The build-decision guide for the AI era
Should it be a
prompt?
Or a skill, an MCP server, an agent, a real app, or nothing at all. Answer eight questions and get a verdict with the reasoning, the trade-offs and a first step.
By Mahmoud Halat, founder of Space & Story. Free, no sign-up, no model call: the logic is published.
- 6App / systemWeeks to months
- 5AgentDays to weeks
- 4MCP / connectorHours to days
- 3SkillAn afternoon
- 2Saved promptUnder an hour
- 1PromptMinutes
Step one
Run your task up the ladder
Eight questions. The cheapest rung that works wins, and every verdict shows the rule that set it. How the rules work.
The decision tool
Picture one task you're thinking of handing to AI. Answer eight questions about it and you'll get one rung of the ladder, or a plain "don't build it", with the rules that put it there, what it costs you and a concrete first step. The logic is fixed and published, so the same answers always give the same verdict.
Question 1 of 8
How often does this task come up?
Count the real occurrences over the last couple of months, not the ones you hope for.
Pick an answer to beginTip: press A, B, C, D or 1–4
What the patterns say
15 of 38 everyday tasks are skills. Only 7 need an agent or an app.
We ran 38 common work tasks, from meeting notes to invoice matching, through the same rules as the tool. Most land low on the ladder: packaged instructions beat software far more often than the hype suggests. Each one has its reasoning and a starting point.
Open the pattern library- PromptAlt text for a site's image backlog · Explaining one clause you don't understand
- Saved promptWeekly status update to your manager · Social posts from a new blog article
- SkillMeeting transcript to summary and action items · Proposal and quote drafting
- MCP / connectorWeekly KPI report from your analytics · Sales call prep brief
- AgentInbound lead triage and routing · Churn-risk alerts for account managers
- App / systemThree-way invoice matching (PO, receipt, invoice) · Personalised cold outreach at scale
- Don't buildCategorising expenses and bank transactions · Drafting replies in your own inbox
The explainer
Prompt vs skill vs MCP server vs agent vs app: what each one is, what it costs, when it breaks.
Read the definitions| Rung | First version | Upkeep |
|---|---|---|
| Prompt | Minutes | None |
| Saved prompt | Under an hour | Edit when your process changes |
| Skill | An afternoon | An owner, a version, a test case or two |
| MCP / connector | Hours to days | Auth, permissions, API changes |
| Agent | Days to weeks | Evals, logs, cost limits, someone on call |
| App / system | Weeks to months | A team or a vendor, forever |
Reviews
If it has to be an app: ten builders and coding agents, scored the same way
Six criteria, one published rubric, every price linked to its source. None of these products is ours. How we score.
App buildersA hosted app from a description
- Base44Wix-owned builder with an integrated backend on every plan, priced in message and integration credits.
- BoltBrowser-based builder metered in tokens, with Bolt Cloud hosting and database, and two-way GitHub sync.
- LovablePrompt-to-app builder with its own managed backend (Lovable Cloud) and Git sync on every plan.
- Replit AgentAn agent inside Replit's cloud workspace, with database, hosting, rollback and an IDE in one place.
- v0Vercel's agent for Next.js and React apps, with GitHub sync on Free and one-click deploys to Vercel.
Coding agentsWork inside your own repository
- Claude CodeAnthropic's coding agent for the terminal, IDEs, desktop, web and Slack, included in paid Claude plans.
- CursorAI-first code editor with an agent, cloud agents, Bugbot reviews and its own Composer models.
- GitHub CopilotCopilot's cloud agent takes a GitHub issue and returns a pull request, on every paid plan from $10.
- OpenAI CodexOpenAI's coding agent: an open-source CLI, IDE extension, desktop app and cloud tasks, included in paid ChatGPT plans.
- Windsurf (Devin Desktop)The Windsurf editor, renamed Devin Desktop in June 2026, with an agent command center and Devin cloud agents.
Comparisons
Comparisons and guides, from published specs
Prices and features checked against official pages, with a link next to every claim.
- 01AI app builders compared: Lovable, Bolt, v0, Replit, Base44, and the coding agents that come nextLovable, Bolt, v0, Replit and Base44 plus four coding agents compared on price (checked Nov 2026), billing units, lock-in and who should pick which.
- 02Best AI app builder for non-technical founders: ranked for getting an MVP liveLovable, Replit, Bolt, Base44 and v0 ranked for founders who don't code: our rubric re-weighted for speed, a managed backend and a predictable bill. Nov 2026.
- 03Best AI app builder for internal tools: ranked for company dashboards and admin appsReplit, Lovable, v0, Bolt and Base44 ranked for internal tools, re-weighted for SSO, data access and ownership, plus when Retool is the better answer. Nov 2026.
- 04How much does it cost to build an app with Lovable, Bolt, Replit or Base44?Estimate a month of building with Lovable or Base44 from their published per-message credit costs, and plug in your own Bolt and Replit figures. Nov 2026.
- 05Vibe coding security checklist: 10 checks before you launch an AI-built appTen checks for apps built with Lovable, Bolt, Base44, Replit or v0 before real users arrive, each tied to a real incident: RLS, exposed keys, auth, backups.
- 06AI agent frameworks compared: OpenAI Agents SDK, Claude Agent SDK, LangGraph, CrewAI, Mastra and Google ADKOpenAI Agents SDK, Claude Agent SDK, LangGraph, CrewAI, Mastra and Google ADK compared on language, licence, models, MCP and when to pick each. Nov 2026.
- 07Claude Skills vs custom GPTs vs Gemini Gems vs Copilot agentsWhat Claude Skills, custom GPTs, Gemini Gems (now Gemini skills) and Copilot agents can do, who can share them, and the plan you need. Checked Oct 2026.
- 08Where to find Claude skills and MCP serversAnthropic's skills repo, skills.sh, Skillspace, the official MCP Registry, Glama, Smithery and mcp.so compared on size, quality signals and safety.
- 09How to vet an MCP server (or a skill) before you install itA 12-point checklist for MCP servers and agent skills, from the MCP spec, OWASP's MCP Top 10 and four real incidents, with safe settings for popular servers.
Essays
The Moat Line essays
The other half of the question: once you can build it in a weekend, can anyone else?
Start here · 8 minThe Moat Line: what AI made cheap, and what it didn'tThe Moat Line separates software you can type into existence from software you can't. A four-rung test for whether an AI-era idea is defensible.Essay · 6 minIs my startup an AI wrapper? 7 honest signsIs your startup just an AI wrapper? Score it on 8 criteria in a free 2-minute test, then see the seven signs and the move that climbs off each.Essay · 4 minKinetic vs structural moatsA structural moat compounds while you sleep; a kinetic moat evaporates the day you slow down. How to tell which one your company is running on.Essay · 4 minFounder-market-fit is the moat AI can't cloneWhen tools clone any below-the-line product in a weekend, founder-market-fit is the moat left, because the right person on the problem can't be typed.Essay · 5 minBest AI idea validators: an honest comparisonValidatorAI, IdeaProof, DimeADozen and a skeptical ChatGPT or Claude prompt compared, with prices checked Oct 2026 and the one test most of them skip.