borghei/Claude-Skills / engineering/feature-flags-architect
feature-flags-architect
VerifiedFeature flag strategy, lifecycle, and operations. Use when designing a flag taxonomy, planning a gradual rollout, building kill switches, auditing flag debt, defining governance, running progressive delivery, or rolling back via flags.
Feature Flags Architect
End-to-end feature flag design, rollout, and lifecycle management. Covers flag taxonomy (release vs ops vs experiment vs permission), gradual rollout patterns with blast-radius math, kill-switch runbooks, governance (who can flip what), and the flag-debt cleanup loop that keeps a codebase from drowning in stale flags.
This skill is provider-agnostic: the patterns work whether you run LaunchDarkly, Statsig, Unleash, Flagsmith, ConfigCat, GrowthBook, OpenFeature, or a homegrown system backed by Redis/DynamoDB/Postgres.
Core Capabilities
- Flag taxonomy — classify every flag as one of four types (release, ops, experiment, permission); the type sets lifetime, who flips it, and auto-expiry.
- Rollout ramps — dogfood → canary → 100% schedules with per-step blast-radius math and ramp templates by change type.
- Kill switches — ops-flag pattern with fail-open/closed decisions, runbooks, and alerting for risky dependencies or code paths.
- Governance — role-based who-can-flip-what matrix plus audit-log/retention requirements, mappable onto any flag system.
- Flag debt cleanup — quarterly inventory → classify → decide → execute → measure loop to keep flag count flat or shrinking.
- End-to-end workflows — add a release flag, build a kill switch, audit flag debt, flag-based rollback after a bad release.
Flag taxonomy — the four types
Every flag belongs to exactly one of these four types. The type determines the lifecycle, who can flip it, and whether it should auto-expire. Mixing types in one flag is a root cause of flag debt.
| Type | Purpose | Lifetime | Who flips | Auto-expire? |
|---|---|---|---|---|
| Release | Decouple deploy from release. Ship code dark, ramp to users. | Days to weeks | Engineer who owns the feature | Yes — remove after 100% rollout + 1 release |
| Ops | Kill switches, circuit breakers, throttles. Turn off risky behavior fast. | Permanent or long-lived | Oncall / SRE / platform team | No — but review quarterly |
| Experiment | A/B test, multi-arm bandit, hold-out group. Measure causal impact. | Weeks to months (test duration) | Product / data / growth | Yes — remove after winner is shipped |
| Permission | Entitle users to features based on plan, role, beta-list. | Permanent | Product / billing | No — but consolidate into entitlement system |
Decision rule when adding a flag: name the type. If you can't, the flag shouldn't exist yet — clarify intent first.
When to Use
- Designing a flag system from scratch — start with the flag taxonomy.
- Planning a risky release (DB migration, payment-provider swap, framework upgrade) — use the rollout ramp + kill switch.
- Auditing a codebase with hundreds of growing flags — use flag debt cleanup +
scripts/flag_audit.py. - Setting up an A/B test or hold-out — covers the flag side; use
data-analytics/statistical-analystfor sample size and significance. - Building a kill switch for a third-party dependency outage — use the kill switch runbook generator.
- Defining who in the org can flip production flags — use the governance matrix.
- Building entitlements / paid-tier gating — use this for flag mechanics;
business-growth/paywall-upgrade-crofor the UX.
Clarify First
Before designing or operating flags, settle the inputs below. When one is missing or vague and would change the result, ask about it; when it wouldn't, use a sensible default and say which one you used.
- Flag type: release, ops, experiment, or permission (sets the lifecycle, who can flip it, and auto-expiry — the root decision).
- Task: design a taxonomy, simulate a rollout ramp, generate a kill-switch runbook, or audit flag debt (selects
flag_audit.pyvsrollout_simulator.pyvskill_switch_runbook.py). - Blast radius / population: user count and ramp profile, or the dependency a kill switch guards (drives
--users/--profileand the runbook content).
Ask about the two or three inputs that matter most for this request, not the whole list. If the user wants a result straight away, produce it and state your assumptions at the top.
Workflow
- Name the type of every new flag (release, ops, experiment, permission) and give it an owner and, for release and experiment flags, a removal date.
- Plan the ramp with
rollout_simulator.pyfor the user count and risk profile; each step needs the metric that would stop it. - Build the kill switch for risky dependencies with
kill_switch_runbook.py, passing the real owner team, oncall rotation and flag-console URL so the runbook has no placeholders. - Audit debt quarterly with
flag_audit.pyagainst a control-plane export: delete dead flags, remove stranded 100% release flags and their dead branches, investigate orphan references. - Verify that the kill switch actually flips in staging (or under fault injection) and that the audit's debt list shrinks quarter over quarter.
Quick Start
python3 scripts/flag_audit.py --path assets --flag-export assets/sample_flag_export.json --as-of 2026-06-01 # 1 dead, 1 orphan, 1 stranded at 100%
python3 scripts/rollout_simulator.py --users 500000 --profile standard
python3 scripts/kill_switch_runbook.py --flag ops.recs.kill_switch --feature Recommendations --dependency recs-svc \
--owner-team recs-team --oncall recs-oncall --flag-console-url https://flags.example.com/flags/ops.recs.kill_switch
Tools Overview
| Tool | What it does | Key flags | Exit codes |
|---|---|---|---|
flag_audit.py | Find flag reads in code, join with a control-plane export, classify each flag (dead, orphan, stranded at 0/100%, should-be-config, ghost owner, healthy) with an action | --path, --flag-export, --as-of, --pattern <regex> (repeatable), --owner-team, --include-active, `--format json | markdown |
rollout_simulator.py | Per-step blast radius, time to detect, recommended schedule | --users, --profile, --format | 0, 2 bad arguments |
kill_switch_runbook.py | Runbook: default state, fail mode, flip procedure, validation, escalation | --flag, --feature, --dependency, --owner-team, --oncall, --flag-console-url, `--fail-mode open | closed`, thresholds |
flag_audit.py recognises LaunchDarkly (variation, boolVariation), OpenFeature (getBooleanValue, useBooleanFlagValue), Statsig (checkGate), Unleash (isEnabled), generic flags.isEnabled/evaluate/getVariant, homegrown helpers (getFlag, useFlag, isFeatureEnabled, get_feature_flag) and flag decorators. For other SDKs pass --pattern with a regex whose first group is the flag key. Flags read through variables built at runtime are not found; grep for those by hand. The runbook defaults (unknown-team, a placeholder console URL) are there to be overridden.
Anti-Patterns
Mistake: Release flags without a removal date. Why it happens: Once the feature is at 100%, removing the flag has no visible payoff. Instead: Set the date when the flag is created and let the quarterly audit enforce it; stranded flags double the code paths to test.
Mistake: A kill switch that has never been flipped. Why it happens: It exists for emergencies, so nobody tries it. Instead: Flip it in staging on a schedule (or under fault injection) and record the time to effect.
Mistake: One flag doing two jobs (a release ramp that later becomes a permission gate). Why it happens: Reusing an existing flag is quicker than creating one. Instead: Create a new flag of the right type. Mixed-type flags cannot be removed without breaking someone.
Files
scripts/flag_audit.py,scripts/rollout_simulator.py,scripts/kill_switch_runbook.py— the tools above.assets/sample_checkout.ts(flag reads through three SDK shapes),assets/sample_flag_export.json(control-plane export).
Reference Documentation
Load the reference that matches the task — keep this file lean and pull detail on demand:
- references/flag-types-and-patterns.md — the four flag types in depth, code patterns per type, evaluation-context schema, naming conventions, server-vs-client trade-offs, and storage models. Read when designing a flag system or writing the gating code.
- references/operations-governance-and-workflows.md — the standard ramp table, blast-radius math, kill-switch design, the governance matrix + audit-log requirements, the flag-debt cleanup loop, all four end-to-end workflows, the anti-patterns list, and the tooling input/output reference. Read when running a rollout, defining governance, or executing a workflow.
- references/rollout-and-kill-switch-playbook.md — rollout decision tree, ramp templates by change type, cohort/segment selection, kill-switch design, and monitoring/alerting hookup. Read when planning a specific ramp or wiring kill-switch alerting.
- references/flag-debt-and-cleanup.md — what flag debt costs, the cleanup cadence, dead-code detection, safe removal procedures, PR templates, and governance for flag count. Read when paying down stale flags.
Scope & Limitations
This skill covers: flag taxonomy and lifecycle, rollout ramps with blast-radius math, kill-switch design and runbooks, governance/audit requirements, and flag-debt cleanup — provider-agnostic across LaunchDarkly, Statsig, Unleash, Flagsmith, ConfigCat, GrowthBook, OpenFeature, and homegrown systems.
This skill does NOT cover: the statistical design of A/B tests (data-analytics/statistical-analyst), the paywall/upgrade UX for entitlements (business-growth/paywall-upgrade-cro), or the metric instrumentation needed to make ramp decisions (engineering/observability-designer).
Related skills
engineering/observability-designer— wire the metrics needed to make ramp decisionsengineering/incident-commander— flag-based rollback during incidentsengineering/chaos-engineering— verify kill switches actually work under fault injectiondata-analytics/statistical-analyst— sample size and significance for A/B tests behind flags
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 469k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 469k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 469k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |