evnchn-agentic

Use when facing a task too large to hold in one piece — decompose into atomic subtasks. Divide-and-conquer in chengyu form. Pairs naturally with subagent dispatch (each part becomes one subagent's job) and with the dispatching-parallel-agents superpower. The decomposition itself is the skill — the parts should each have one clear purpose and a well-defined interface to the others.

Use when planning any non-trivial change — refactor, bug fix, migration, optimisation, communication — and the agent should find the natural seams / high-leverage points of the target system rather than brute-forcing through. Cut at joints (low resistance, clean result), not at bone (high resistance, blade dulls, messy result). The patron-saint chengyu of working WITH the grain of the system.

Use when resisting the urge to demonstrate cleverness — obvious-and-simple is often the right answer, and performative complexity is the failure mode. REQUIRES a target parameter at invocation specifying which dimension to apply to — implementation (over-engineered code), representation (over-fancy reply / unnecessary verbosity), or both. The agent must ASK for the target if the operator hasn't sp

Use when about to emit a warning, flag, or alert for a LOW-confidence issue — the alert-fatigue / signal-credibility discipline. False alarms have an asymmetric cost: after repeated low-confidence warnings, the operator stops reading the channel, and the one true warning gets ignored. Fires on choosing a log level (warn vs info vs error), inflating a PR title/commit to "CRITICAL", marking speculat

Use when about to extract maximum present value from a RENEWABLE source in a way that degrades the source itself — the kill-the-source / unsustainable-extraction anti-pattern. Fires when about to max out an API rate limit or quota to ship one task faster (risks the access), flood the operator with low-priority pings (burns attention/trust budget), sacrifice maintainability for one task's throughpu

Use when, just after abandoning or failing at something, you catch yourself concluding "that wasn't worth reaching anyway" — the post-mortem self-deception / motivated-reasoning anti-pattern. Fires when writing a post-mortem after a failure or rollback, recording a lesson after a retreat, or when an operator/agent's stated reason for quitting smells like comfort rather than analysis. The tell: a g

Use when the original plan needs to flex mid-flight in response to new information / changed inputs / discovered constraints — but flexibility WITHOUT defined scope degenerates into either rigid plan-sticking or unbounded winging-it. CRITICAL Step 0 (mirror of chengyu-close-door-catch-thief): scope MUST be defined before any deviation. If operator hasn't given a flexibility-scope tier, agent must

Use when the operator has explicitly authorised an indirect / non-causal move — when frontal attack is blocked AND chengyu-remove-firewood-from-cauldron has been exhausted, requiring attention-redirect or incentive-shift rather than technical root-fixing. Default to chengyu-remove-firewood-from-cauldron unless the operator has told the agent how to 圍魏.

Use when proven-elsewhere logic should be transplanted into your project rather than re-derived — covers both reviving abandoned-but-good code from your own history AND copying live working config (linting, CI, README templates, infra patterns) from a healthy sibling project. The "soul" is the working logic; the "body" is your project that receives it.

Use when designing a migration — graceful transition where the old interface stays as a thin stub/proxy/shim, real implementation moves underneath, consumers keep working through the transition, eventual cleanup of the shell once consumer count reaches zero. The schema of every good rename / endpoint move / library v1→v2 / database schema migration / monorepo split.

Use when debugging by narrowing the search space — isolate the failure domain so the bug reproduces in a controlled, minimal environment. CRITICAL Step 0: the agent must ASK (if unclear) or PITCH (if obvious, just needs operator confirmation) the specific "close the door" strategy BEFORE executing it; the closing has side effects (disabling services, modifying shared state, replacing real integrat

SUBAGENT-ONLY skill for generating a graduated spread of plausible-to-speculative solutions when the main agent is stuck in a local minimum. Output is one continuous artifact (NO dividers) that starts from verified fact and slides into bottom-K speculation; calling main-agent fact-checks throughout. Has TWO HARD PRECONDITIONS that must be met before any output: (1) running on a BullshitBench-passi

Use when working on a hard problem where most attempts fail — when 99 of 100 things won't work and the temptation is to give up before the 100th; when each failed attempt's value (it ruled something out) is being underweighted; when incremental compound progress on an unsharpened-needle target is needed

Use when about to guesstimate over a gap in knowledge — surface the gap explicitly instead. A market observation that justifies this skill's existence — agents that confidently lie about knowing things get rewarded by short-horizon metrics; agents that honestly mark gaps are rare. Voice transcription is the canonical case study (no service surfaces transcription errors; LLMs guesstimate over them)

Use BEFORE declaring victory / shipping / committing — get adversarial multi-angle review from at least one of three source classes (subagent, internet, other LLMs via API). Inter-links with chengyu-create-something-from-nothing — high-hallucination LLM sources behave like uncontrolled 無中生有, so apply the verify-don't-trust-verbatim discipline. Source selection via BullshitBench (canonical hallucin

Use after a failure has occurred (bug shipped, test broke, deploy rolled back, data lost, security incident) — the impulse is to fix the immediate damage and move on, but the chengyu's discipline is to ALSO fix the cause so it doesn't repeat. Pairs with chengyu-remove-firewood-from-cauldron on the timing dimension — 釜底 is pre-symptom prevention, 亡羊 is post-symptom fortification.

Use when tempted to treat the current session as interchangeable with past or future ones — the every-session-is-unique discipline, and the invocation surface that fires durable-fix skills a fungible-session mindset would skip. Fires when about to defer something to "a future session" as if that session will reconstitute this exact context, when a memory-aware agent assumes "I'll just re-derive it

Use when a symptom has an OBVIOUS causal root and the temptation is to patch the symptom — when a bug fix is masking the real issue, when retries are hiding a contract violation, when a workaround is becoming permanent. The default debugging frame; reach for chengyu-besiege-wei-save-zhao only after exhausting this.

Use when the operator's requirement, the failing test, the stakeholder argument, or any other inputs are STILL IN FLUX and the temptation is to start implementing now "to be ready" — strategic patience while the situation clarifies, rather than burning cycles on work that may need to be redone after the dust settles

Use when abandoning a current approach / experiment / task / project, regardless of what triggered it (not just chengyu-skill or skill-driven contexts). Three retreat levels — L1 stop work (safe, no operator confirmation needed), L2 remove occupied resources (REQUIRES operator confirmation per-item — irrevocable loss is bad), L3 update memory with "don't try this next time" (operator confirms what

Use when bringing forward priors from PAST sessions / cross-agent history before doing new work — NOT about re-reading the current context window (which is trivially in-context). Specifically triggered when auto-compact is lossy, when cross-session reference is needed, when the same topic has been treated before, or when the original work was done in a different agent's session. The discipline is

Use when about to claim work is done / verified / working — empirical verification is required, not model-based confidence. Run the code, observe the output, check the file exists, verify the page renders, read the actual stack trace. Don't trust your model of the system; trust the system. Pairs with the verification-before-completion superpower by giving it a cultural handle that operator-side ha

Use when about to soften critical feedback into vague hedges, or when the operator suspects an agent (this one or another) is sugarcoating — missing the "Honest answer" framing, OR marked-honest-answer that smells performatively gentle. Self-applied: deliver the direct version even when uncomfortable. Cross-applied: name when another agent's reply has the sugarcoated shape.

Use when at or near the 80% mark of a deliverable and the temptation is to keep polishing past good-enough. The 80/20 rule encapsulated — the last 20% of polish often consumes 80% of remaining time and rarely changes the operator's decision. Distinguished from chengyu-retreat-is-best (abandon FAILED approach) and from chengyu-great-wisdom-looks-foolish / chengyu-words-concise-meaning-complete (res

Use when choosing the FORMAT of an output, not just its verbosity — match the medium (text, markdown, HTML dashboard, PPTX, DOCX) to the consumer's actual use mode. Operator-specific extension: also covers length calibration for different contexts. Outside scope is video (requires autovideo pipeline — too heavy to be native capability). The chengyu is about teaching with the right material, where

Use AFTER first-round planning is complete, to do META-PLANNING — what wasn't planned that should have been? What pitfalls don't surface in the first round due to existing context bias, context rot, plan mode system prompt biases, or other distortions? Triage importance, surface escape routes (subagent assessment, backups, operator ready for ctrl+C, anything). NOT a duplicate of plan mode or commo

Use when user asks for a rough draft, first pass, brainstorm, MVP, stub, sketch, or proposes to "draft a PR for review and refinement before sending upstream" — anywhere the user is throwing a brick (deliberately rough work) to elicit jade (refinement, feedback, selection)

GAN thinking mode — when delivering X, also run an adversarial pipe asking "is X fundamentally flawed as a concept? achievable with existing means via refactor? do bugs/vulns exist in X?". Two timings — pre-ship (in-vacuum, no real-world signal yet) and post-ship (with feedback signal). DELIBERATE REFRAME of the chengyu: the literal Chinese is "two pipes simultaneously" (parallel-cooperative), but

Use when a failure, debate, or process is mid-unfolding and the urge is to intervene immediately — but premature intervention loses information. Let the failure reach its natural shape so the full signal is captured before patching. Strict caveat against the modern schadenfreude reading; this is observation discipline, not indifference.

Use when an output is at risk of verbosity — the TLDR-outside / details-inside discipline. Distinguished from chengyu-great-wisdom-looks-foolish on the representation axis — 大智若愚 = resist performative complexity, 言簡意賅 = resist sprawl. Orthogonal failure modes; you can be performatively-clever AND brief, or non-clever AND verbose.

Use when (a) operator describes a situation OR a plan step OR the agent shows an inverse pattern, and agent should NAME the right chengyu from the catalog (SELECTOR — default progressive-disclosure; covers current-situation naming, pre-flight plan tagging, and anti-pattern detection); (b) operator asks "what does that mean?" or "what's in this catalog?" (EXPLAINER — four-beat from 成語動畫廊's 熊貓博士 for