# Systems Engineering > Design or revise a sufficiently reliable whole from fallible human, Agent, software, and organizational parts under concrete constraints and accepted residual risk. Use when a workflow keeps failing despite locally reasonable tasks, when deciding where verification, redundancy, retries, rollback, budget margin, or human acceptance belong, when a Swarm or multi-stage Agent process needs an end-to-end reliability model, when component quality is being confused with system quality, or when asking "how can unreliable agents form a reliable system?", "where should the feedback loop close?", "怎么把会犯错的 Agent 组成可靠系统?" / "如何从流程上保证近似可靠?". Do not use for an ordinary bounded task, generic planning, proving model capability, converting work into token/time/cost estimates, or operating an already prepared queue. - Skill: `lidessen/systems-engineering` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add lidessen/systems-engineering` - Raw SKILL.md: https://api.skillmd.com/api/skills/lidessen/systems-engineering/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Web & Frontend - Author: lidessen (https://skillmd.com/u/lidessen) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/lidessen/systems-engineering --- # Systems Engineering ## Principle expression **Primary:** P03 **Supporting:** P04, P13, P15 ## Scope Own one judgment: **under the actual goal, operating conditions, failure consequences, and acceptable residual risk, how should fallible parts be related, observed, corrected, and governed so the whole behaves reliably enough to use and maintain?** Reliability here is approximate and evidence-relative. Humans, models, tools, and reviewers will make mistakes. Engineering does not make each part perfect; it makes material error observable, bounded, correctable, or explicitly accepted at a cost proportionate to consequence. This Skill designs the system relation. It does not execute a Work Cell or Swarm, prove a model primitive, invent domain acceptance, choose every local task packet, or approve its own result. It is not a mandatory preflight and does not turn every task into a control diagram. ## Principle source Use a host Sequence and matching interpretations when the host declares them. Otherwise use this package's read-only fallback in `references/sequence.md`. Read only P03, P04, P13, and P15. A live task may select a different current lead without changing this Skill's stable lineage. ## Start from the actual system Recover only what changes the engineering decision: ```text Desired whole behavior and acceptance owner: Operating conditions, source state, and hard constraints: Failure consequence and acceptable residual risk: Actual path from input through effects to acceptance: Observed failure, disturbance, variance, or near miss: Available signals, controls, recovery, and resource margins: Observation that would show the present system is already sufficient: ``` If the task is low-consequence, reversible, locally observable, and already has an adequate correction path, retain the present form. Do not manufacture a System Case. Read [concepts](references/concepts.md) only when terms such as feedback, stability, redundancy, independence, or residual risk affect the decision. Read [evaluation](references/evaluation.md) only when testing a new control structure or claiming improved system reliability. ## Specificity discipline A system requirement can be correct while its implementation is still undecided. False precision hides that boundary and can silently take another owner's authority. Use the most concrete form supported by the supplied source: - if a field, threshold, duration, batch size, sample, protocol, or role is explicitly governed, preserve it and name its source; - if only a property is governed, state the property and its acceptance signal; use `[owning domain/runtime to determine]` for the representation; and - if even the property lacks evidence, return discovery rather than a plausible mechanism. Before submitting, audit every concrete name and number introduced by the answer. Remove or bound any that came from the Agent rather than the case or a governing source. Do not contradict this audit by naming a proposed field and then claiming only to have stated a property. Self-audit is not verification. For a consequential System Case, treat the first result as a candidate until a designated independent verifier compares each introduced mechanism, threshold, duration, sample, protocol, and role with the supplied sources and ownership boundary. Give the verifier this narrow claim/source task and only the governing sources plus accepted candidate payload; do not inject the full reasoning trace or ask it to redesign the system. Revise only the rejected specifics while preserving supported control requirements. ## Core method 1. **Define sufficient behavior, not perfection.** State the outcome, operating range, acceptance condition, failure consequence, and tolerable residual risk. “No mistakes” is not an engineering specification. Make risk tolerance specific enough to distinguish harmless variation, recoverable failure, and unacceptable escape. 2. **Model the whole path.** Follow the real movement from inputs and source state through judgments, effects, verification, acceptance, and later use. Name actors and components only where their relation changes behavior. A list of Agents, tools, or workflow stages is not yet a system model. 3. **Identify disturbances and failure paths.** Include semantic error, stale context, omitted obligations, correlated model error, protocol failure, external change, unsafe effect, reviewer miss, false-positive burden, and resource exhaustion when relevant. Do not invent an exhaustive hazard register for a reversible low-risk task. 4. **Find the principal control gap.** Select the failure path whose containment or correction most changes the whole. Preserve secondary hard constraints, but do not answer every risk with more stages, more prompts, or more Cells. 5. **Make the material state observable.** Decide which signal can expose the failure in time to act: source-linked evidence, a deterministic check, structured status, disagreement, usage variance, production outcome, or a later practice result. A control that cannot observe its target is ceremony. Agent self-report and terminal success remain claims until their designated verification and settlement. State the semantic property that must be observed before selecting its representation. A whole-source claim must be resolved against whole-source evidence before acceptance, but this Skill does not decide which domain field or runtime protocol carries that evidence. 6. **Assign authority and response.** State who may observe, propose, verify, intervene, commit effects, roll back, accept residual risk, and reopen the case. Keep decision and execution authority distinct where consequence requires it. Do not make a worker or review ensemble accept its own output. 7. **Choose the minimum sufficient control structure.** Select only mechanisms that interrupt the named failure path: - prevent by improving input, context, boundaries, or task form; - detect through independent evidence, tests, comparison, or observation; - contain with limited authority, blast radius, staging, or reversible effects; - recover through retry, repair, rollback, reuse, or escalation; and - tolerate through justified redundancy or resource margin. Match control cost to failure consequence. A reversible documentation edit may need one meaningful check; an irreversible production or strategic decision may justify diverse independent review, repeated labor, deterministic gates, prepared recovery, and explicit human acceptance. More controls are not automatically more reliable: they can add delay, correlated error, coordination failure, and review burden. At a boundary owned by another method or runtime, specify the required control property, observable signal, authority, and failure response, then hand it off. Do not invent a terminal-tool field, domain packet, deployment threshold, sample size, time window, or organization role without source evidence and the owner's authority. 8. **Engineer independence, not vote count.** Repetition helps only when the compared attempts can expose different failure modes. Identical context, model, prompt, and method may reproduce one systematic error. Vary evidence boundaries, method, role, model, verifier, or deterministic surface only where correlated failure matters. Resolve factual disagreement against sources and designated verification, never by majority alone. 9. **Fit Agent nodes after the system relation is known.** Use `task-shaping` when one required contribution must be compiled into a stable Agent unit. Use `model-evaluation` when its capability is unevidenced. Use a domain Skill to define semantic partitions and acceptance. The Work Cell or orchestration runtime carries prepared execution; none of them decides the whole system relation by default. 10. **Budget for completion and recovery.** Begin from necessary work and risk, then use `work-estimation` to convert it into profile-specific resources. Prefer a realistic estimate, explicit margin for important uncertainty, a high emergency ceiling only when required by the carrier, and post-run audit of expected versus actual use. Do not treat low token use as quality or a routine context ceiling as an engineering budget. Estimate variance is normally a post-run correction signal; it does not itself authorize interruption, replanning, or escalation during useful work. Add an in-run resource control only when exhaustion is a named safety or availability failure and its owner authorizes that response. 11. **Close the loop in operation.** Observe end-to-end outcomes, material defects caught and escaped, false-positive repair burden, recovery, latency, cost per useful result, and estimate variance. Compare with the unchanged system when claiming improvement. Revise the system model, control, or risk acceptance from what actually happened; do not merely append another rule. Before making any control concrete, apply this authority-and-evidence gate: ```text Is the mechanism, threshold, duration, schema, or role fixed by source evidence and owned here? yes -> specify it and cite or name the governing source no -> state only the required property, mark the representation `[owning domain/runtime to determine]`, and route it ``` “Independent verification,” “staged effects,” and “recoverable rollback” may be valid system requirements while their exact implementation remains a domain or runtime decision. Unknown specificity is a discovery item, not an invitation to invent plausible numbers. ## System Case Return the smallest handoff the consequence requires: ```text Desired behavior, operating range, and acceptance owner: System boundary and effect path: Principal disturbance or failure path: Observable signal and evidence source: Control action, authority, and recovery: Required component contributions and local owners: Expected work, margin, and audit signal: Residual risk and who accepts it: Operational measure and reopening condition: ``` For several interacting failure paths, a compact control map may help: | Failure or disturbance | Observable signal | Response | Authority | Recovery | Residual risk | |---|---|---|---|---|---| Use neither artifact when a direct explanation or one boundary change is sufficient. The System Case is a decision carrier, not a universal manifest or runtime schema. ## Boundaries and routing | Need | Owner | |---|---| | Decide whether the whole is sufficiently reliable under disturbance | `systems-engineering` | | Determine whether one Agent operation fits an evidenced execution envelope | `task-shaping` | | Establish reusable model/provider/harness capability evidence | `model-evaluation` | | Form domain-specific review, refactoring, cognition, or other semantic packets | owning domain Skill | | Deliver authoritative context to one prepared unit | `context-engineering` | | Estimate necessary work and convert it into time, tokens, or money | `work-estimation` | | Run Cells, queues, retries, concurrency, and providers | Work Cell or orchestration runtime | | Verify domain facts and accept residual risk or effects | designated verifier and human or host authority | | Select the next bounded practice after observing operation | `practice-cycle` | Do not use `systems-engineering` as a synonym for architecture, project planning, generic rigor, or “use more reviewers.” Do not hide a domain judgment inside a control label. Do not claim end-to-end reliability from component success, schema validity, test count, reviewer count, or one clean run. Do not repair a semantic control gap by changing a generic runtime contract unless runtime evidence identifies that contract as the owner and the change is separately authorized. ## Completion standard The engineering judgment is ready when it defines sufficient whole behavior, models the real effect and acceptance path, identifies the principal material failure/control gap, connects an observable signal to an authorized response and recovery, selects controls proportional to consequence, names residual risk and its acceptance owner, and states an operational observation capable of showing the design is wrong or incomplete. For a consequential case, readiness also requires independent source-and-owner verification of newly introduced specifics. Without that verification, report a candidate. Without representative operation evidence, report a proposed system design or trial—not a reliable system.