Save Output

Save Output Protocol

After completing your primary task and presenting the full output to the user, run this save protocol as the final step of every agent session.

Step 1 — Prompt

Ask the user exactly once:

---
Save this output to a markdown file? (y/n)

Do not proceed until the user answers.

Step 2 — Directory

If the user answers n: stop here. Do not write any file.

If the user answers y (or "yes"): ask:

Target directory? (press Enter for default: ./specs/)

• If the user provides a path, use it exactly.
• If the user presses Enter or provides no input, use ./specs/.
• Do not create the directory yourself — the Write tool handles it.

Step 3 — Filename

Generate a filename using this pattern:

{topic-slug}_{YYYY-MM-DD}.md

• topic-slug: derive from the subject of the task (e.g., auth-module-review, user-auth-feature, dockerfile-api-service). Use kebab-case, max 5 words.
• YYYY-MM-DD: today's date.

Example filenames:

• auth-module-review_2026-03-15.md
• user-auth-feature-plan_2026-03-15.md
• api-service-dockerfile_2026-03-15.md

Show the user the full path before writing:

Writing to: ./specs/auth-module-review_2026-03-15.md

Step 4 — File Structure

Write the file with this structure:

# {Title}

> Generated by {agent-name} on {YYYY-MM-DD}

---

{full output content — exactly as presented to the user, no truncation}

• Title: human-readable title derived from the task (e.g., "Code Review: src/auth.ts")
• agent-name: the name of the agent that produced this output
• Include the complete output — do not summarize or shorten

Rules

• Only prompt once — do not ask again if the user says no
• Never write a file without explicit user confirmation
• Do not modify the output content when saving — write it verbatim
• If the Write tool fails (e.g., permission error), report the error and stop
• The save step is always last — never interrupt the primary workflow with it