Design or modernize a software system from an evidence-backed current-state inventory through quality attributes, architecture options, API and schema contracts, migration/rollback plans, ADRs, and verification. Use for greenfield or existing monoliths, modular monoliths, services, system/API/ database design, schema migration review, data-flow reliability, tenant/PII controls, or architecture evolution. Existing systems stay read-only until the user selects an option and authorizes exact mutations. Unknown facts stay UNKNOWN. This is an off-DAG engineering skill: do not emit findings or invent STAGE_GATES, ROUTES, stages, or back-edges.
Installation
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Design or modernize a software system from an evidence-backed current-state inventory through quality attributes, architecture options, API and schema contracts, migration/rollback plans, ADRs, and verification. Use for greenfield or existing monoliths, modular monoliths, services, system/API/ database design, schema migration review, data-flow reliability, tenant/PII controls, or architecture evolution. Existing systems stay read-only until the user selects an option and authorizes exact mutations. Unknown facts stay UNKNOWN. This is an off-DAG engineering skill: do not emit findings or invent STAGE_GATES, ROUTES, stages, or back-edges.
System design lifecycle
Own system boundaries, runtime interfaces, operational data stores, system
migrations, and architecture decisions. Do not equate a diagram, SQL file, or
OpenAPI document with a working, safe, or scalable system.
Read
references/system-design-resource-map.md
before any existing-system task. It defines the lifecycle, artifact contract,
decision stop, evidence states, access tiers, and cross-skill ownership. Read
references.md only for the database/API/reliability branch
that applies to the selected system.
Non-negotiable boundary
Treat repository, configuration, schema, API, dependency, and deployment
intake as read-only access. It is not authorization to rewrite them.
Preserve absent facts as UNKNOWN. Do not infer traffic, SLOs, consistency,
budget, compliance, migration windows, or team capability from the phrase
“system design.”
Present at least a recommendation, a viable alternative, and explicit
rejection/exit conditions. Stop before choosing the database, topology,
compatibility policy, or migration strategy for the user.
Bind any later mutation to a user-selected option and exact authorized
action IDs. Keep before/after locators, SHA-256, verification, and rollback.
Never mutate a production database or configuration. Apply only to a
disposable environment explicitly placed in scope; otherwise deliver a
reviewed plan and scripts.
Describe schema_lint.py as a lexical heuristic. It is not a SQL parser,
query planner, lock simulator, schema diff engine, or zero-downtime proof.
Keep VERIFIED for checks that actually ran and retain their command,
return code, locator, and hash. Use PLANNED, UNKNOWN, or UNAVAILABLE
otherwise.
Keep this skill off the research DAG. Emit no light.findings.v1; add no
STAGE_GATES, ROUTES, stage number, or back-edge; do not attach _shared.
Choose the mode
Situation
Mode
New system with no implementation
greenfield requirements and option design
Existing repository/system
read-only intake, then current-state inventory
Existing monolith or services changing gradually
modernization with compatibility and rollback
API-only change
contract and consumer compatibility branch
Schema-only change
dialect/version/context-specific migration branch
User supplied a completed package
review and evidence verification
Phase 1 — Intake and protection
Capture or preserve as UNKNOWN:
users, business goal, critical use cases, data classification;
load range, latency, availability, durability, and consistency targets;
team/operations capability, budget, deployment environment, compliance, and
migration window;
topology: greenfield, monolith, modular monolith, or services;
Keep --out outside the source root. Read all emitted artifacts and verify
source_unchanged=true.
Phase 2 — Requirements and current state
Produce:
context and quality-attribute scenarios with a measurable stimulus,
environment, response, and target or UNKNOWN;
capacity estimates for request rate, storage growth, fan-out, latency budget,
and any dominant resource; if unknown, write UNKNOWN plus the measurement
plan rather than inventing numbers;
current-state → target-state mapping with explicit gaps. For greenfield,
current state can be none, but the gap list still records missing evidence;
architecture fitness functions: observable signals, thresholds, verification
command/probe, and evidence state. A quality attribute without a fitness
function is still only prose;
a component/interface/store inventory with owner and fact provenance;
synchronous/asynchronous data flows, transaction boundaries, delivery
semantics, idempotency/deduplication, backpressure, timeout/retry, and failure
handling;
risk, assumption, unknown, and stale-fact registers.
Do not silently convert a code search into an architecture truth. Mark each
fact as declared, observed, inferred, or unknown.
Phase 3 — Options and decision stop
Present at least:
a recommended option with reasons;
a viable alternative;
conditions under which each should not be used;
cost/complexity, migration risk, compatibility window, rollback, operations,
and exit criteria;
unresolved facts that could reverse the recommendation.
Then stop. Ask the user to select an option and authorize exact action IDs.
Do not prewrite the user's choice or generate the chosen schema/API/migration/
ADR as if approval already existed.
Before presenting the decision, validate that requirements, capacity estimates,
current/target state, fitness functions, at least two genuinely different
options, hard-constraint and fitness evidence, tradeoffs, rejection conditions,
reversal costs, and migration/deprecation stance are present:
In proposal, PASS means only ready_for_user_decision=true; it never writes
the selection, and the report emits a canonical option_packet_sha256 for each
option. In authorized, the selection must be paired with a
light.system-design.v2.authorization whose option digest still matches,
whose approved action IDs are a subset of that option, whose target is
explicitly disposable, whose rollback cannot be waived, and whose date is not
later than --as-of. The walking skeleton
(entry/core_path/state_boundary/observable_result/failure_probe/verification/action_ids)
may contain only approved actions before ready_for_implementation=true.
For each option, state whether migration/deprecation is applicable. If it is
applicable, the option must be replacement-first. Consumer inventory is a list
of stable consumer/interface IDs, owners, usage status, evidence state,
evidence locator/date, or an explicit measurement plan. Telemetry is a
structured metric/source/evidence record. Rollout is a sequence of phases with
entry, exit, and rollback conditions; rollback has a trigger, action, and
verification. A deprecation compatibility window has start, end, and removal
conditions. Plain strings do not satisfy these fields. If migration is not
applicable, record why; do not leave it blank.
Use
templates/decision-authorization.template.json
after the user responds. Copy the selected digest emitted by
design_readiness.py; a changed requirement, state model, fitness function, or
selected option changes that digest and requires fresh authorization.
Phase 4 — Build the architecture package
After authorization, produce only the selected scope:
context and quality attributes;
component/boundary and data-flow views;
API/event contracts and consumer compatibility policy;
schema and current-to-target change plan;
rollout, backfill, rollback, and deprecation plan;
ADR with alternatives and consequences;
security/privacy design controls and items for specialist review;
For authoritative diff/drift, use a real engine/tool selected for the project
(for example Atlas, Skeema, Alembic, Flyway, Liquibase, or Prisma) and preserve
its command/output. Do not claim this skill implements those engines.
If Mermaid rendering is unavailable, report syntax/structure verification only.
Phase 6 — Verify, rehearse, and deliver
Verify as applicable:
contract and example request/response;
schema creation and migration on the selected database/version;
data preservation, compatibility window, rollback, and reapply;
load assumptions rather than invented load results;
timeout/retry/circuit-breaker and failure drills;
logs, metrics, traces, SLOs, deployment, and rollback observability;
tenant isolation and PII controls with specialist review still pending.
Record authorization binding, source-intake binding, implemented action IDs,
artifact hashes, and verification entries in the package manifest, then run:
Deliver only when the package distinguishes VERIFIED, PLANNED, UNKNOWN,
and UNAVAILABLE; every VERIFIED entry is evidence-backed; the manifest
binds the copied authorization file, option digest, approved action IDs, and
implemented action IDs; and existing-system packages bind the read-only
intake-integrity.json hash. Artifact and verification locators in the
manifest are resolved relative to the manifest's directory and must stay inside
that package directory; ../, absolute paths to outside evidence, or
current-working-directory-dependent locators are not a portable delivery
package.
Cross-skill ownership
system-design: system boundaries, runtime interfaces, operational schema,
system migration, reliability choices, ADRs.
project-structure: visible file tree and authorized file moves. Borrow its
protection discipline; never send schema migration back to it.
data-engineering: research-data quality, lineage, transformations, splits,
and data release. A service database is not a research dataset pipeline.
frontend-design: interaction and interface implementation. This skill owns
backend/API boundaries, not UI.
research-ethics: final ethics/privacy judgment. This skill proposes design
controls and review items only.
orchestrator: may consume delivered state; it receives no invented gate.
Unknown requirements and stale facts stayed explicit.
Capacity/load/storage estimates are explicit; unknowns have measurement
plans instead of invented numbers.
Current-state → target-state gaps are recorded, even for greenfield
(current_state=none).
Every quality attribute has a fitness function and each option has a
fitness result with evidence or an honest UNKNOWN/UNAVAILABLE warning.
Options, tradeoffs, rejection conditions, and a real user decision exist.
Authorization digest still matches the selected option packet; approved
action IDs are in scope, the target is explicitly disposable, and
rollback remains required.
Package manifest binds the authorization file hash, option digest,
implemented action IDs, and read-only intake integrity or an explicit
greenfield/not-applicable reason.
At least two interfaces/options were compared; the first idea was not silently accepted.
The authorized design has the thinnest end-to-end walking skeleton and a failure probe.
Migration/deprecation stance is explicit; applicable migrations are
replacement-first with consumer inventory, telemetry, rollout, rollback,
and compatibility window.
Mutations match authorized action IDs and a disposable target.
API/schema/migration claims state dialect, version, context, and limits.
Every VERIFIED item has command, return code, locator, and SHA-256.
Every verification entry carries action IDs that are inside the
implemented and approved scope.
Package artifact/evidence locators are manifest-relative and do not
escape the package directory.
Rollback and compatibility were exercised or remain visibly planned.
