Prompting

Prompting

Single entry point for prompt engineering workflows. Parse the first positional argument as the subcommand; forward remaining arguments.

Subcommand	Purpose
craft [goal] [--research]	Generate a new optimized prompt via the llm-prompt-engineer agent
cache <list|add|prune|refresh> [args]	Manage the research cache (wraps claude-prove prompting cache CLI)
token-count <patterns...>	Estimate token counts (wraps claude-prove prompting token-count CLI)

Default to craft when the user's intent is "make a prompt" without naming a subcommand. Default to cache list when the user says "cache" with no further args. token-count requires at least one pattern.

---

Subcommand: craft

Delegate to the llm-prompt-engineer agent to generate an optimized prompt from the user's requirements.

Before Delegating

Gather these inputs (ask if missing):

1. Target model — Claude, GPT, open-source, or general-purpose
2. Task type — classification, generation, analysis, agent behavior, system prompt
3. Output expectations — what a successful response looks like
4. Constraints — token budget, latency, cost, safety
5. Consumption context — one-shot, system instruction, or agent definition

Do NOT delegate until intent and constraints are clear.

Research Mode

• Default (no flag): Agent uses bundled references/prompt-engineering-guide.md and any cached research. No WebSearch/WebFetch.
• --research flag or explicit user request: Agent does live web research and caches results.
• Agent-detected gap: If the bundled guide and cache don't cover the ask, agent asks the user before researching.

Delegation Instructions

Pass gathered requirements to llm-prompt-engineer with these directives:

1. Read references/prompt-engineering-guide.md
2. Check the research cache in priority order — later tiers override earlier: plugin (cache/prompting/ in plugin dir), global (~/.claude/cache/prompting/), project (.prove/cache/prompting/)
3. Apply research mode (default: guide + cache; --research: live web, cache results)
4. Apply all relevant optimization techniques
5. Embed brief inline comments in the generated prompt explaining key design choices (e.g., <!-- Primacy effect: critical constraint placed first -->)
6. Present the result with:
   • Techniques applied — each technique used and why
   • Guide sections referenced — which parts of the bundled guide informed the design
   • Token estimate — approximate token count
   • Trade-offs — what was optimized for vs deprioritized

After Generation

Offer to:

• Iterate on specific sections
• Adjust token-efficiency vs clarity balance
• Add or remove techniques
• Test with example inputs
• Research a specific aspect (--research)

If the user provides a file path, write the final prompt there. Otherwise, output directly.

---

Subcommand: cache

Manage the prompt engineering research cache that supplements the bundled guide. Wraps claude-prove prompting cache CLI for filesystem operations; research (add/refresh) is performed by this skill using WebSearch/WebFetch.

Cache Tiers (Priority Order)

Later tiers override earlier tiers for entries with the same filename.

1. Plugin (read-only): cache/prompting/ in the plugin directory. Ships with seed entries. Not user-editable.
2. Global: ~/.claude/cache/prompting/ — user-managed, shared across projects.
3. Project: .prove/cache/prompting/ — project-specific overrides.

Cache Entry Format

Each entry is a markdown file named by topic slug (e.g., claude-tool-use.md):

---
topic: Claude tool use prompting
source: Anthropic docs, blog posts
fetched: 2026-03-29
---

<distilled research content>

Actions

Parse the first argument after cache. Default to list if none.

list

Show all entries from all three tiers, grouped by location. Read frontmatter to show topic + fetch date. Mark plugin entries "read-only".

Plugin (read-only):
  claude-prompting-2026.md    — Claude prompting best practices 2026 (2026-03-29)

Global (~/.claude/cache/prompting/):
  few-shot-patterns.md        — Few-shot example patterns (2026-03-15)

Project (.prove/cache/prompting/):
  gemini-grounding.md         — Gemini search grounding (2026-03-28)

add <topic>

Research a topic and save to cache.

1. AskUserQuestion with header "Cache scope" and options: "Global" / "Project"
2. Research via WebSearch + WebFetch — gather 3-5 authoritative sources
3. Distill into a focused, actionable reference (not a link dump)
4. Write to the chosen cache directory with frontmatter (topic, source, fetched date)
5. Report what was cached and the file path

prune

Remove stale entries from global and project tiers. Plugin tier is read-only.

1. List entries from global + project with fetch dates
2. AskUserQuestion with header "Prune" and multiSelect options listing each entry
3. Delete selected entries
4. Report what was removed

refresh <topic>

Re-research an existing entry in place.

1. Find the existing entry (project first, then global)
2. If not found, suggest add instead
3. Research fresh via WebSearch + WebFetch
4. Overwrite the entry, updating the fetched date
5. Show a brief diff summary of what changed

---

Subcommand: token-count

Estimate token counts for files via regex-based heuristic tokenizer. Works on any text file.

Usage

Run from the user's project root. Invoke the CLI directly:

claude-prove prompting token-count <patterns...> [flags]

Positional arguments are glob patterns or literal file paths. Multiple patterns combinable.

Flag	Effect
--sort tokens	Sort by token count descending (default)
--sort name	Sort alphabetically by path
--sort lines	Sort by line count descending
--json	Machine-readable JSON output
--no-strip	Include YAML frontmatter in count (stripped by default)

Examples

claude-prove prompting token-count '**/*.md'
claude-prove prompting token-count agents/llm-prompt-engineer.md
claude-prove prompting token-count 'agents/**/*.md' 'skills/**/SKILL.md'
claude-prove prompting token-count '**/*.md' --json

Interpreting Results

• Tokens: Heuristic, typically within 10-15% of Claude's BPE tokenizer. Slightly overcounts.
• Lines/Chars: After frontmatter stripping (unless --no-strip).
• Frontmatter stripping: YAML between --- markers excluded by default — Claude Code doesn't send frontmatter as prompt content.

When to Use

• Before/after prompt optimization — measure the delta
• Auditing prompt budgets across a project
• Comparing agent definitions to find outliers
• Checking if a reference file is too large for context