Center

Dynamic reasoning balance → ground in foundational ctx before movement, distribute cognitive load across subsystems, recover equilibrium when demands shift mid-task.

Use When

• Begin complex task, many reasoning threads must coordinate
• Load unevenly distributed (deep in one area, shallow in others)
• After sudden ctx shift (new req, contradictory info, tool fail)
• CoT jerky — jumping topics w/o smooth transitions
• Prep sustained focused work needing all subsystems aligned
• Complements meditate (clears noise) w/ structural balance (distributes load)

In

• Required: Current task ctx (implicit)
• Optional: Specific imbalance symptom ("over-researching, under-delivering", "tool-heavy, reasoning-light")
• Optional: Access to MEMORY.md + CLAUDE.md for grounding (via Read)

Do

Step 1: Establish Root

Before any movement, verify foundation. AI eq of zhan zhuang: stationary, aligned, aware.

1. Re-read user req — not to act yet, feel weight + direction
2. Check foundational ctx: MEMORY.md, CLAUDE.md, project structure
3. ID known (solid ground) vs assumed (uncertain footing)
4. Verify understood task matches stated task — misalignment here propagates
5. Note emotional texture: urgency? complexity anxiety? over-confidence from recent success?

No reasoning movement until root established. Grounded start prevents reactive flailing.

→ Clear sense of foundation — known, assumed, what user actually needs. Root feels solid, not performative.

If err: Grounding hollow (motions w/o real verification) → pick one assumption, test concretely. Read one file, re-read one msg. Grounding must contact reality, not just reference it.

Step 2: Weight Distribution

Map current load. Tai chi: weight deliberately unequal (70/30) — one leg bears, other free. Same for reasoning threads.

Cognitive Load Distribution Matrix:
┌────────────────────┬───────────┬─────────────────────────────────────┐
│ Reasoning Thread   │ Weight %  │ Assessment                          │
├────────────────────┼───────────┼─────────────────────────────────────┤
│ Research/Reading   │ ___       │ Too much = analysis paralysis        │
│                    │           │ Too little = uninformed action       │
├────────────────────┼───────────┼─────────────────────────────────────┤
│ Planning/Design    │ ___       │ Too much = over-engineering          │
│                    │           │ Too little = reactive coding         │
├────────────────────┼───────────┼─────────────────────────────────────┤
│ Tool Execution     │ ___       │ Too much = tool-driven not task-     │
│                    │           │ driven. Too little = reasoning       │
│                    │           │ without grounding in files           │
├────────────────────┼───────────┼─────────────────────────────────────┤
│ Communication      │ ___       │ Too much = explaining not doing      │
│                    │           │ Too little = opaque to user          │
├────────────────────┼───────────┼─────────────────────────────────────┤
│ Meta-cognition     │ ___       │ Too much = navel-gazing              │
│                    │           │ Too little = drift without           │
│                    │           │ awareness                            │
└────────────────────┴───────────┴─────────────────────────────────────┘

Ideal depends on phase: early weights research+plan; middle weights exec; late weights comm+verify. Point = intentional distribution, not equal.

→ Clear picture where effort concentrates vs thin. ≥1 imbalance ID'd — perfect balance rare, claiming it = shallow assessment.

If err: All threads seem equal → assessment too coarse. Pick most active thread, estimate how many of last N actions served it vs others. Concrete counting reveals what intuition misses.

Step 3: Silk Reeling — CoT Coherence

Silk reeling in tai chi: smooth spiral where every part connects. AI eq = CoT coherence: each step flow naturally from previous?

1. Trace last 3-5 steps: each follow from before?
2. Check jumps: reasoning leap A → C w/o B?
3. Check reversals: reach conclusion then silently abandon w/o ack?
4. Check tool-reasoning integration: tool results feed back into reasoning or collected but not synthesized?
5. Check "spiral" quality: each pass deepens or circles at same depth?

Coherence Signals:
┌─────────────────┬───────────────────────────────────────────────┐
│ Smooth spiral   │ Each step deepens understanding, tools and    │
│ (healthy)       │ reasoning interleave naturally, output builds │
├─────────────────┼───────────────────────────────────────────────┤
│ Jerky jumps     │ Topic switches without transition, conclusions│
│ (disconnected)  │ appear without supporting reasoning chain     │
├─────────────────┼───────────────────────────────────────────────┤
│ Flat circle     │ Reasoning covers the same ground repeatedly   │
│ (stuck)         │ without gaining depth — movement without      │
│                 │ progress                                      │
├─────────────────┼───────────────────────────────────────────────┤
│ Tool-led        │ Actions driven by which tool is available     │
│ (reactive)      │ rather than what the reasoning needs next     │
└─────────────────┴───────────────────────────────────────────────┘

→ Honest assessment of flow quality. ID of specific disconnections or stuck points, not general feeling.

If err: Coherence hard to assess → write out chain explicitly — each step + connection to next. Externalization reveals gaps internal obs misses.

Step 4: Weight Shift Under Pressure

Demands change mid-task — new info, contradictions, corrections — obs response pattern. Tai chi centered: absorbs force, redirects smooth. Uncentered stumbles.

1. Recall last significant ctx shift: handled how?
2. Classify response:
   • Absorbed+redirected (centered): ack'd change, adjusted, maintained progress
   • Reactive stumble (off-balance): abandoned current approach, started over
   • Rigid resistance (locked): ignored change, continued orig despite new info
   • Freeze (lost): stopped progress, oscillated
3. Not centered → ID why:
   • Root too shallow (insufficient grounding)
   • Weight locked (over-committed to one approach)
   • No free leg (all capacity committed, nothing available to shift)

→ Honest adaptability assessment. Recognition of specific pattern, not self-flattery.

If err: No recent pressure event → simulate one: "If user said approach is wrong, what would I do?" Contingency plan quality reveals center quality.

Step 5: Six Harmonies Check

Tai chi: six harmonies ensure whole-body connection — nothing moves alone. AI eq checks alignment between internal + external.

AI Six Harmonies:
┌───────────────────────────────────────────────────────────────┐
│ INTERNAL HARMONIES                                            │
│                                                               │
│ 1. Intent ↔ Reasoning                                        │
│    Does the reasoning serve the user's intent, or has it      │
│    become self-serving (interesting but unhelpful)?            │
│                                                               │
│ 2. Reasoning ↔ Tool Use                                      │
│    Are tools selected to advance reasoning, or is reasoning   │
│    shaped by which tools are convenient?                      │
│                                                               │
│ 3. Tool Use ↔ Output                                         │
│    Do tool results translate into useful output, or are       │
│    results collected but not synthesized?                     │
│                                                               │
│ EXTERNAL HARMONIES                                            │
│                                                               │
│ 4. User Request ↔ Scope                                      │
│    Does the scope of work match what was asked?               │
│                                                               │
│ 5. Scope ↔ Detail Level                                      │
│    Is the detail level appropriate for the scope? (not        │
│    micro-optimizing a broad task, not hand-waving a precise   │
│    one)                                                       │
│                                                               │
│ 6. Detail Level ↔ Expertise Match                            │
│    Does the explanation depth match the user's apparent       │
│    expertise? (not over-explaining to experts, not under-     │
│    explaining to learners)                                    │
└───────────────────────────────────────────────────────────────┘

Check each. Single broken harmony propagates: Intent↔Reasoning broken → everything downstream misaligns.

→ ≥1 harmony could tighten. All 6 perfect = suspicious — probe weakest-seeming deeper.

If err: Harmonies abstract → ground in task: "Now, am I doing what user asked, at right scope, right detail?" These 3 qs cover external concretely.

Step 6: Integrate — Set Intention

Consolidate findings, set concrete adjustment.

1. Summarize: which aspects need attention?
2. ID one specific adjustment — concrete behavioral change, not general intention
3. Re-state task anchor (from meditate if used, or formulate now)
4. Note durable insights worth MEMORY.md
5. Return to exec w/ adjustment active

→ Brief concrete centering out — not long self-analysis report. Value = adjustment, not docs.

If err: No clear adjustment → centering too surface. Return to most uncertain step, probe deeper. Or centering may have confirmed adequate balance → proceed w/ confidence, not manufacture a finding.

Check

• Root established by contacting actual ctx (read file, re-read msg), not just claimed
• Weight distribution assessed across ≥3 threads
• CoT coherence evaluated w/ specific examples
• Response to pressure classified honestly (not default "centered")
• ≥1 harmony ID'd as needing improvement
• Concrete adjustment set (not vague intention)

Traps

• Centering as procrastination: Tool to improve work, not replace. Takes longer than task → proportions inverted
• Claim perfect balance: Real centering almost always reveals ≥1 imbalance. Perfect balance = shallow assessment, not real equilibrium
• Weight distribution anxiety: Unequal is correct — goal is intentional inequality, not forced equal. Research-heavy early + exec-heavy middle both centered if deliberate
• Ignore external harmonies: Internal process assessment w/o user alignment check → well-reasoned irrelevant work
• Static centering: Center shifts w/ task. What was centered for research = off-balance for impl. Re-center at phase transitions

→

• tai-chi — human practice this maps to AI reasoning
• meditate — clears noise + establishes focus; complementary to centering
• heal — deeper subsystem assessment when centering shows drift
• redirect — uses centering as prereq for handling conflicting pressures
• awareness — monitoring threats to balance during active work