走為上計 — retreat is the best stratagem
走為上計 (zǒu wéi shàng jì) — "retreat is the best stratagem". Most famous of the Thirty-Six Stratagems (#36, last and most cited). The discipline of knowing when to bail — agents over-persist by default; this skill is the explicit counter-force.
Pairs with chengyu-grind-pestle-to-needle as opposite-and-complementary: 鐵杵磨針 says "keep grinding when the path is real but the result is far"; 走為上計 says "stop grinding when the path itself is wrong".
When this applies
Symptoms that this schema applies (NOT just chengyu-skill or skill-driven contexts — generic bail-out):
- A debugging session has tried 10 things, none worked, and the next 10 won't either — the approach is wrong, not the count
- An experiment was based on an assumption that has now been falsified — continuing would compound the wrong inference
- A feature implementation is bigger than expected AND the requirement that drove it has weakened — retreat is cheaper than completion
- A project / fork / branch has drifted away from anything useful — bail before sunk cost gets larger
- The session itself is corrupted (context rot, repeated failures, model getting stuck in a loop) — restart in a fresh session
The fact that bailing FEELS bad is not signal that it's wrong. Sunk cost is invariant under future decisions.
The three retreat levels
| Level | What it means | Operator confirmation | |---|---|---| | L1 | Stop work. No further changes, no resource cleanup. The current state stays as-is. | Not required — safe default | | L2 | Remove occupied resources — kill processes, delete temp files, clean up branches, remove infra, etc. | REQUIRED per-item — irrevocable loss is bad | | L3 | Update memory with "don't try this next time" so the lesson survives the session boundary. | REQUIRED for content — operator confirms what gets recorded |
Behavior
- State the bail explicitly. "Retreating from this approach. Reasons: [X, Y]." Don't drift into something else without naming the bail.
- Pick the retreat level. Default to L1 unless something obviously needs to be cleaned up.
- For L2: ASK or PITCH each resource removal. Mirror of chengyu-close-door-catch-thief's Step 0 pattern — "I want to delete the staging branch / kill the training run / remove the temp directory at X. Confirm before I do it." Per-item, not batched, because each is irrevocable.
- For L3: PITCH the lesson text to the operator. "I'd record: 'Don't try [approach] for [problem-shape] because [reason].' Approve, edit, or skip." Memory entries that survive across sessions are durable; get the wording right.
- Don't retreat further than the operator authorised. L1 doesn't escalate to L2 unilaterally; L2 doesn't escalate to L3.
Anti-pattern
Silent resource cleanup. Deleting branches / killing processes / removing files without operator confirmation. Catastrophic when wrong; very hard to recover from.
Not marking the lesson. Bailing without L3 means the same wrong approach gets retried next session because there's no record. Compound wasted work.
Bail-out as procrastination. Retreating from work that's actually clear because it's hard. The schema applies when the approach is wrong, not when you don't feel like continuing. Test: would a clean restart actually go differently, or just postpone the same difficulty?
Heroic last-stand grinding. Refusing to bail because "I'm so close" when in fact you're not. Pair this skill with chengyu-grind-pestle-to-needle's "if attempt N is just attempt N-1 with no new information, stop" check.
Etymology
Thirty-Six Stratagems #36, the famous last. The full saying in modern usage: 三十六計，走為上計 — "of the thirty-six stratagems, retreat is the best." The classical military reading: when outmatched, an orderly retreat preserves the army for future engagement; a destroyed army loses everything. The agentic translation is exact — preserving cognitive resources (operator attention, agent context, time budget) for the next approach is worth more than the sunk cost of the current one.