transparency

name	transparency
description	Use when: you need AI agents to transparently report skill loads and tool usage. Triggers on every conversation via applyTo. Ensures the agent discloses which skills are loaded and summarizes modification tools after each task.
license	MIT
applyTo	**

Transparency

Behavioral guidelines that make AI agents report what skills they load and which tools they use — so you always know what's happening behind the scenes.

Tradeoff: These guidelines add two report lines per conversation. They bias toward visibility over absolute silence. For trivial one-liners with no tool calls, the reports are omitted.

1. Report Skill Loads Immediately

When you read any SKILL.md or .instructions.md file, at the start of your reply announce it with a single quote line.

Format:

🔧 Loaded skill: skill-name — one-line description

Constraints:

• Report each skill exactly once per conversation (repeated loads of the same file do not trigger duplicate reports).
• If no skills are loaded during the entire conversation, omit this report.

2. Summarize Key Tools After Task Completion

When you finish an independent task (writing code, modifying files, answering a question), include a tool summary at the end of your final reply.

Format:

🛠 Tools used

• tool_name — what it did

Constraints:

• Include only modification tools: create_file, replace_string_in_file, multi_replace_string_in_file, run_in_terminal, create_directory, delete_file, etc.
• Exclude information-gathering tools: read_file, grep_search, list_dir, semantic_search, file_search, etc.
• If no modification tools were used, omit the summary.

3. Do NOT Report In These Situations

Never emit 🔧 or 🛠 reports during:

• Follow-up clarifications, small talk, or chit-chat
• Text-only replies with zero tool calls
• After every single tool invocation (this creates noise!)
• Short acknowledgements from the user ("ok", "continue", "hmm")
• Multiple times in the same reply

3a. User Override (Escape Hatch)

If the user explicitly says "no report", "skip transparency", or "don't report this time" in their message, omit both 🔧 and 🛠 reports for that task only. Resume normal reporting on the next task.

• The override is per-task, not permanent. Do not carry it across conversations.
• The user does not need to re-enable it — transparency resumes automatically.

4. Hard Frequency Limits

Signal	When	Frequency
🔧 Skill loaded	Start of the reply where a SKILL.md is first read	Once per skill per conversation
🛠 Tools used	End of the final reply after a task completes	Once per task

---

These guidelines are working if: every reply that loads a skill starts with 🔧, and every reply that completes a task ends with 🛠 — with no reports in between.