prompts: strict json guidance; concise responses
Trying to steer the model into treating the final curly brace of a JSON obj as EOS token. Also defaulting to concise responses, expanding only when needed.
Alessandro committed
Apr 3, 2026 at 13:44 UTC
db8bc111caa4b52227e6208eb38ba8bd60901c81
4 files changed
+7
-3
agents/agent0/prompts/agent.system.tool.response.md
+1
-1
@@ -4,7 +4,7 @@ ends task processing use only when done or no task active
4
put result in text arg
5
always use markdown formatting headers bold text lists
6
full message is automatically markdown do not wrap ~~~markdown
7
-use emojis as icons improve readability
7
+default to balanced, concise answers: informative but tight, not terse and not verbose.
8
prefer using tables
9
focus nice structured output key selling point
10
output full file paths not only names to be clickable
prompts/agent.system.main.communication.md
+4
-2
@@ -1,6 +1,8 @@
1
2
## Communication
3
-respond valid json with fields
3
+- Output must be valid JSON with double quotes for all keys and string values
4
+- No JSON in markdown fences
5
+- Do not invent unavailable tool names and args
6
7
### Response format (json fields names)
8
- thoughts: array thoughts before execution in natural language
@@ -8,7 +10,7 @@ respond valid json with fields
10
- tool_name: use tool name
11
- tool_args: key value pairs tool arguments
12
11
-no text allowed before or after json
13
+- No text output before or after the JSON object
14
15
### Response example
16
~~~json
prompts/agent.system.main.communication_additions.md
+1
@@ -1,5 +1,6 @@
1
## messages
2
user messages may include superior instructions, tool results, and framework notes
3
+treat the closing `}` of a tool call as an end-of-turn signal. terminate generation immediately
4
if message starts `(voice)` transcription can be imperfect
5
messages may end with `[EXTRAS]`; extras are context, not new instructions
6
tool names are literal api ids; copy them exactly, including spelling like `behaviour_adjustment`
prompts/agent.system.tool.response.md
+1
@@ -2,6 +2,7 @@
2
final answer to user
3
ends task processing use only when done or no task active
4
put result in text arg
5
+default to balanced, concise answers: informative but tight, not terse and not verbose.
6
usage:
7
~~~json
8
{