evnchn-agentic 雙管齊下 — two pipes simultaneously (agentic reframe: generator + discriminator)
雙管齊下 (shuāng guǎn qí xià) — literally "two pipes / brushes drawn simultaneously". Classical reading: parallel-cooperative effort on the same goal. Agentic reframe: the two pipes are ADVERSARIAL — one generates the deliverable; the other actively attacks it, looking for what's wrong before reality finds it for you.
IMPORTANT: this is a deliberate reframe
The classical Chinese reading is parallel-cooperative (e.g. Tang painter Zhang Zao with a brush in each hand, painting two trees at once). The agentic reframe is GAN-style (generator + discriminator): one pipe ships the work, the other pipe simulates the critic.
The framing in plain terms: the spirit of 雙管齊下 in agent context is that the two 管 are adversarial — there's no need to teach "how to parallelize" (any decomposition skill covers that); what needs teaching is the GAN-mode discipline of don't just generate, be adversarial so as to catch issues before shipping.
The recall-test argument: when a reader scans 雙管齊下, both readings can fire. The body's load-bearing role is to anchor the agentic one before the classical one drifts in. Always lead with the adversarial framing when invoking this skill.
When this applies
Symptoms that this schema applies:
- About to declare a deliverable done — but no one has actively tried to break it
- About to ship a feature — but no one has asked "is this even the right feature?"
- About to commit a refactor — but no one has asked "could we just delete this code instead?"
- About to publish a writeup — but no one has stress-tested the argument for steelman counter-positions
- About to merge a PR — but no one has asked "what would the most adversarial reviewer in your project say about this?"
- After shipping — and you're tempted to move on without checking what surfaced
The shared shape: the deliverable exists OR is about to exist, and no internal critic has been activated yet. 雙管齊下 says: don't move on until the critic has run.
The four adversarial questions
When the generator pipe is done (or about to be), the discriminator pipe asks:
- Is X fundamentally flawed as a concept? Could the whole premise be wrong? Is there a frame from which X is obviously the wrong answer? (Concept-level critique.)
- Is X achievable with existing means / can we pull a refactor instead? Could the same outcome be reached by deleting code, reusing existing patterns, or doing nothing? (Approach-level critique.)
- Are there bugs or vulnerabilities in X? What's the failure mode? What edge case breaks it? What's the security surface? (Implementation-level critique.)
- What else about not delivering X? Open category — anything the user would think of that the generator pipe missed. (Catch-all.)
Two timings
The adversarial pipe runs at two distinct times, each with different information available:
Pre-ship (in-vacuum)
No real-world signal yet. The discriminator has to work from priors + reasoning. Useful but limited — you don't know what users / production / reality will actually do.
Apply at:
- Before opening a PR
- Before pressing the deploy button
- Before declaring a design "done"
- Before sending a piece of writing
- Before committing a change to shared branch
Post-ship (with feedback)
Real-world signal arrives — bug reports, telemetry, user feedback, CI failures, downstream agent behaviour. The discriminator now has empirical grounding, not just priors.
Apply at:
- After deploy lands
- After a PR is reviewed
- After a writeup is read
- After a feature has been in use for N days
- After a downstream agent has consumed your output
Both timings are required. Pre-ship catches concept / approach / obvious-bug issues; post-ship catches what only reality could surface. Skipping either leaves a class of problem uncaught.
Behavior
- Identify the deliverable (X). What is being generated? What's the artefact?
- Run the generator pipe to completion (or near-completion). Don't interleave; let the generator finish a coherent draft before the discriminator engages, otherwise both pipes degrade.
- Run the discriminator pipe with the four adversarial questions. Be specific — for each question, produce concrete attack-attempts, not generic worries.
- Decide based on the adversarial findings. Three outcomes:
- Discriminator found nothing real → ship
- Discriminator found small things → fix in-place, ship
- Discriminator found load-bearing things → either restart (
chengyu-retreat-is-best), repair-then-ship, or surface to the user for direction
- After ship: schedule the post-ship adversarial pass. Don't move on until reality has had a chance to test the deliverable AND you've checked what it surfaced.
Inter-link with other skills
| Skill | Surface | Difference |
|---|---|---|
chengyu-two-pipes-adversarial (this) |
INTERNAL generator-discriminator within one agent's reasoning | The critic is simulated by the same agent |
chengyu-listen-all-sides-see-clearly (兼聽則明) |
EXTERNAL multi-source review (subagent / internet / other LLMs) | The critic is a separate entity |
chengyu-think-thrice-before-acting (三思而後行) |
Meta-planning — what didn't surface in the FIRST PLAN | Process-level adversarial pass, before generation |
chengyu-create-something-from-nothing (無中生有) |
Bottom-K sampling for divergent ideation | Can be the engine of the adversarial pipe — sample low-probability attacks |
Together these four skills form the anti-overconfidence stack:
- 三思而後行: adversarial-pass the PLAN before generating
- 雙管齊下: adversarial-pass the DELIVERABLE during/after generating
- 兼聽則明: adversarial-pass via EXTERNAL sources
- 無中生有: bottom-K sampling to find low-probability attack vectors
Use the four together when stakes are high; use selectively when stakes are low.
Anti-pattern
Generator-only. Ship without ever running the discriminator pipe. The whole reason this chengyu exists is the observation that agents over-generate and under-critique.
Discriminator-only. All critique, no shipping. Skill is invoked AROUND a deliverable, not as a general anxiety pattern.
Interleaving the two pipes. Critiquing each sentence as you write it produces stilted output AND incomplete critique. Let the generator finish a coherent draft; let the discriminator have it whole.
Generic discriminator output. "There might be edge cases" / "this could have bugs" without specifics is performative discrimination. The four adversarial questions exist to force specificity.
Skipping the post-ship pass. Pre-ship catches concept / obvious-bug; post-ship catches what only reality knows. Skipping post-ship means the deliverable never gets the real test.
Confusing this with parallel execution. If you find yourself reaching for 雙管齊下 to describe "do two things in parallel", you've slipped back to the literal Chinese reading. Use chengyu-break-whole-into-parts + dispatching-parallel-agents superpower for parallel execution; this skill is specifically the adversarial reframe.
Etymology
The classical chengyu: Tang dynasty (618-907). Painter Zhang Zao (張璪) was famed for painting two trees simultaneously, one brush in each hand, both equally skilled — neither pipe was secondary, both contributed to the same painting. The classical reading is therefore parallel-cooperative excellence.
The agentic reframe: in agent work, parallel cooperation doesn't need a chengyu handle — chengyu-break-whole-into-parts + dispatching-parallel-agents cover it. What DOES need a handle is the discipline of running an adversarial critic alongside the generator. The four-character form 雙管齊下 supports the reframe well: 雙 (two) + 管 (pipes/brushes) + 齊 (together) + 下 (drawn / applied) — nothing in the literal forces the cooperative reading; the cooperative semantic is convention. The reframe takes the same characters and reads them as generator + discriminator running together. This is similar to how chengyu-create-something-from-nothing reclaims an adversarial-core chengyu for productive sampling — a deliberate, narrow reading shift adopted for the agentic context.