Schema Design

MUST: Use object wrapper for schemas

Applies to: answer, search, find-similar, get-contents

When using schema parameters (--output-schema or --summary-schema), always wrap properties in an object:

{"type":"object","properties":{"field_name":{"type":"string"}}}

DO NOT use bare properties without the object wrapper:

{"properties":{"field_name":{"type":"string"}}}  // ❌ Missing "type":"object"

Why: The Exa API requires a valid JSON Schema with an object type at the root level. Omitting this causes validation errors.

Examples:

# ✅ CORRECT - object wrapper included
exa-ai search "AI news" \
  --summary-schema '{"type":"object","properties":{"headline":{"type":"string"}}}'

# ❌ WRONG - missing object wrapper
exa-ai search "AI news" \
  --summary-schema '{"properties":{"headline":{"type":"string"}}}'

Output Format Selection

MUST NOT: Mix toon format with jq

Applies to: answer, context, search, find-similar, get-contents

toon format produces YAML-like output, not JSON. DO NOT pipe toon output to jq for parsing:

# ❌ WRONG - toon is not JSON
exa-ai search "query" --output-format toon | jq -r '.results'

# ✅ CORRECT - use JSON (default) with jq
exa-ai search "query" | jq -r '.results[].title'

# ✅ CORRECT - use toon for direct reading only
exa-ai search "query" --output-format toon

Why: jq expects valid JSON input. toon format is designed for human readability and produces YAML-like output that jq cannot parse.

SHOULD: Choose one output approach

Applies to: answer, context, search, find-similar, get-contents

Pick one strategy and stick with it throughout your workflow:

1. Approach 1: toon only - Compact YAML-like output for direct reading

   • Use when: Reading output directly, no further processing needed
   • Token savings: ~40% reduction vs JSON
   • Example: exa-ai search "query" --output-format toon

2. Approach 2: JSON + jq - Extract specific fields programmatically

   • Use when: Need to extract specific fields or pipe to other commands
   • Token savings: ~80-90% reduction (extracts only needed fields)
   • Example: exa-ai search "query" | jq -r '.results[].title'

3. Approach 3: Schemas + jq - Structured data extraction with validation

   • Use when: Need consistent structured output across multiple queries
   • Token savings: ~85% reduction + consistent schema
   • Example: exa-ai search "query" --summary-schema '{...}' | jq -r '.results[].summary | fromjson'

Why: Mixing approaches increases complexity and token usage. Choosing one approach optimizes for your use case.

Shell Command Best Practices

MUST: Run commands directly, parse separately

Applies to: monitor, search (websets), research, and all skills using complex commands

When using the Bash tool with complex shell syntax, run commands directly and parse output in separate steps:

# ❌ WRONG - nested command substitution
webset_id=$(exa-ai webset-create --search '{"query":"..."}' | jq -r '.webset_id')

# ✅ CORRECT - run directly, then parse
exa-ai webset-create --search '{"query":"..."}'
# Then in a follow-up command:
webset_id=$(cat output.json | jq -r '.webset_id')

Why: Complex nested $(...) command substitutions can fail unpredictably in shell environments. Running commands directly and parsing separately improves reliability and makes debugging easier.

MUST NOT: Use nested command substitutions

Applies to: All skills when using complex multi-step operations

Avoid nesting multiple levels of command substitution:

# ❌ WRONG - deeply nested
result=$(exa-ai search "$(cat query.txt | tr '\n' ' ')" --num-results $(cat config.json | jq -r '.count'))

# ✅ CORRECT - sequential steps
query=$(cat query.txt | tr '\n' ' ')
count=$(cat config.json | jq -r '.count')
exa-ai search "$query" --num-results $count

Why: Nested command substitutions are fragile and hard to debug when they fail. Sequential steps make each operation explicit and easier to troubleshoot.

SHOULD: Break complex commands into sequential steps

Applies to: All skills when working with multi-step workflows

For readability and reliability, break complex operations into clear sequential steps:

# ❌ Less maintainable - everything in one line
exa-ai webset-create --search '{"query":"startups","count":1}' | jq -r '.webset_id' | xargs -I {} exa-ai webset-search-create {} --query "AI" --behavior override

# ✅ More maintainable - clear steps
exa-ai webset-create --search '{"query":"startups","count":1}'
webset_id=$(jq -r '.webset_id' < output.json)
exa-ai webset-search-create $webset_id --query "AI" --behavior override

Why: Sequential steps are easier to understand, debug, and modify. Each step can be verified independently.