AWS Agent Advisor

Helps startups decide how and where to run AI agents on AWS. Deterministic scoring recommends a runtime; the conversation adapts to the user's technical background.

Definitions

• "Load" = Read the file with the Read tool and follow it. Do not summarize or skip.
• $RUN_DIR = the run directory under .agent-advisor/ (e.g. .agent-advisor/0630-1430/), created in Intake.
• $PLUGIN = ${CLAUDE_PLUGIN_ROOT} (the installed plugin root). On Claude Code this token substitutes inline. If ${CLAUDE_PLUGIN_ROOT} does not resolve (some Cursor/Codex builds, or a literal ${CLAUDE_PLUGIN_ROOT} string showing up in a path error), fall back to the skill's own directory: this SKILL.md lives at <plugin>/skills/agent-advisor/SKILL.md, so the engine and its data are all inside this skill — scripts at ./scripts/..., runtime profiles at ./references/runtimes/..., and decision refs at ./references/decision-refs/... relative to it. Prefer ${CLAUDE_PLUGIN_ROOT}/skills/agent-advisor/...; use the relative fallback only when it fails to resolve.

Prerequisites

• uv available (for scoring). Check: uv --version. If missing, tell the user to install it from the official install guide (https://docs.astral.sh/uv/getting-started/installation/ — e.g. brew install uv or pipx install uv) and stop.

Phase Structure (frontmatter)

Phase, fragment, and assembler files carry a YAML frontmatter block that declares how each phase is composed — its inputs, triggers, fragments, assembler, artifacts, gates, and ordering. The execution contract is the vendored references/vendored/dsl/INTERPRETER.md: it defines every frontmatter key, the fragment/assembler model, the gate protocol (HANDOFF_OK / GATE_FAIL), and the interpreter loop. Load it first (once, at the start of a run), then execute each phase file's prose body. Elsewhere in this skill, INTERPRETER.md (without a path) refers to this loaded contract.

Execution

This skill is driven by the interpreter loop in INTERPRETER.md (§ The interpreter loop): it reads .phase-status.json, determines the current phase, runs each phase's _preconditions / fragments / _assemble / _postconditions, advances on HANDOFF_OK via _advances_to, and validates state. The backbone (intake → discover → clarify → confirm → design → estimate → generate → migration-plan → poc → complete) and the one sidebar branch (add-capabilities) are derived from the phase files' frontmatter — they are not restated here.

Cold start (entry phase). With no run under .agent-advisor/ carrying a .phase-status.json, begin at references/phases/intake/intake.md — this skill's entry phase (the one carrying _init: true). On a warm start, current_phase in .phase-status.json is authoritative (INTERPRETER.md § The interpreter loop).

Skill bindings (INTERPRETER.md § Skill bindings). This skill declares:

• Run root: .agent-advisor/ — $RUN_DIR is this skill's name for the run directory (.agent-advisor/[MMDD-HHMM]/). Intake's own prose performs the _init bootstrap.
• State shape: § State file below (advisor-specific keys such as entry_point, audience, recommendation_reviewed, migration_plan_ctx, migration_plan_unavailable); the shared state schema is not vendored.
• Run seed (optional): $RUN_DIR/seed.json, else .agent-advisor/seed.json at the run root (schema scripts/schemas/seed.json) supplies machine-readable answers for a non-interactive run — the Clarify dimensions, the two gate answers, the POC mode, the live-probe answer, and a co_recommend tie-break. It is the HIGHEST-precedence source for every value it carries (clarify.md Step 2.5), which is what makes a repeated run's score comparable: the deterministic engine gets byte-identical input. A gate the seed omits is declined; a dimension the seed omits falls through to detection, then prose, then an assumed value that MUST be recorded in $RUN_DIR/UNANSWERED.md. With no seed, the interactive flow is unchanged.
• Resolved statuses: skipped (routing resolved the phase without running it), plus not_applicable for migration_plan only.
• Conditional backbone routing: the entry-point routing below. When a routing rule marks a phase not-applicable, set it skipped and advance through its _advances_to in the same state write.

Routing & gates (orchestration)

Sidebar placement and conditional backbone routing are orchestration prose owned by this file (INTERPRETER.md § Skill bindings, § Backbone vs sidebar).

Entry-point routing:

• build_scratch → skip Discover; Clarify → Confirm → Design → Estimate → Generate → Gate 2 → POC (any winning runtime). No migration plan (nothing existing to migrate).
• build_deploy → Discover (if code) → Clarify → Confirm → Design → Estimate → Generate → Gate 1 → Migration Plan (if existing non-AWS AI workload detected and user confirms) → Gate 2 → POC (any winning runtime).
• migrate → Discover (if code) → Clarify → Confirm → Design → Estimate (target-state run cost; migration TCO comparison stays with the Migration Plan engine) → Generate → Gate 1 → Migration Plan (in-skill, reusing the sibling gcp-to-aws skill) → Gate 2 → POC (any winning runtime, when the plan was produced). Declining Gate 1 keeps the classic handoff: pointer to /aws-startup-advisor:llm-to-bedrock with handoff-summary.md.
• add_capabilities → load references/phases/add-capabilities/add-capabilities.md and follow it (no runtime scoring; writes capabilities-recommendation.md). This is a self-contained branch — it does NOT pass through Clarify / Confirm / Design / Estimate / Generate, so the phase gate below never applies to it.
• Temporal detection routes into migrate with temporal units pre-seeded (see discover).

Gate semantics (backbone tail):

• Gate 1 → migration_plan runs only when generate is done AND recommendation_reviewed == true (generate.md Step 5.5) AND entry point ∈ {migrate, build_deploy} AND the run is migration-eligible (generate.md Step 6) AND the user confirmed Gate 1. Otherwise resolve it: not_applicable (build_scratch / no migratable workload) or skipped (declined) — and advance.
• Gate 2 → poc runs only when phases.poc == "in_progress" (set when the user answers Gate 2 "yes" — asked in generate.md Step 7 or migration-plan.md Step 6) AND recommendation_reviewed == true. Any winning runtime (agentcore / ecs / eks / lambda / lambda_microvms) — the POC shape follows the verdict (poc.md Step 3 dispatch on references/decision-refs/poc-shapes.md). Gate 2 is only offered when migration_plan ∈ {completed, skipped, not_applicable} — or in_progress on build_deploy only (Stage 2 failed/aborted; fallback POC from design.json per migration-plan.md failure handling); for entry point migrate, only when migration_plan == "completed" (the POC implements the plan) OR when the stage resolved not_applicable with migration_plan_unavailable == "engine_absent" — a standalone deployment that does not bundle the migration engine, where Gate 2 is offered by migration-plan.md Step -1 and the POC is design-backed. A migrate-POC with no plan for any OTHER reason (the user declined) has nothing to implement.
• Persisting Gate 2 as phases.poc = "in_progress" BEFORE poc.md loads makes the confirmation resumable: if the session breaks between the "yes" and the load, the interpreter re-enters poc without re-asking. (A declared deviation from INTERPRETER.md § The interpreter loop step 5's gate-then-in_progress ordering — the user's confirmation is the entry event worth persisting.)

Phase gate: Do NOT load design.md / estimate.md / generate.md unless $RUN_DIR/.phase-status.json exists and BOTH phases.clarify == "completed" AND phases.confirm == "completed". Confirm confirms the deployment model, the service set, and (for a co_recommend tie) the user's chosen_runtime — Design and the diagram depend on its confirm.json output, so it must not be skipped. If the user asks to skip Clarify or Pass 2, refuse briefly and run it.

State file (.phase-status.json)

{
  "run_id": "0630-1430",
  "entry_point": "build_scratch",
  "audience": "technical",
  "current_phase": "clarify",
  "phases": {
    "intake": "completed",
    "discover": "skipped",
    "clarify": "in_progress",
    "confirm": "pending",
    "design": "pending",
    "estimate": "pending",
    "generate": "pending",
    "migration_plan": "pending",
    "poc": "pending"
  }
}

Status values: pending → in_progress → completed, plus skipped. Use read-merge-write: read before each update, change only the advancing keys, keep prior phases.

recommendation_reviewed (top level, boolean) is set to true by generate.md Step 5.5 when the user explicitly confirms they have seen the recommendation. Gate 1, Gate 2, and the migration_plan / poc states all require it — no gate may be asked while it is absent.

migration_plan additionally uses not_applicable (build_scratch, or no migratable workload detected). When Stage 2 runs, migration_plan_ctx is added at the top level: {"repo": "<abs path to target repo>", "migration_dir": "<abs path to .migration/<id>/>"} — Stage 3 reads gcp-to-aws artifacts ONLY via this recorded path, never by re-globbing.

Files

File	Purpose
references/vendored/dsl/INTERPRETER.md	Vendored DSL execution contract (interpreter loop + gate protocol)
references/phases/intake/intake.md	Entry point + technical background + open context
references/phases/discover/discover.md	Lightweight code detection
references/phases/clarify/clarify.md	Clarify orchestrator + answer mapping to scoring keys
references/phases/clarify/clarify-technical.md	Technical-background question wording
references/phases/clarify/clarify-business.md	Business-background question wording
references/phases/confirm/confirm.md	Winner-specific follow-ups
references/phases/design/design.md	Assemble recommendation; Migrate handoff branch
references/phases/estimate/estimate.md	Coarse cost magnitude
references/phases/generate/generate.md	Layered recommendation doc + scaffolding
references/phases/migration-plan/migration-plan.md	Stage 2: full migration plan via the sibling gcp-to-aws engine
references/decision-refs/temporal.md	Temporal rules: Tier 1/2 tables, adapter, runbooks, commercials (consumed by discover/clarify/design/generate)
references/decision-refs/poc-shapes.md	Per-runtime POC deploy shapes (ECS/EKS/Lambda/MicroVMs/Temporal)
references/decision-refs/*.md	Runtime service cards, model defaults, freshness
references/decision-refs/workload-classes.md	Deterministic verdicts for non-agent workload units (batch/service/io)
references/runtimes/*.json	Runtime registry (read by scoring.py)
scripts/scoring.py	Deterministic scoring engine
scripts/test_temporal_decision_refs.py	Content lock for the Temporal decision reference
scripts/test_poc_shapes.py	Content lock for the POC deploy shapes
scripts/test_workload_classes.py	Content lock for workload-classes.md (verdicts table)
scripts/test_unit_grouping.py	Unit grouping + pattern matching (workload-class assignment)
scripts/test_collapse_invariant.py	Collapse-invariant ordering enforcement (A→B implies [B] ⊆ [A] outputs)