ai-native-tool-stack-selection

Selects the smallest workable AI-native tool stack using a 12-category decision tree (thinking partner, research, design, app builder, coding agent, hosting, backend, skills, project management, notes

SKILL.md

GitHub

AI-Native Tool Stack Selection

Purpose

Help a pre-startup founder or small team choose the smallest workable AI-native tool stack without drowning in a market of hundreds of tools. This skill structures the choice around categories — not product names — because categories (what purpose you need the tool for) hold up over time, while individual products and their features go stale quickly in this market.

Based on

  • The owner's AI-native Business Design workshop (the owner's own workshop), tools.md — "2026 AI-Native Stack": a 12-category breakdown organized by what you're trying to do, not by vendor; the "minimum viable stack" principle ("3–6 tools, not 30"); the three-tier maturity path for agent tools (no-code platforms → open- source runtimes → developer frameworks).
  • See ../../references/tool-category-map.md (12 categories with examples, a time-stamped snapshot) and ../../references/workshop-source.md.

Method

  1. Go through the 12 categories (see ../../references/tool-category-map.md) and identify which are necessary for YOUR own case right now — not all of them at once:
    1. AI thinking partner (general chat/project AI)
    2. Research and information retrieval
    3. Design sketching
    4. App builder (prompt → working app)
    5. Coding agent (when moving from prototype to production)
    6. Version control / code storage
    7. Hosting and deployment
    8. Backend and database
    9. Skills (packaging and reusing agent capabilities)
    10. Project management
    11. Meeting/note-taking tool (turning conversations into machine-readable text)
    12. Workflow automation and agent building
  2. Choose one default tool for each necessary category. Resist the urge to use multiple tools in the same category simultaneously — it fragments context and slows things down instead of speeding them up.
  3. Apply the minimum-stack rule of thumb. A typical working pre-startup stack is 3–6 tools, not 30. Start with the minimum: thinking partner + research + design sketch + app builder + code storage. Add categories only once a genuine need arises, not proactively.
  4. As the team or need grows: add project management and a meeting notes tool only once multiple people are working on the same thing regularly — not right at the start.
  5. Once the first workflow has proven valuable and is genuinely closed-loop in shape (see ../closed-loop-process-and-human-oversight-design/SKILL.md): consider wrapping it as an agent. Start with a no-code agent platform; move to an open-source runtime or a developer framework only once technical skill and genuine need require it — not by default.
  6. Check lock-in before committing. What infrastructure (database, hosting) does the tool tie you to, and can code/data be exported if needed? For a prototype this often doesn't matter much; for a product meant to scale, it matters a lot.
  7. Remember this is a snapshot. Tool lists, pricing, and free tiers change weekly in this market. Always check a tool's current status before committing — don't treat the named examples in ../../references/tool-category-map.md as an up-to-date truth.

Gotchas

  • Adopting two tools in the same category "just to compare" is explicitly against step 2's rule of thumb — it fragments context and slows things down rather than speeding them up, even though it feels like hedging risk.
  • Adding project management or meeting-notes tooling before multiple people are actually working on the same thing regularly (step 4) is premature scaffolding for a stack that's supposed to start at 3-6 tools, not 30.
  • The named tools in ../../references/tool-category-map.md are a time-stamped snapshot, not a current recommendation — step 7 warns that pricing, features, and free tiers in this market shift weekly, so re-verify a tool's status before committing rather than trusting the reference file as-is.
  • Jumping straight to an open-source runtime or developer framework (LangGraph, CrewAI, etc.) for agent-building skips the required progression in step 5: start with a no-code platform, and only move up once a workflow has proven genuinely closed-loop-shaped and real technical need exists — not by default enthusiasm.
  • Skipping the lock-in check (step 6) is low-risk for a throwaway prototype but becomes a real liability if the product is meant to scale — the skill only flags this as optional for prototypes, not for anything intended to grow past MVP.

What this skill does NOT do

  • Does not recommend specific product names as a permanent truth — categories and the selection principle hold up, individual products go stale quickly (see the timestamp in ../../references/tool-category-map.md).
  • Does not perform technical due diligence on a tool's security, contract terms, or the safety of skills/agents — check that separately before business-critical use; install skills/agents only from trusted sources.
  • Does not assess which developer framework (LangGraph, CrewAI, Claude Agent SDK, etc.) a dev team should choose for production code — that's the dev team's decision; this skill only gives the founder context on what's involved before that conversation.

Continue from here

  • Related skill in this pack: ../ai-buildable-prd-writing/SKILL.md (who the PRD is handed to for building), ../closed-loop-process-and-human-oversight-design/SKILL.md (when it makes sense to move to agents).
  • Related skill in another pack: ../../../../ai-strategy-and-governance/skills/build-vs-buy-vs-partner-ai/SKILL.md — a larger-scale build/buy/partner decision; this skill is a lighter, tactical choice for the pre-startup stage.
  • The pack's shared guardrails: ../../CLAUDE.md

References

  • ../../references/tool-category-map.md — 12 categories with examples (a time-stamped snapshot)
  • ../../references/workshop-source.md — source information
  • ../../CLAUDE.md — the pack's shared guardrails