Hone

Capture the original session prompt, analyze it, and generate an optimized version. Builds prompt engineering intuition over time.

Archive root: Resolve $SESSION_KIT_ROOT (default: ~/.stoobz). All ~/.stoobz/ paths below use this root.

Session Check-In (silent — before main process)

On first invocation of any session-kit skill in this session, register the active session in the manifest. See session-checkin.md for the full protocol. Summary:

1. Detect session ID from most recently modified .jsonl in ~/.claude/projects/$(pwd | tr '/' '-')/ (fallback: git root encoding). If detection fails, skip silently.
2. Read $SESSION_KIT_ROOT/manifest.json (create if missing).
3. If no entry with this session_id exists → create active registration (status: "active", session_id, return_to, started_at, last_activity, last_exchange, skills_used, nulls for label/summary/archive_path).
4. If entry exists → update last_activity, last_exchange, append this skill to skills_used.
5. Write manifest. Proceed to main process. No output about check-in.

Process

1. Check for existing file — Read ./.stoobz/HONE.md if it exists. If found:

   • Preserve previous entries under a ## Previous Prompts heading
   • Add new entry as the primary (top) section with updated timestamp

2. Identify the initial prompt — Find the user's first substantive message that defined the session's work. This is the raw prompt exactly as written.

3. Analyze the prompt against these dimensions:

   • Clarity — Was the goal unambiguous?
   • Scope — Was the scope well-defined or open-ended?
   • Context provided — Did it give enough background?
   • Constraints — Were boundaries/requirements stated?
   • Actionability — Could Claude act immediately or needed clarification?

4. Generate the optimized prompt — Rewrite incorporating:

   • Everything learned during the session
   • Context that had to be discovered/clarified mid-session
   • Specific file paths, module names, or technical details that emerged
   • Explicit scope boundaries
   • Skills to load
   • Clear success criteria

5. Add coaching notes — Brief, specific tips about what made the difference between the original and optimized versions.

6. Write .stoobz/HONE.md in the current working directory.

Output Format

# Hone

**Date:** {YYYY-MM-DD}
**Session:** {branch or topic}

---

## Original Prompt

> {Exact verbatim copy of the user's initial prompt, in a blockquote}

## Analysis

| Dimension     | Rating                                     | Notes                           |
| ------------- | ------------------------------------------ | ------------------------------- |
| Clarity       | {good/fair/vague}                          | {specific observation}          |
| Scope         | {focused/broad/undefined}                  | {specific observation}          |
| Context       | {sufficient/partial/missing}               | {what was missing}              |
| Constraints   | {explicit/implicit/none}                   | {what should have been stated}  |
| Actionability | {immediate/needed-clarification/ambiguous} | {what blocked immediate action} |

**What worked well:** {1-2 things the prompt did right}

**What cost time:** {things that required clarification or discovery mid-session}

## Optimized Prompt

> {The rewritten prompt — ready to copy-paste into a new session}
>
> {This should be self-contained: include context, constraints, relevant paths,
> skills to load, and clear success criteria. A new Claude session reading only
> this prompt should be able to start working immediately.}

## Prompt Tips

{2-4 bullet points — specific, actionable lessons from this prompt. Not generic advice.}

- {Tip grounded in what actually happened this session}
- {Tip about what context/constraint would have saved time}

---

_Generated by /hone — building prompt intuition one session at a time._

Rules

• Verbatim original — Copy the exact original prompt, warts and all. No editing.
• Honest analysis — If the prompt was great, say so. If it was vague, say that too. No flattery.
• Optimized != longer — The best prompts are concise. Add specificity, not word count.
• Session-specific tips — Generic prompt advice ("be specific") is useless. Ground every tip in what actually happened.
• The optimized prompt must be self-contained — Someone should be able to paste it cold into a new session and get productive immediately.
• Write to ./.stoobz/HONE.md unless the user specifies a different path