| name | gstack-plan-eng-review |
| description | Use when a plan needs technical stress-testing, when sequencing is unclear, when the user asks for a tech lead's view, or when dependencies may block the build. Triggers: "tech lead review", "is this buildable", "stress-test the architecture", "what's the load-bearing assumption", "is the sequencing right", "what's the rollback plan", "what's the smoke test", "is this an irreversible decision". |
Plan — Engineering Review
The counterpart to gstack-plan-ceo-review. CEO answers "should we build this?" — this answers "is the proposed order even possible, and what's the technical risk?"
When to run
- After a CEO review, before kickoff.
- When the user is about to commit engineering time to a sequence and is unsure of the order.
- When dependencies between phases are unstated.
- When the plan has hand-wavy "and then we'll integrate" steps.
The 10 questions
1. What is the one thing that has to be true for the rest to work?
Every non-trivial plan has a load-bearing assumption. Find it. If it's wrong, the rest collapses.
State it: "This plan assumes X. If X is false, the plan needs Y as a fallback."
2. What's the build order, and why is phase 1 first?
Force a sequence. Most plans are written in parallel ("we'll build A, B, and C") when only some of A is needed to start B.
State the dependency graph explicitly: "B requires A. C requires B. D requires nothing." If a phase has no upstream blocker, it can move earlier.
3. Where is the unrecoverable error path?
What, if done wrong, would force a complete rewrite? These are the decisions to make slowly and first. Everything else is fine to discover.
Examples: data model, auth boundary, deployment target, schema migration strategy.
4. Build vs buy vs adapt — for each major dependency?
For every external dependency (auth, payments, ML model, code analysis, search), state: build, buy, or adapt from an existing piece. Justify in one sentence.
Default answer: adapt. Most "build it" answers fail the time-to-value test.
5. What's the scale target for the first 1,000 users?
Plans that target "millions of users on day one" are usually plans that target zero users. Be concrete: requests/sec, storage GB, concurrent sessions, deployment region.
Numbers reveal where the actual risk is. "5 req/s" means database choice doesn't matter. "5,000 req/s" means it's the only thing that matters.
6. What is the rollback plan?
If phase 3 ships and breaks, can you roll back to phase 2? Is the data model additive-only? Are the migrations safe to undo?
Plans that can't roll back are plans that don't get to ship until they're perfect, which is never.
7. What's the test surface?
What's the smoke test that proves the plan works end-to-end? Not unit tests — the one test a non-engineer could run that proves the user-facing outcome.
If you can't name it, the plan doesn't have a definition of done.
8. What's the security boundary?
What does the user trust the system with? Where is that trust enforced? What's the worst-case if a boundary is breached?
For AI-agent systems, this is the question. State the boundary: which tool calls are user-approved, which are sandboxed, which are gated behind a human-in-the-loop.
9. What is the observability story?
When this is live and something goes wrong at 2am, what does the on-call see? Logs, metrics, traces, alerts. Be specific.
"No observability yet" is a fine answer for phase 1. It is not a fine answer for "go live."
10. What's the smallest end-to-end vertical slice?
The vertical slice that exercises every layer of the stack — UI → API → DB → external service — is the test of the plan's shape. If you can't build it in <2 weeks, the layer breakdown is wrong.
Output format
PLAN: <name>
LOAD-BEARING: <the assumption that, if false, breaks the plan>
SEQUENCE: <ordered list of phases, with explicit dependencies>
IRREVERSIBLE: <decisions to make first / slowly>
BUILD/BUY: <one-liner per major dependency>
SCALE TARGET: <req/s, storage, region, by when>
ROLLBACK: <how to revert, what blocks it>
SMOKE TEST: <the one test a non-engineer could run>
SECURITY: <the trust boundary + what happens if breached>
OBSERVABILITY: <what the on-call sees at 2am>
SLICE: <smallest end-to-end vertical slice + estimate>
VERDICT: <ship / cut / re-scope / sequence-changes-needed>
Source / license
Adapted from JZKK720/gstack · plan-eng-review, MIT.