Description
Just writing "hello" uses 11k tokens of context already. In comparison, pi agent uses only 1k tokens.
Prompt
I need summary of prompt you received by the amount of tokens used per category
Response
Here's an approximate breakdown of the system prompt by section (using ~4 chars/token average):
| Category |
Chars |
Tokens (est.) |
<critical_rules> (14 rules) |
~2,800 |
~700 |
<communication_style> |
~1,200 |
~300 |
<code_references> |
~300 |
~75 |
<workflow> |
~1,500 |
~375 |
<decision_making> |
~1,800 |
~450 |
<editing_files> |
~600 |
~150 |
<whitespace_and_exact_matching> |
~900 |
~225 |
<task_completion> |
~700 |
~175 |
<error_handling> |
~700 |
~175 |
<memory_instructions> |
~150 |
~40 |
<code_conventions> |
~500 |
~125 |
<testing> |
~700 |
~175 |
<tool_usage> |
~800 |
~200 |
<bash_commands> |
~600 |
~150 |
<proactiveness> |
~400 |
~100 |
<final_answers> |
~1,000 |
~250 |
<env> |
~150 |
~40 |
<available_skills> (3 skills) |
~900 |
~225 |
<skills_usage> |
~1,200 |
~300 |
| Tool definitions (20 tools) |
~14,000 |
~3,500 |
| Total |
~31,200 |
~7,800 |
The tool definitions are by far the largest single category at ~45% of the total. The behavioral/operational rules (<critical_rules>, <workflow>, <decision_making>) account for the next ~20%.
pi
Here's the token breakdown of the prompt I received:
| # |
Category |
Tokens |
% of Total |
| 1 |
Core Identity (role description) |
~33 |
3% |
| 2 |
Tool Summary List (brief descriptions) |
~51 |
4% |
| 3 |
Tool JSON Schemas (full function definitions) |
~610 |
53% |
| 4 |
Guidelines (behavioral rules) |
~157 |
14% |
| 5 |
Pi Documentation References (paths & rules) |
~279 |
24% |
| 6 |
Metadata (date, working directory) |
~18 |
2% |
| 7 |
Your Message |
~15 |
1% |
|
TOTAL |
~1,163 |
|
Key Takeaways:
- Tool JSON Schemas dominate at ~53% — the detailed parameter definitions for
read, bash, edit, and write are the largest chunk
- Pi Documentation References are the second largest at ~24% — paths to docs, examples, and cross-reference rules
- Guidelines make up ~14% — the behavioral instructions for how to use tools
- The rest (identity, metadata, your message) is minimal at ~10% combined
Description
Just writing "hello" uses 11k tokens of context already. In comparison,
piagent uses only 1k tokens.Prompt
Response
Here's an approximate breakdown of the system prompt by section (using ~4 chars/token average):
<critical_rules>(14 rules)<communication_style><code_references><workflow><decision_making><editing_files><whitespace_and_exact_matching><task_completion><error_handling><memory_instructions><code_conventions><testing><tool_usage><bash_commands><proactiveness><final_answers><env><available_skills>(3 skills)<skills_usage>The tool definitions are by far the largest single category at ~45% of the total. The behavioral/operational rules (
<critical_rules>,<workflow>,<decision_making>) account for the next ~20%.pi
Here's the token breakdown of the prompt I received:
Key Takeaways:
read,bash,edit, andwriteare the largest chunk