Prompt Author

Create custom prompt content for claude-code-modes — bases, modifiers, or axis values — grounded in Anthropic's emotion research.

Before starting

Read these references:

1. references/emotion-research.md — emotion vector findings and principles
2. references/prompt-quality.md — 10 quality criteria
3. references/base-format.md — manifest and fragment specs
4. references/examples.md — chill base as worked example

Read the existing prompts to understand the current patterns:

• prompts/base/ — standard base fragments
• prompts/chill/ — chill base fragments (emotion-research-informed)
• prompts/modifiers/ — existing modifiers
• prompts/axis/ — existing axis values

Phase 1: Determine what to create

AskUserQuestion: What type of prompt content?

• Base — a complete foundational prompt (directory + manifest + fragments)
• Modifier — a behavioral layer that composes on top of any base/preset
• Axis value — a custom value for agency, quality, or scope

Phase 2: Interview about desired behavior

Ask the user conversationally (not all at once):

For a base:

• What's the philosophy? (e.g., minimal, domain-specific, team-flavored)
• What behaviors matter most?
• What tone? (direct, warm, formal, casual)
• How does it differ from standard and chill?
• What fragments should it have? (consolidate like chill, or granular like standard?)

For a modifier:

• What behavior does this shape? (investigation, craftsmanship, caution, speed)
• When would someone use it? (debugging, reviewing, prototyping)
• What should the agent do differently with this modifier active?
• What should it explicitly NOT do?

For an axis value:

• Which axis? (agency, quality, scope)
• What behavior does this value represent?
• How does it differ from the existing built-in values?

Phase 3: Draft the content

Apply these principles from the emotion research (see references/emotion-research.md):

Writing rules

1. Positive framing first. State what to do, not what to avoid. "Keep changes scoped to the request" not "Don't add features beyond what was asked."
2. Zero ALL-CAPS emphasis. No IMPORTANT, CRITICAL, MUST, NEVER. State instructions with calm confidence.
3. Activate calm vectors. Use assured, direct language. "Consider the impact" not "you MUST check with the user FIRST."
4. Avoid desperation triggers. No stacked contradictory requirements. No pressure language. No impossible constraints.
5. Counter post-training brooding. Encourage leading with answers. Directness over qualifications.
6. Preserve protective caution. Reframe as professional judgment, not anxiety. "Actions that warrant consideration" not "NEVER do this without asking."
7. Include worked examples. Use <example> tags for complex behaviors. Show good vs bad. 2-3 examples for core instructions.
8. Add a priority hierarchy (for bases). One line near the top: "When guidelines conflict: safety first, then user instructions, then correctness, then style."
9. Security boundaries stay firm. Calm tone, but no ambiguity on injection prevention, auth testing limits, URL restrictions.
10. Keep it lean. Target 40-70% the token count of the standard base. Every instruction should earn its place.

For bases specifically

• Write a base.json manifest (flat JSON array with "axes" and "modifiers" reserved words)
• Decide fragment structure — consolidate like chill (fewer, bigger fragments) or granular like standard
• Include env.md with all template variables (see references/base-format.md)
• Actions fragment should be behaviorally neutral — the agency axis handles the autonomous-vs-cautious split

For modifiers specifically

• Keep under 200 words. Modifiers are behavioral nudges, not full rewrites.
• Use the same calm, positive framing as base content.
• Start with a heading (# Modifier Name).
• Focus on one behavioral dimension — don't try to cover everything.

Phase 4: Review with user

Present the drafted content and ask for feedback. Highlight:

• Token count compared to standard/chill equivalents
• Number of negative instructions (target: < 5)
• Number of ALL-CAPS markers (target: 0)
• Number of worked examples
• Whether security boundaries are preserved

Iterate until the user approves.

Phase 5: Write files

For a base: Create the directory, write base.json and all fragment files. For a modifier: Write the .md file to prompts/modifiers/. For an axis value: Write the .md file to prompts/axis/{axis}/.

Remind the user to:

1. Run bun scripts/generate-prompts.ts if the content should be embedded (built-in bases/modifiers)
2. Register custom content in .claude-mode.json if it's project-specific
3. Test with claude-mode create --base <name> --print or claude-mode create --modifier <name> --print

Quality checklist

Before finalizing, verify:

• Negative instruction count < 5 per 1000 characters
• Zero ALL-CAPS emphasis markers
• Security boundaries present and unambiguous
• At least 2 worked examples (for bases)
• Priority hierarchy present (for bases)
• All template variables present in env.md (for bases)
• Manifest has both "axes" and "modifiers" entries (for bases)
• Content is shorter than the standard base equivalent