Theory of Constraints (TOC)
Core principle: Every system has exactly one constraint limiting its throughput at any moment. Improving anything that is *not* the constraint is waste. Find the constraint, exploit it, then repeat.
The 5 Focusing Steps
Step 1: Identify the Constraint
Find the single bottleneck — the resource, process, or step with the least capacity relative to demand.
How to find it:
- Where does work pile up? (queue buildup = upstream of the constraint)
- Where is there idle time? (downstream of the constraint — starved for input)
- What does everyone wait on? (the constraint is often "that one person" or "that one step")
- What's the longest step in the value stream?
Signal: *The constraint is where WIP accumulates and where delays originate.*
Step 2: Exploit the Constraint
Get maximum throughput from the constraint without spending money or making major changes first.
Tactics:
- Eliminate waste at the constraint (don't let it sit idle, don't let it process low-value work)
- Protect it with a buffer upstream (so it's never starved)
- Remove it from non-constraint work (free it up to do only what only it can do)
- Reduce defects feeding into it (rework at the constraint is doubly expensive)
Step 3: Subordinate Everything Else
Make all non-constraint steps serve the constraint, not optimize themselves.
This is counterintuitive: a non-constraint running at 100% is harmful if it floods the constraint with WIP.
Key shift: Stop measuring local efficiency. Measure constraint throughput as the system metric.
Step 4: Elevate the Constraint
If steps 2–3 aren't enough, now invest: add capacity, add people, add tooling — but only at the constraint.
Step 5: Repeat
Once the constraint is resolved, it moves. Find the new one. Never let inertia become the constraint.
Output Format
🔍 Constraint Identification
- Identified constraint: [Name the specific step, role, resource, or decision]
- Evidence: What signals point to this being the constraint?
- WIP accumulation point: Where does work pile up?
- Downstream starvation: What's blocked waiting for the constraint?
📊 Throughput Analysis
- Current flow: [Input → Step A → Step B →... → Output]
- Constraint location in the flow
- Estimated throughput loss due to constraint
⚡ Exploit Actions (no-cost first)
- What waste can be removed at the constraint immediately?
- What low-value work can be offloaded from the constraint?
- What buffers should be added upstream?
🔄 Subordination Changes
- Which non-constraint steps are currently "optimizing locally" and harming system throughput?
- What should they slow down or stop doing?
📈 Elevation Options (if needed)
- Investment options to increase constraint capacity
- Cost/benefit relative to throughput gain
⚠️ False Bottlenecks to Avoid
- Which steps *look* slow but are actually downstream of the real constraint?
- What "improvements" would be wasted effort?
Common Constraint Types
| Type | Example | Signal |
|---|---|---|
| Capacity constraint | One engineer reviews all PRs | PRs queue up, reviewer is always busy |
| Knowledge constraint | Only one person knows the system | Everything waits on that person |
| Decision constraint | Approvals bottleneck execution | Teams idle waiting for sign-off |
| Handoff constraint | Work crosses team boundaries | Long wait times at team interfaces |
| Policy constraint | Rules prevent fast action | Work is done but can't ship |
| Market constraint | Demand is the limit | System has capacity but no demand |
Key Mental Shifts
- Don't: Optimize every step. Do: Protect and exploit the constraint.
- Don't: Measure local efficiency. Do: Measure global throughput.
- Don't: Start with investment. Do: Exploit first, then elevate.
- Don't: Assume the constraint is fixed. Do: Expect it to move after you fix it.