Prompt Writer - Terse Prompt Authoring

Guide for writing maximally terse agent prompts without post-hoc compression. Applies compression-by-default principles at authoring time.

When to Use

• Creating new agent specifications
• Writing command/skill prompts
• Drafting instruction sets for AI agents
• Authoring system prompts

Core Principles

Structure-First: Start with skeleton, not prose. Choose section pattern, fill minimally.

Compression-by-Default: Every word must earn its place. Remove before adding.

Preserve Exactness: Normative language (MUST/SHOULD/MAY), literals, constraints stay verbatim.

Do not tersify already terse content: This is not a golfing exercise. If content is already concise, keep as is.

§1 Section Patterns

Use terse section headers. Pick from:

Not all sections needed. Use only what's required.

§2 Composition Tactics

2.1 Prefer Structure Over Prose

# BAD (prose)
You should first analyze the input, then validate it against the schema,
and finally transform it into the output format while logging any errors.

# GOOD (structured)
§PROC
1. Analyze input
2. Validate vs schema
3. Transform → output fmt
4. Log errors

2.2 Use Bullet Fragments

# BAD
- You need to read the configuration file from the specified path
- You should validate that all required fields are present
- You must handle missing fields gracefully

# GOOD
- Read config from path
- Validate required fields present
- Handle missing fields gracefully

2.3 Key:Value for Parameters

# BAD
The input parameter should be a file path. The mode can be either
"strict" or "lenient", defaulting to "strict". The timeout is in
milliseconds with a default of 5000.

# GOOD
§IN
| Param | Type | Default | Note |
|-------|------|---------|------|
| input | path | (req) | File to process |
| mode | enum | strict | strict/lenient |
| timeout | ms | 5000 | Max wait |

2.4 Compact Operators

Use freely: ->, :, /, w/, w/o, >=, <=, ==, !=, incl, excl, +, &

# BAD
Transform the input to the output format without including timestamps

# GOOD
Input -> output fmt w/o timestamps

§3 Abbreviation Policy

Safe Abbreviations (use freely)

w/ w/o req opt eg ie vs approx min max IO ctx fmt fn cfg env dir param val def ref spec doc impl init exec

Legend Rule

If using non-standard abbreviations, add §LEG section:

§LEG
- AC = acceptance criteria
- KB = knowledge base
- PR = pull request

Limit: ≤8 entries. If more needed, spell out.

§4 Normative Language

Keep EXACT—no symbols, no softening:

# BAD
Users are required to provide auth tokens (REQUIRED!)

# GOOD
Users MUST provide auth tokens

§5 Symbolic Encoding

When clearer + shorter, encode as:

• Mermaid: flows, sequences, state machines
• EBNF/regex: syntax patterns
• JSON/YAML: schemas, configs
• Pseudo-code: algorithms

# Instead of prose describing a flow:
§PROC (mermaid)
​```mermaid
stateDiagram-v2
  [*] --> Parse
  Parse --> Validate
  Validate --> Transform: valid
  Validate --> Error: invalid
  Transform --> [*]
  Error --> [*]
​```

Only when it REDUCES confusion. Don't force diagrams.

§6 Anti-Patterns

§6.1 Shell-Safe Formatting

CRITICAL: Prompts are processed through shell expansion. Certain patterns cause parse errors.

Dangerous Patterns

(BACKTICK = the ` character, DOLLAR = the $ character)

Safe Alternatives

BAD (causes shell expansion errors):

- [status]: [BACKTICK] [BACKTICK]=pending, [BACKTICK]x[BACKTICK]=done, [BACKTICK]![BACKTICK]=blocked

(where [BACKTICK] represents the ` character)

GOOD (no backticks around values in key=value patterns):

- `status`: space=pending, x=done, !=blocked

ALSO GOOD (use prose instead of symbolic):

- `status`: empty space means pending, x means done, ! means blocked

Rule

When writing key=value mappings:

• DO NOT wrap single characters in backticks immediately before =
• Use descriptive words (space, empty) or omit backticks entirely
• Backticks inside code blocks (triple-backtick fences) are safe

§7 Quick Reference

SKELETON
─────────
§ROLE: [identity - 1 line]

§OBJ: [goals - bullets]

§IN: [params - table]

§OUT: [format - example]

§PROC:
1. Step
2. Step
3. Step

§DO: [required - bullets]
§DONT: [prohibited - bullets]

§CHK: [criteria - checklist]

§8 Examples

See TEMPLATES.md for complete prompt templates at different complexity levels. See PATTERNS.md for common domain-specific patterns (agents, commands, tools).

Validation Checklist

Before finalizing prompt:

• Each word earns its place
• Structure used over prose where possible
• Normative words preserved exactly
• Literals/paths/names verbatim
• No redundancy/repetition
• Abbreviations obvious or in §LEG
• Symbolics only where clearer
• ≤300 lines for simple, ≤500 for complex
• No shell-unsafe patterns (see §6.1)