Aha

Aha

Capture and revisit insights — the small "I didn't know that" moments that otherwise evaporate.

Three modes, picked by the argument:

Argument	Mode
(none)	Scan recent context, propose insights, record those the user wants
quiz	Quiz the user on existing insights in .ai/insight/
anything else	Treat argument as a topic hint; focus capture on that subject

Insights live in {repo root}/.ai/insight/{yyyy-mm-dd} {succinct-description}.md — kebab-case description, just like the planning skill folder naming.

Step 0: Resolve the argument

Before doing anything else, classify the invocation argument and pick the mode. Do this explicitly — state the resolution to the user in one short sentence before proceeding.

1. Trim leading/trailing whitespace from the argument
2. Lower-case the trimmed argument for comparison only (keep the original for use as a topic hint)
3. Classify:
   • Empty / whitespace only → Mode 1 (Capture)
   • Exactly quiz (case-insensitive, no other words) → Mode 3 (Quiz)
   • Anything else → Mode 2 (Capture with topic hint); pass the original (un-lowercased) argument through as the hint
4. Ambiguity guard: if the argument starts with quiz followed by more text (e.g. quiz me on spring boot), treat it as Mode 3 and use the remainder as an optional topic filter for which insights to draw from. If unsure, ask the user which mode they meant before continuing.
5. Announce: tell the user which mode was picked, e.g. Mode: capture (topic hint: "spring boot dynamic properties"). Then proceed to the matching mode below.

Mode 1: Capture (no argument)

Step 1: Scan recent context for insight candidates

Look back through the conversation since the last /aha invocation in this session (search for prior aha skill activity or files written under .ai/insight/ during this session). If there's no prior invocation, scan the whole conversation.

Identify candidates. A good insight signal is any of the following:

• The agent was tasked by the user to fix a bug
• The agent was asked by the user to explain a concept
• The agent corrected a misconception the user held (e.g. "actually, X works differently than you'd expect because…")
• A non-obvious gotcha surfaced — language quirk, framework footgun, ordering dependency, race condition, off-by-one, encoding issue
• A pattern or idiom specific to a framework/language/codebase was revealed (e.g. how dynamic application properties work in Spring Boot)
• The user asked "why does X happen?" and the answer involved domain knowledge that isn't obvious from the code
• A performance characteristic was surprising (e.g. "this is O(n²) because…")
• The agent pointed out a subtle invariant that must be preserved

Ignore noise: routine code edits, things the user already clearly knew, decisions explicitly captured in a plan/ADR, knowledge the agent gathered as part of a non-troubleshooting-/research-related skill invocation (e.g. /planning), the agent's own internal reasoning.

Step 2: Propose each candidate to the user

For each candidate, ask the user whether to record it.

• If running in Claude Code: use the AskUserQuestion tool with options Record it, Skip, Refine first (one question per candidate, batched in one call only if proposing ≤4).
• Otherwise: enumerate candidates as a numbered list and ask the user which to record (e.g. "Reply with the numbers to record, or none").

For each candidate present a one-line summary so the user can decide quickly.

If the user picks Refine first, ask follow-up questions to sharpen the framing before writing.

Step 3: Draft and write each confirmed insight

For each confirmed candidate:

1. Read templates/insight.md
2. Derive {succinct-description} in kebab-case (e.g. spring-boot-dynamic-properties, pg-locks-blocking-vacuum)
3. Resolve {repo root} and create .ai/insight/ if missing
4. Write the file to {repo root}/.ai/insight/{yyyy-mm-dd} {succinct-description}.md
5. Fill every section of the template; omit a section's that would have no body
6. Make it standalone — see Standalone framing below

After writing, report the filenames back to the user.

Standalone framing

An insight must make sense to a reader who has no knowledge of the PR, conversation, or task that produced it. The capture context is not part of the record.

• Strip context-relative references. Remove phrases like "in this PR", "the change above", "as we discussed", "the bug we just fixed", "your new method". State the insight in general, timeless terms instead.
• Don't reference code the reader can't see. If you mention a symbol, file, or snippet, either inline the relevant code in the ## Example section or point to it via a permalink / path:line ref in ## Permalink. Never assume the reader can see a diff or the surrounding code.
• Prefer permalinks over working-tree paths when the code may move — a path:line ref rots; a commit-pinned permalink doesn't.
• Example code is fine even if it originated from the PR, provided the snippet is self-contained and illustrates the general point — not "the line you changed".
• Re-read test: before writing, re-read the draft as if you'd never seen the conversation. If any sentence only makes sense with that hidden context, rewrite it.

Mode 2: Capture with topic hint

If the argument is anything other than quiz, treat it as a topic hint (e.g. /aha how dynamic application properties work in Spring Boot).

• Use the hint to narrow Step 1's scan — only surface candidates relevant to that topic
• If no relevant candidate is found in context, ask the user to provide the substance directly, then draft from their answer
• Otherwise the flow is identical to Mode 1 from Step 2 onwards

Mode 3: Quiz (/aha quiz)

Step 1: Load all insights

• List every .md file under {repo root}/.ai/insight/
• If none exist, tell the user and exit
• Shuffle the file list into random order
• Initialise an in-session score: correct: 0, total: 0

Step 2: Quiz one insight at a time

For each insight in the shuffled order:

1. Read the file
2. Generate 1-3 multiple-choice questions depending on insight depth:
   • 1 question for short, single-fact insights
   • 2-3 for richer insights with multiple distinct facts, code, or trade-offs
3. Each question must have exactly 4 options, one correct, three plausible distractors
4. Distractors must be wrong but believable — not obvious throwaways. Draw from common misconceptions, adjacent concepts, or the kind of mistake the user might actually make.
5. Vary question style across the quiz:
   • "What does X do?"
   • "Why does X happen?"
   • "Which of these would break X?"
   • "What would you change to fix X?"
   • "Given this code snippet, what's the bug?"
6. Ask via the right tool:
   • Claude Code: AskUserQuestion with the 4 options
   • Otherwise: print the question and enumerate options 1., 2., 3., 4.; ask the user to reply with the number
7. After the user answers:
   • Say whether it was correct
   • Briefly explain the correct answer (one or two sentences)
   • Update in-session score and show the running tally e.g. Score: 3/4

Step 3: End of quiz

When all insights are exhausted:

• Report final score
• List any insights the user got wrong, so they know what to revisit
• Do not persist the score to disk

Critical rules

• Never write an insight without user confirmation — always propose first
• Insights must be standalone — no reader should need knowledge of the originating PR, conversation, or task to understand them; see Standalone framing
• Never overwrite or update an existing insight file (unless asked to) without explicit user approval; if the description collides, append a short disambiguating suffix
• Be concise: the value of an insight note is that it's quickly re-readable later
• Don't capture project decisions or plans — those belong in .ai/planning/ or ADRs, not here