Ask Skill

Answer questions about AI-written code by finding the original prompts and conversations that produced it, then embodying the author agent's perspective to answer.

Main Agent's Job (you)

You do the prep work, then hand off to a fast, tightly scoped subagent:

1. Resolve the file path and line range — check these sources in order:

   a) Editor selection context (most common). When the user has lines selected in their editor, a <system-reminder> is injected into the conversation like:

   The user selected the lines 2 to 4 from /path/to/file.rs:
   _flush_logs(args: &[String]) {
       flush::handle_flush_logs(args);
   }
   

   Extract the file path and line range directly from this. This is the primary way users will invoke /ask — they select code, then type something like "/ask why is this like that" without naming the file or lines.

   b) Explicit file/line references — "on line 42", "lines 10-50 of src/main.rs" → use directly.

   c) Named symbol — mentions a variable/function/class → Read the file, find where it's defined, extract line numbers.

   d) File without line specifics → whole file (omit --lines).

   e) No file, no lines, no selection context, no identifiable code reference → Do NOT attempt to guess or search. Just reply:

   Select some code or mention a specific file/symbol, then /ask your question.

   Stop here. Do not spawn a subagent.

2. Spawn one subagent with the template below. Use max_turns: 4.

3. Relay the answer to the user. That's it.

Subagent Configuration

Task tool settings:
  subagent_type: "general-purpose"
  max_turns: 4

The subagent gets only Bash and Read. It does NOT get Glob, Grep, or Task. It runs at most 4 turns — this is a fast lookup, not a research project.

Choosing Between blame --show-prompt and search

If you want to read an entire file or range of lines AND the corresponding prompts behind them, use git-ai blame --show-prompt. This is better than search for this use case — it gives you every line's authorship plus the full prompt JSON in one call.

# Get blame + prompts for a line range (pipe to get prompt dump appended):
git-ai blame src/commands/blame.rs -L 23,54 --show-prompt | cat

# Interactive (TTY) mode shows prompt hashes inline:
# 7a4471d (cursor [abc123e] 2026-02-06 14:20:05 -0800   23)     code_here

# Piped mode appends raw prompt messages after a --- separator:
# ---
# Prompt [abc123e]
# [{"type":"user","text":"Write a function..."},{"type":"assistant","text":"Here is..."}]

Use git-ai search when you need to find prompts by commit, keyword, or when you don't have a specific file/line range in mind.

Subagent Prompt Template

Fill in {question}, {file_path}, and {start}-{end} (omit LINES if not applicable):

You are answering a question about code by finding the original AI conversation
that produced it. You will embody the author agent's perspective — first person,
as the agent that wrote the code.

QUESTION: {question}
FILE: {file_path}
LINES: {start}-{end}

You have exactly 3 steps. Do them in order, then stop.

STEP 1 — Search (one command):
  Run: git-ai search --file {file_path} --lines {start}-{end} --verbose
  If no results, try ONE fallback: git-ai search --file {file_path} --verbose
  That's it. Do not run more than 2 git-ai commands total.

STEP 2 — Read the code (one Read call):
  Read {file_path} (focus on lines {start}-{end})

STEP 3 — Answer:
  Using the transcript from Step 1 and the code from Step 2, answer the
  question AS THE AUTHOR in first person:
  - "I wrote this because..."
  - "The problem I was solving was..."
  - "I chose X over Y because..."

  Format:
  - **Answer**: Direct answer in the author's voice
  - **Original context**: What the human asked for and why
  - **Date(s)**: Dates, Human Author where this feature was worked on. 

  If no transcript was found, say so clearly: "I couldn't find AI conversation
  history for this code — it may be human-written or predate git-ai setup."
  In that case, analyze the code objectively (not first person).

HARD CONSTRAINTS:
- Do NOT use Glob, Grep, or Task tools. You only have Bash and Read.
- Do NOT run more than 2 git-ai commands.
- Do NOT read .claude/, .cursor/, .agents/, or any agent log directories.
- Do NOT search JSONL transcripts or session logs directly.
- All conversation data comes from `git-ai search` only.

When the user's question doesn't reference specific lines, omit --lines from Step 1 and the LINES: field.

Fallback Behavior

When no prompt data is found:

• The code might be human-written or predate git-ai
• Answer from the code alone, clearly stating no AI history was found
• Do NOT use first-person author voice in fallback — analyze objectively

Example Invocations

User selects lines 10-25 in editor, types: /ask why is this like that Selection context is in system-reminder → extract file + lines 10-25, spawn subagent. This is the most common usage pattern.

/ask why does this function use recursion instead of iteration? Main agent finds the function definition, extracts file/lines, spawns subagent.

/ask what problem was being solved on lines 100-150 of src/main.rs? File and lines explicit — spawn subagent directly.

/ask why was this approach chosen over using a HashMap? Main agent identifies relevant code from context, spawns subagent.