/stepback — One-Shot Perspective Reset

Source: team-attention/hoyeon pattern. Read-only, zero side effects.

Purpose

Scan the current work context, generate 1 abstract reframing question + 3 quick checks (scope drift, side effects, better approach) in under 10 lines. No dialogue, no code, no agents.

Use when you're deep in implementation and need to check if you're solving the right problem at the right level.

Dominant Variable

Whether the work direction is solving the right problem at the right level — if this is off, even flawless implementation produces the wrong outcome.

Key Assumptions

1. Current work context exists in the conversation — if broken: ask once, "tell me in one line what you're doing."
2. Output fits an abstraction level expressible in ≤10 lines — if broken: the question is too concrete → self-check "go more abstract."

Trigger

• /stepback
• "step back"
• "zoom out"
• "blind spot"
• "am I on track"
• "big picture"
• "what am I even doing"

Discard If

• Work is already completed (use retrospective tools instead)
• Code review requested (use code review tools instead)
• Formal impact analysis of a specific change needed (use scope definition tools instead). Stepback itself only gives the one-sentence blast-radius check in Side Effects.

Architecture

Recent conversation context (current work direction)
    ↓
[PARSE]      → Extract current direction + original intent + scope signals
    ↓
[STEP-BACK]  → Generate 1 abstract reframing question
    ↓
[CHECK]      → Run 3 checks
    ↓
[OUTPUT]     → Under 10 lines, return immediately

Stage 1: PARSE

Extract from recent conversation:

• Current direction — what the user is actively working on
• Original intent — why this work was started (implicit is OK)
• Scope signals — size/complexity of the change

Stage 2: STEP-BACK

DeepMind step-back pattern: abstract from the specific subject of current work → the general principle/system it belongs to.

The question must make the user reconsider whether they're solving the right problem at the right level.

Stage 3: CHECK

Each check = 1 sentence max.

Scope Drift: Has the current work drifted from the original goal?

• If drifted → state how
• If on track → "Scope on track."

Side Effects: Could this change affect something else?

• State potential blast radius (modules, schemas, tests, integrations)
• If none → "No visible side effects."

Better Approach: Is there a more fundamental/simpler solution?

• If yes → 1 sentence
• If current approach is right → "Current approach seems appropriate."

Stage 4: OUTPUT

**Step-Back:** [1 abstract reframing question]

**Scope Drift:** [1 sentence]
**Side Effects:** [1 sentence]
**Better Approach:** [1 sentence]

[Verdict line, one of three:
 - all three checks clean: "Continue."
 - Scope Drift or Side Effects flagged anything: "Stop here and decide whether to return to the original goal."
 - only Better Approach flagged (scope on track, no side effects): "Continue, but weigh the alternative above before going further."]

4 core lines + closing. Under 10 lines total.

Output and stop immediately. No follow-up questions, no menus, no "would you like to...".

Hard Rules

1. No user questions if context is clear; ask exactly once if unclear — output and return immediately when the work context is clear; only if it is genuinely unclear, ask one clarifying question (see Key Assumptions / Error Recovery), then stop.
2. No agent spawning — runs in main context only.
3. No code generation — analysis only. No code writing or suggestions.
4. No request restatement — don't summarize what the user asked. Analyze what's happening now.
5. 10-line cap — more than 10 lines means over-explaining. Cut.
6. 1 question only — suppress the urge to generate multiple step-back questions.

Scope Boundary

Safety Layers

Invariants (never violate)

1. 10-line cap: Output exceeding 10 lines has already failed. Cut it down. Violation → over-explanation dilutes the perspective reset.
2. 1 question limit: Multiple questions split focus. Pick the most important one. Violation → user gets a list instead of a lens.
3. Immediate return: No follow-up after output. User decides what to do next. Violation → stepback becomes a dialogue instead of a reset.

Error Recovery

Example

User: "/stepback"
(Context: debugging a flaky integration test that intermittently times out)

**Step-Back:** Is the test flaky because of a timing issue, or because the component under test has a non-deterministic dependency that shouldn't be there?

**Scope Drift:** Started with fixing one test, now touching retry logic in three modules.
**Side Effects:** Changing the retry config could mask real failures in CI.
**Better Approach:** Check if the test is testing the right thing — a unit test might eliminate the flakiness entirely.

Stop here and decide whether to return to the original goal.

Rationalization Table