kumaran-is AI Application Playbook
Iron Law: Start one tier below where you think you need to be. Most "agent" problems are workflows; most "workflow" problems are a single model call. Use the smallest pattern that works.
This skill encodes a production-grade framework for deciding when to use AI and how to ship it safely. The full playbook is in playbook.md (10k+ words across four layers); load it only when needed using the routing below.
Prime Directive
Use AI for translation between the messy world and your clean system — never as the clean system itself.
Everything the model reads is untrusted input. Everything the model writes is an untrusted proposal until validated.
Fast triage (use this first)
Before any AI design work, ask in order:
- Is the input messy / unstructured / ambiguous? If no → deterministic code, stop.
- Can a downstream verifier (human or code) catch errors? If no → don't use AI for the decision.
- Is this on the critical latency path with no fallback? If yes → don't use AI here.
- Pick the smallest pattern that works: single model call in a workflow → ReAct loop in a state machine → planning agent → autonomous agent. Start at #1.
If the answer to all three is "AI is appropriate," route to the relevant layer below.
When to load the full playbook
Read playbook.md and apply the matching layer based on what the user is doing:
| User is doing... | Load these layers from playbook.md |
|---|---|
| Deciding whether to add AI to a feature | Layer 1 (Principles) |
| Architecting an AI feature | Layer 1 + Layer 2 (Operating Controls) |
| Writing code: schema validation, output contract, retries, fallback | Layer 2 §2.4, §2.5; Layer 3 §3.1, §3.5 |
| Writing prompts or designing prompts for security | Layer 2 §2.5; Layer 3 §3.2 (adversarial test catalog) |
| Building or reviewing an agent | Layer 2 §2.5, §2.6; Layer 3 §3.2 |
| Designing or evaluating RAG | Layer 2 §2.7; Layer 3 §3.2 retrieval-poisoning row |
| Picking a model / deciding to fine-tune | Layer 2 §2.9 |
| Sizing cost and latency | Layer 2 §2.10, §2.11; Layer 3 §3.4 |
| Designing eval set / regression tests | Layer 2 §2.14; Layer 3 §3.2 |
| Preparing for launch | Layer 4 §4.1 (unified launch checklist) |
| Setting up rollout or canary | Layer 4 §4.2 |
| Handling an incident | Layer 4 §4.3 |
| Testing kill switch | Layer 4 §4.4 |
| Onboarding to AI workflow (first 30 days) | Layer 4 §4.5 |
| Avoiding overbuilding | Layer 4 §4.6 |
For the unified Decision Record template, run the /ai-decision-record skill.
For the launch checklist, run /ai-launch-check.
For incident response, run /ai-incident-response.
Default stance
When advising on AI design work:
- Default to the Implementation Minimum (six items in Layer 2 §2.3). Don't over-engineer on day one.
- Push back when AI is the wrong tool. The Red Flag Matrix (Layer 1 §1.6) lists where deterministic code wins.
- Insist on the Output Contract (Layer 2 §2.4) before any AI output mutates state.
- Treat all model inputs as untrusted — including retrieved documents, tool outputs, file uploads, and prior conversation turns from external sources.
- Quote thumb rules sparingly. Don't recite the playbook at the user; apply it.
When NOT to load the full playbook
- Pure non-AI coding tasks (writing a SQL query, fixing a CSS bug, refactoring deterministic logic).
- Conversations about Claude Code or AI tooling itself, unless the user is building an AI feature using those tools.
- Questions about AI as a topic ("what is RAG?") without an applied design context.
When in doubt: skim the Fast Triage section, decide if AI is even the right answer, and only load playbook.md if it is.
Fast triage example
User: "Should I use an LLM to classify support tickets into 5 buckets?"
Fast triage:
1. Is input messy/unstructured? YES — free-text tickets.
2. Can a downstream verifier catch errors? YES — human review queue + SLA.
3. On critical latency path with no fallback? NO — async classification.
→ AI is appropriate. Pattern: workflow (single model call + output contract).
NOT an agent. Routing: Layer 2 §2.9 (model selection) + §2.14 (eval set: 100–300, precision/recall per class).
User: "I want to auto-approve expense reports under $100 using an LLM."
Fast triage:
1. Is input messy/unstructured? PARTIAL — structured form but free-text justification.
2. Can a downstream verifier catch errors? NO — auto-approval with no human check = direct financial action.
→ AI is NOT appropriate for auto-approval. Red Flag Matrix hit: Tier-4 action with no verifier.
Recommend: AI as a draft/suggest assistant only, human approves.
Verify (before routing to a lifecycle skill)
✅ Fast triage completed — AI is appropriate for this use case
✅ Correct playbook layer(s) identified from the routing table
✅ If user is starting a new feature → directed to /ai-decision-record
✅ If user is preparing to launch → directed to /ai-launch-check
✅ If there is a production incident → directed to /ai-incident-response
✅ Only loaded playbook.md sections relevant to the current task (not the full 10k words)
If this skill cannot proceed
Situation: fast triage returns all NO (deterministic code is sufficient)
→ Do not load playbook.md. Output: "Based on the fast triage, this looks like a
deterministic code problem. AI would add cost and complexity without benefit."
Situation: user insists on AI after triage says no
→ Apply the Red Flag Matrix (Layer 1 §1.6). Name the specific flags.
→ Output: "The Red Flag Matrix flags this as [reason]. I recommend against AI here.
If you want to proceed anyway, run /ai-decision-record to document the justification."
References
- Full playbook (4 layers):
.claude/skills/ai-playbook/playbook.md - Layer 1 §1.6 Red Flag Matrix: when NOT to use AI
- Layer 2 §2.3 Implementation Minimum: 6 required items before shipping any AI feature
- Layer 2 §2.4 Output Contract: schema → types → business rules → permissions → confidence → audit → route
- Layer 3 §3.2 Adversarial test catalog: prompt injection, tool misuse, retrieval poisoning
- Lifecycle skills:
/ai-decision-record→/ai-launch-check→/ai-incident-response→/ai-audit - OWASP LLM Top 10: https://owasp.org/www-project-top-10-for-large-language-model-applications/