melech-buy-vs-build
Decide whether to adopt an existing tool or build a capability yourself, grounded in verified research.
AdirD · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add AdirD/agent-shell-hamelech --skill melech-buy-vs-build技能说明
name: melech-buy-vs-build description: Decide whether to adopt an existing tool or build a capability yourself, grounded in verified research. disable-model-invocation: true
Buy vs Build
An agent asked to "add background jobs" will happily write a tasks table, a
polling worker, a retry column, and a dead-letter flag. It works. It is also
BullMQ, or Trigger.dev, or Cloud Tasks, or the queue library already sitting in
package.json — rebuilt badly, and now owned forever.
Nobody makes the call deliberately. Repo-level review only ever asks "does this already exist in our code?", and model memory answers external questions with confident, stale, sometimes invented package names. The expensive, hard-to-reverse decision — own the implementation, or hand the capability to a third party — gets made by default instead of on evidence.
melech-buy-vs-build owns that decision. It looks outward — finding what the
rest of the world already shipped for the capability, verified against live
sources — and it runs one inward check first: adopt-vs-rebuild, is this
already covered by a dependency or vendor you already have? Then it brings back
a shortlist and a verdict: adopt, or keep building.
Think "is there an AI for that?" — but for the build-vs-buy call, and grounded in evidence instead of vibes.
Two Intents
Both intents are outward research feeding one build-vs-buy verdict. Whichever fires, run the inward adopt-vs-rebuild check first (step 2) — a capability you already own beats anything you would go find or build.
| Intent | Trigger | Goal |
|---|---|---|
| Intercept | Code, a plan, or a diff exists and may be reinventing something. | Name what was (re)built, find the incumbents, decide adopt vs. keep building. |
| Explore | Open-ended: "what's out there for X", "any new tools for X", "how do people do X now". | Map the space, surface the notable and the recent, arm the user to choose. |
Do not ask which intent applies when it is obvious from the input. A diff or a proposal means intercept. A bare capability question means explore.
The One Rule That Decides Everything
Search the capability, not the noun.
The single reason this research fails is searching the user's own vocabulary.
An agent that greps the web for "postgres polling table worker" finds
blog posts. An agent that recognizes the capability as durable background job
execution finds the entire category.
So before any search: translate the implementation into the canonical term the industry uses for it.
| What the agent built | What to actually search |
|---|---|
| Rows in a table + a polling worker + retry count | job queue, background jobs, durable task execution |
| Custom event bus with handler registry | pub/sub, message broker, event bus |
| Hand-rolled retry-with-jitter wrapper | retry library, resilience/backoff library |
| Bespoke role/permission matcher | authorization / policy engine, RBAC / ReBAC |
Cron-ish setInterval + lockfile | job scheduler, distributed cron, workflow orchestration |
| Diff-and-apply state machine for a wizard | state machine library, workflow engine |
| Custom CSV/Excel export writer | spreadsheet / serialization library |
If unsure of the canonical term, the first search is for the term itself: "what is it called when …". Get the vocabulary right, then fan out.
Workflow
1. Name the capability ─► 2. Read the ground truth ─► 3. Fan out lanes ─────────► 4. Reduce ─► 5. Shortlist + verdict ─► 6. Hand off
(canonical vocabulary) (stack, constraints) adopt-vs-rebuild (inward) ┐ (dedupe, kill dead) (ask_question)
outward lanes (web) ┴─ same concurrent batch;
inward hit short-circuits
1. Name the capability
State in one line what the thing does, stripped of the local implementation. List 2–5 canonical search terms and any obvious synonyms. Show this to the user before fanning out — a wrong capability name wastes the whole run.
2. Read the ground truth (inline, must be first)
This quick read stays on the main thread because the outward lanes cannot be briefed without it. Gather the constraints that will decide fit:
- Language and runtime, and what is already in the manifests
(
package.json,requirements.txt/pyproject.toml,go.mod,Cargo.toml,Gemfile,pom.xml) - Infrastructure already paid for: cloud provider, Postgres/Redis/Kafka, vendors already in the bill
- Hard constraints: self-host only, data residency, licensing policy, air-gapped, no new vendors, budget
- Scale reality: 100 jobs/day and 100k jobs/second do not shortlist the same tools
If this cheap read already surfaces a blatant hit — the capability is literally a direct dependency or an obviously enabled vendor feature — stop here and report it. Do not fan out to buy something you already own. Otherwise, carry the constraints into step 3 and let the deep adopt-vs-rebuild check run concurrently.
3. Fan out parallel lanes
Dispatch independent research lanes concurrently — one subagent per lane, each running multiple web searches along its own path (except adopt-vs-rebuild, which is mostly local). Lanes hunt different kinds of answers, not different keywords. The inward lane and the outward lanes launch in the same concurrent batch, so you do not pay the inward check as serial latency.
| Lane | What it hunts |
|---|---|
| Adopt-vs-rebuild (inward, privileged) | The deep version of the step-2 check: transitive deps, framework/stdlib built-ins, a vendor plan that already covers this, an internal monorepo library. Mostly local work plus targeted doc lookups. Short-circuit authority — see below. |
| Canon | The category's standard name and the 3–8 options every comparison lists. |