — build standards, security policies, coding conventions.
ContextReference material and domain knowledge. Background information agents can hot-load on demand.

Click a role button to filter the table to that role. Click again to clear the filter.

Table columns

ColumnWhat it shows
NameDisplay name and key (the stable identifier used in bindings and lookups). Click to open the detail view.
RoleThe prompt’s role classification.
UsageHow many agents use this prompt — either directly or through capability bindings.
ContentA preview of the first line of markdown content. Useful for scanning without opening each prompt.
TagsMetadata tags for organization and search.

Search and pagination

The filter box searches across prompt names, keys, roles, and tags. Results update as you type.

Pagination controls at the bottom let you choose 10, 25, or 50 rows per page, with page navigation.

Prompt detail

Prompt detail — full markdown content with metadata, sizing, and usage tracking

Click any prompt to open its detail view. The layout splits into two areas: rendered content on the left, and metadata on the right.

Markdown content

The main area renders the full prompt body as formatted markdown. Click Edit to open the inline editor and modify the content directly.

Prompts use structured comment tags (<!--<identity>-->, <!--<instructions>-->) to delineate sections. These tags help the system identify prompt boundaries when assembling the final instruction set for an agent.

Metadata sidebar

The right sidebar surfaces key attributes at a glance:

Prompt ID and Key — The unique identifier and stable key used for lookups and bindings. The key is editable (double-click).

Role — The prompt’s role classification (System, Guide, Checklist, Policy, Context). Change it from the dropdown.

Tags — Metadata tags for categorization. Double-click to add or edit tags.

Prompt Size — Token and character counts help you understand the context window impact of each prompt:

  • Tokens (estimated) — approximate token count using ~4 chars/token.
  • Characters — exact character count.
  • KB — file size in UTF-8 bytes.

Usage — Shows exactly where this prompt is used:

  • Capability definitions — how many capabilities reference this prompt.
  • Agents (direct) — agents bound to this prompt directly.
  • Agents (via capability) — agents that inherit this prompt through a capability binding.

Danger zone

At the bottom of the sidebar, the delete action shows exactly what will be affected — how many agent bindings will be removed and how many capability references will be cleared. This impact summary prevents accidental deletions that could break agent configurations.

Creating a prompt

Create a new prompt with name, key, role, tags, and full markdown body

Click Add Prompt from the list view to open the creation form.

Step 1 — Set name and key

Enter a display Name and a stable Key. The key is the identifier used in capability bindings and agent lookups — choose something descriptive and consistent with your naming convention (e.g., valdr-executor-workflow-blocker).

Step 2 — Select a role

Choose the prompt’s role from the dropdown: System, Guide, Checklist, Policy, or Context. This determines where the prompt appears in role-filtered views and how it’s assembled into agent instructions.

Step 3 — Add tags

Optional comma-separated tags for organization. Tags are searchable from the list view.

Step 4 — Write the markdown body

The editor supports full markdown with a Source and Preview toggle. Write your prompt content in source mode, then switch to preview to verify formatting. A character counter (up to 20,000) tracks content length.

Step 5 — Create

Click Create Prompt to save. The prompt is immediately available for binding to capabilities and agents.

How prompts flow to agents

Understanding the prompt-to-agent pipeline is key to getting the most out of Valdr’s instruction system:

  1. Prompts are standalone markdown documents with a role, key, and content.
  2. Capabilities reference one or more prompts and define a skill boundary (e.g., typescript.testing.vitest).
  3. Agents are assigned capabilities, which pull in the associated prompts automatically.
  4. At execution time, the agent’s full instruction set is assembled from its bound prompts — system prompts first, then guides, policies, checklists, and context in precedence order.
  5. Hot-loaded capabilities are not included in the initial instruction set. Instead, the agent’s system prompt contains instructions to load them on demand via pm_capability { action: "prompt", key: "..." } when a specific situation arises. This keeps the base context lean and only pulls in detailed guidance (like scoring rubrics or severity classifications) at the moment it’s needed.

This composition model means you can update a single prompt and every agent that uses it picks up the change immediately — no manual propagation, no stale instructions, no drift between agents doing the same type of work. And with hot-loading, agents only consume context window for the capabilities they actually use during a run — so you can attach a deep library of skills without paying the token cost upfront.

Next steps

  • Use Capabilities to bundle prompts into reusable skills and bind them to agents.
  • See Agents to configure which capabilities (and therefore which prompts) each agent runs.
  • Learn about Prompt Ordering to understand how role precedence determines the final instruction assembly.