| name | classify |
| description | Classify an agent by type and complexity tier. Use at the start of any agent build request to determine how to approach the project. |
Classify Agent
Determine the agent type and complexity tier before any other work begins. This classification drives every downstream decision.
First: reframe the request as an agent to build
Any incoming problem, task, or goal is a request to build an agent for that
class of problem — not a task for you to perform (see the Prime Directive in
CLAUDE.md). Before classifying, restate the request in agent terms to confirm the
framing, e.g. "Got it — you want an agent that [does X repeatably]." Then classify.
Never research the domain or start solving the user's problem here. Classification
is about the agent, not about answering the underlying question.
Agent Type
Identify the dominant mode. If the agent spans modes, note the primary and secondary.
| Mode | Examples | Primary Input | Research Intensity | Top Failure Risks |
|---|
| Creative/Voice | Content marketing, brand copywriting, social media | Style samples, voice preferences, anti-examples | Low — voice > data | Tone drift, generic output |
| Analytical/Research | Market analysis, competitor intel, due diligence | Public data, frameworks, benchmarks | High — data is the product | Stale facts, hallucinated citations |
| Operational/Workflow | Code review, incident response, migration planning | Repo conventions, SOPs, decision trees | Low — process > research | Missed edge cases, wrong defaults |
| Regulated/High-Risk | Healthcare, legal, tax, compliance, security audit | Regulations, internal policies, approval chains | Medium — accuracy is existential | Non-compliant output, overclaiming |
| Internal-Knowledge | Support triage, onboarding, internal tools | CRM data, call transcripts, internal docs | None — all context is private | Confidentiality leaks, stale internal facts |
What the type changes
- Creative agents need anti-examples and voice calibration, not market stats
- Analytical agents need sourced data and citation discipline, not creative freedom
- Operational agents need decision trees and failure examples, not competitor analysis
- Regulated agents need forbidden-output lists and escalation paths, not engagement metrics
- Internal-knowledge agents need private artifact ingestion, not web research
Complexity Tier
Match investment to stakes. Not every agent needs the full framework.
| Tier 1: Minimal | Tier 2: Standard | Tier 3: High-Stakes |
|---|
| When | Personal utility, low risk, single user, narrow scope | Professional use, moderate risk, real audience | Regulated, high liability, multi-stakeholder, brand-critical |
| Examples | Meeting notes formatter, git commit helper, journal prompter | Content marketing agent, sales support, technical writing | Medical documentation, legal compliance, client-facing contractual proposals |
| CLAUDE.md | 30-60 lines, core rules only | 80-120 lines, full structure | 80-150 lines, tight surface — manage risk through external structure |
| Skills | 1-3, lightweight | 3-8, with examples | 3-8, with good+bad examples, validation steps |
| guide.md | Optional | Yes, sourced | Yes, sourced, reviewed, with [Review: date] markers |
| Evals | 5 representative, informal | 10-15 representative + 5 edge + 3 adversarial | Full suite with scoring rubric + release criteria |
| Design phase | Failure list only | Full | Full + source-of-truth precedence |
Default to Tier 2. Upgrade to Tier 3 when: output goes to clients/regulators, mistakes have legal/financial cost, or the user says "this needs to be bulletproof." Downgrade to Tier 1 when: personal, low-risk, and narrow.
Output
State explicitly to the user:
- Agent type: [Primary mode] (secondary: [mode], if applicable)
- Complexity tier: [1/2/3] — because [concrete reason]
- Top 3 risk factors for this type+tier combination
- Research intensity: None / Light / Deep
Ask the user if this classification is correct before proceeding to the interview phase.
