session-reflection

Session-End Reflection

Catch blind spots and under-investigated areas before concluding a session by asking two questions:

1. Confidence audit: What am I least confident about right now?
2. Blind-spot check (Sam Altman): What's the biggest thing I'm missing about this situation? What don't I realize?

~1 in 4 sessions, one of the answers reveals a critical gap that would silently invalidate work. This skill catches those gaps at the cheapest possible moment: before the user has walked away.

Prerequisites

• A session that appears to be concluding (user signals completion, asks for summary, or starts wrap-up language)
• A project-level behavioural rule for "always conduct session-end reflection" should already be active, if this project has one -- check the project's own rules file
• For persisting uncovered findings: the context-file skill, and optionally adr-capture if a binding decision emerges
• For genuine gaps that aren't being fixed in this session: the context-file skill's known-issue document type -- this is the primary, intended source of those entries; see Workflow step 6

When to Use

• A session appears to be concluding: the user signals completion, asks for a summary, or starts wrap-up language.
• All identified tasks are marked complete and the agent is about to hand back to the user.

When NOT to Use

• During a brief query that is clearly complete (e.g., "what's the capital of France?") -- the reflection overhead is not justified.
• When the user has explicitly said "don't do the reflection this time" or similar -- honour it, don't insist.
• In automated/CI contexts -- this is a human-interactive skill only.
• In the middle of active work -- only at session-end boundaries.

Workflow

1. Detect session-end signals: "we're done", "thanks", "that's all", a request for a summary or next steps, or all identified tasks marked complete.
2. Choose the reflection mode -- inline (default, for short sessions) or sub-agent spawn (preferred for deep sessions with significant work; see Sub-Agent Spawn Pattern).
3. Initiate the reflection with a natural opening, e.g. "Before we wrap up, I'd like to do a quick reflection." Ask the two questions sequentially, never both at once -- wait for the user's response to each.
4. Question 1, confidence audit: "What am I least confident about right now?" Generate 3-7 specific items, each naming what was done, what was not verified, and why confidence is low -- not a vague feeling.
5. Question 2, blind-spot check: "What's the biggest thing I'm missing about this situation?" Target assumptions the user stated but that went unverified, alternatives not explored, and signals dropped mid-conversation.
6. Follow up. If the user flags an item, investigate before concluding. If a finding warrants preservation, use context-file. If an item is a verified, concrete gap that is NOT being fixed in this session, create a known-issue document (status: ACTIVE, severity: CRITICAL | HIGH | MEDIUM | LOW) via context-file -- this is the primary source of those entries; a reflection item that only lives in chat scrollback is the failure mode this step exists to prevent. Skip this only when the item is being fixed right now instead.
7. Conclude only once the investigation loop is resolved. Note clearly if new work was spawned.
8. Verify outcome: confirm with the user that the reflection addressed their concerns, summarize any investigation performed, confirm any known-issue file was actually created, and explicitly ask whether anything else from the reflection needs addressing.