Skip to content

PROMPTS.md

Latest commit

 

History

History
110 lines (78 loc) · 4.87 KB

PROMPTS.md

File metadata and controls

110 lines (78 loc) · 4.87 KB

Prompts

keep uses three families of prompt docs, all stored under .prompt/:

Family IDs Used by
.prompt/summarize/ default, conversation, ... keep put (summarization)
.prompt/analyze/ default, conversation, ... keep analyze (decomposition)
.prompt/agent/ reflect, query, conversation, session-start, ... keep prompt CLI and keep_prompt MCP tool

Summarize and analyze prompts replace the LLM system prompt for their respective operations. Agent prompts are rendered with context injection and returned as text — see KEEP-PROMPT.md for the CLI and KEEP-MCP.md for the MCP tool.

Summarize and analyze prompts

Each summarize/analyze prompt doc has two parts:

  1. Match rules — tag queries that determine when this prompt applies (same DSL as .meta/* docs)
  2. ## Prompt section — the actual system prompt text sent to the LLM

When a document is summarized or analyzed, keep scans the matching .prompt/{type}/ docs, finds those whose match rules match the document's tags, and selects the most specific match (most rules matched). The ## Prompt section from the winner is the system prompt sent to the model.

The prompt docs in the store are authoritative at runtime. Bundled prompt files are only the seed content used during system-doc migration. If the matching store prompt is missing or lacks a ## Prompt section, the operation fails instead of silently falling back to Python defaults.

Bundled

ID Match rule Purpose
.prompt/summarize/default (none — fallback) Default summarization prompt
.prompt/summarize/conversation type=conversation Preserves dates, names, facts from conversations
.prompt/analyze/default (none — fallback) Section-by-section extraction for arbitrary documents
.prompt/analyze/conversation type=conversation Fact extraction from conversations
.prompt/analyze/paper type=paper Structural decomposition of research papers along actual sections

Creating custom prompts

Create a new prompt doc with match rules targeting specific tags:

# Custom summarization for code documentation
keep put "$(cat <<'EOF'
topic=code

## Prompt

Summarize this code documentation in under 200 words.
Focus on: what the API does, key parameters, return values, and common pitfalls.
Begin with the function or class name.
EOF
)" --id .prompt/summarize/code

Match rules can combine multiple tags for higher specificity:

# Prompt for meeting notes in a specific project
keep put "$(cat <<'EOF'
type=meeting project=myapp

## Prompt

Summarize this meeting in under 300 words.
Focus on decisions made, action items assigned, and deadlines mentioned.
List each action item with its owner.
EOF
)" --id .prompt/summarize/myapp-meetings

The most specific match wins — a prompt matching type=meeting project=myapp (2 rules) beats one matching just type=meeting (1 rule), which beats the default (0 rules).

Agent prompts

Agent prompts live at .prompt/agent/ and contain template placeholders ({get}, {find}, {text}) that are expanded with store context at render time. They are returned as text for the calling agent to follow — not sent to an LLM as a system prompt.

keep prompt reflect                      # Render the reflect prompt
keep prompt reflect "auth flow"          # With search context injected
keep prompt --list                       # List available agent prompts

State-doc backed prompts

Dynamic agent prompts must reference a state doc via a state tag. The state doc flow runs and its bindings become {binding_name} placeholders in the template. This separates retrieval logic (state doc) from presentation (prompt template). A prompt that uses {get} or {find} without a state tag is invalid and fails at render time.

keep put "## Prompt
{similar}
Question: {text}
" --id .prompt/agent/my-query -t state=my-retrieval-flow -t context=prompt

See KEEP-PROMPT.md for full usage, template placeholders, and state-doc binding details.

Viewing prompt docs

keep get .prompt/summarize/default       # See a summarize/analyze prompt
keep get .prompt/agent/reflect           # See an agent prompt
keep list .prompt                        # All prompt docs

Prompt docs are editable and versioned like any other store document.

See Also