assess

Assess — Project Type Recommendation

Purpose

When a client gives a raw requirement, this skill classifies the project type and recommends the right approach BEFORE decomposition begins. It answers the questions that /decompose and chief-ai-po don't ask: what KIND of project is this, what agents will it need, and what's the right architecture approach?

Without /assess, the pipeline treats every project identically. With it, North Starr adapts its agent activation, pipeline configuration, and complexity expectations to the specific project type.

When to Use

• Before /decompose — when a client gives a raw requirement and you need to understand what kind of project it is
• Before /discover — when you want to validate your understanding of the project type before eliciting detailed requirements
• On demand — when mid-project you realize the project type has shifted (e.g., what started as an automation pipeline now needs multi-agent orchestration)

For projects where the type is already clear and documented, skip directly to /decompose.

Input

The user provides one of:

• Pasted text — a raw requirement, brief, or description
• File path — a document (markdown, text, PDF) containing the requirement
• Conversation summary — output from /discover or a prior discussion

Workflow

Step 1: Read & Understand the Requirement

Actions:

1. If a file path is provided, read the content (use PDF pagination for large PDFs)
2. If text is pasted, use it directly
3. Extract the core intent: what problem is being solved, for whom, at what scale
4. Identify any explicit constraints mentioned (budget, timeline, compliance, existing infrastructure)

Step 2: Classify Project Type

Evaluate the requirement against these project types:

Type	Signals	Example
Automation Pipeline	Event-driven, classification, extraction, routing, batch processing, document processing	"Categorize support tickets and route to the right team"
Agent Harness	Production agent runtime, tool use, memory, safety-critical, long-running tasks	"Build an agent that researches competitors and writes reports"
Multi-Agent System	Multiple agents collaborating, supervisor/sub-agent, peer-to-peer, pipeline of agents	"Research agent feeds writer agent, reviewer checks output"
RAG Application	Knowledge base, document Q&A, search + generation, citations needed	"Let employees ask questions about our internal policies"
Prompt Chain	Sequential LLM calls, transformation pipeline, no agent autonomy	"Summarize → translate → format for each incoming document"
AI OS Component	Shared platform layer, multi-tenant, cross-team, extensibility required	"Build a shared knowledge layer all teams can query"
Hybrid	Combines two or more of the above	"RAG-powered agent with tool use and multi-step reasoning"

If the requirement clearly maps to one type, classify it. If it spans multiple types, classify as Hybrid and list the component types.

Step 3: Determine Architecture Approach

Based on the project type, recommend the high-level architecture:

Automation Pipeline:

• Topology: event source → preprocessor → AI call → post-processor → output sink
• Key decisions: batch vs streaming, model selection, error handling strategy
• Typical components: input parser, prompt template, output validator, routing logic

Agent Harness:

• Topology: agent loop (observe → think → act → reflect) with tool registry and safety layer
• Key decisions: memory strategy (stateless/session/persistent), checkpoint design, tool authorization
• Typical components: agent runtime, tool definitions, safety guardrails, state management

Multi-Agent System:

• Topology selection heuristic:
  • Agents process data in sequence (A→B→C) → Pipeline topology
  • One agent delegates tasks to specialists → Supervisor / sub-agent topology
  • Agents review or critique each other's work → Peer-to-peer topology
  • Complex delegation with sub-teams → Hierarchical topology
• Recommend a specific topology in the sketch based on the requirement signals. Name it.
• Key decisions: state sharing (message passing vs shared memory), loop control (max iterations, cost cap), handoff protocol
• Typical components: per-agent definitions, orchestrator/supervisor (if not pipeline), shared state store, inter-agent message format

RAG Application:

• Topology: ingest pipeline → vector store ← query pipeline → generation
• If the RAG app includes a chat interface → add: conversation manager, session memory, escalation router
• Key decisions: chunking strategy, embedding model, retrieval approach, re-ranking
• Typical components: document ingester, embedding service, vector DB, retrieval layer, prompt with context injection

Prompt Chain:

• Topology: sequential LLM calls with transformation between stages
• Key decisions: chain decomposition, intermediate validation, error propagation
• Typical components: prompt templates per stage, intermediate validators, chain orchestrator

AI OS Component:

• Topology: platform service with API layer, multi-tenant isolation, extensibility hooks
• Key decisions: multi-tenancy model, API versioning, extensibility patterns
• Typical components: API gateway, tenant isolation, plugin system, shared model configs

Hybrid (combines two or more types):

• Identify the primary type (the one that defines the core data flow) and the secondary type(s) that extend it
• Compose the architecture sketch by starting with the primary type's topology and layering in components from secondary types
• Example: RAG Application (primary) + Agent Harness (secondary) = ingest pipeline → vector store ← agent loop (observe → retrieve → reason → act) with safety layer
• Name each component and its source type in the sketch so the ai-architect knows which patterns to apply

Step 4: Map Agent Activation

Based on the project type, determine which pipeline agents are needed:

Agent	Automation	Agent Harness	Multi-Agent	RAG App	Prompt Chain	AI OS
chief-ai-po	Yes	Yes (all 6 safety stories)	Yes (per-agent + orchestration stories)	Yes	Yes	Yes
ai-architect	Yes	Yes	Yes (topology patterns critical)	Yes	Lightweight	Yes (platform concerns)
prompt-engineer	Yes	Yes (agent system prompts)	Yes (per-agent prompts)	Yes	Yes (per-stage prompts)	Yes
rag-advisor	Only if knowledge base	If retrieval-augmented	If shared retrieval	Critical	No	If shared knowledge layer
integration-planner	If external APIs	Yes (tool access)	If external communication	If external data sources	If external APIs	Yes
eval-designer	Yes	Yes	Yes (per-agent + end-to-end)	Yes	Yes (per-stage + end-to-end)	Yes
guardrails-designer	If public-facing	Critical	Yes	Yes	If public-facing	Critical
prompt-adversary	If public-facing	Critical	Yes	Yes	If public-facing	Critical
ai-ops	Lightweight	Yes	Yes (per-agent monitoring)	Yes	Lightweight	Critical
cost-estimator	Yes	Yes	Yes	Yes	Yes	Yes
demo-builder	Yes	Yes	Yes	Yes	Yes	Yes (with onboarding docs)
agentic-designer	If UI needed	If UI needed	If UI needed	If UI needed	No	If UI needed

Step 5: Estimate Complexity & Flag Risks

Complexity estimation:

• S (Small): Single AI call, clear input/output, no external integrations. 1-2 stories.
• M (Medium): Multiple AI calls or RAG pipeline, 1-2 integrations, moderate safety concerns. 2-4 stories.
• L (Large): Agent with tools, multi-step workflows, compliance requirements, multiple integrations. 4-8 stories.
• XL (Extra Large): Multi-agent system, platform component, real-time requirements, regulatory compliance. 8+ stories (recommend phased delivery).

Risk flags — check each. For every flag that applies, you MUST state (a) the concrete impact on this project and (b) a specific mitigation or next step. Do not list a risk without both.

• Needs fine-tuning — domain-specific style or format that prompt engineering alone won't achieve
• Needs real-time — streaming, sub-second latency, WebSocket connections
• Needs compliance review — PII, HIPAA, GDPR, financial regulations, audit trails
• Needs existing infrastructure — depends on databases, APIs, or services not yet available
• Needs credential provisioning — external API keys, OAuth flows, service accounts
• High cost risk — large volume, expensive models, or open-ended generation
• Novel problem — no established pattern in the team's experience
• Multi-team dependency — requires coordination with other teams or external vendors

Example of a properly flagged risk:

Needs credential provisioning — Impact: Zendesk API requires OAuth2 service account; BUILD phase will block until credentials are provisioned. Mitigation: Request credentials in parallel with DESIGN phase; integration-planner will flag BLOCKED if missing at BUILD start.

Step 6: Write Assessment

Write to .plans/ASSESS-<name>.md:

# Assessment: <name>

**Created:** <date>
**Status:** ACTIVE
**Source:** <file path or "requirement text">

## Requirement Summary

<2-3 sentences capturing the core requirement>

## Project Type

**Classification:** <Automation Pipeline / Agent Harness / Multi-Agent System / RAG Application / Prompt Chain / AI OS Component / Hybrid>

**Rationale:** <Why this classification — what signals pointed to this type>

[If Hybrid:]
**Component types:** <list the sub-types>

## Architecture Sketch

**Topology:** <high-level component flow>

**Key components:**
- <component 1> — <responsibility>
- <component 2> — <responsibility>

**Key decisions to make in DESIGN phase:**
- <decision 1>
- <decision 2>

## Technology Recommendations

- **Model(s):** <recommended starting point, to be validated by ai-architect>
- **RAG:** <needed / not needed / conditional — reason>
- **Vector DB:** <if RAG needed, initial recommendation>
- **External integrations:** <list>
- **Infrastructure:** <what the codebase needs that may not exist>

## Agent Activation Map

| Agent | Needed | Priority | Notes |
|---|---|---|---|
| chief-ai-po | Yes/No | Critical/Standard/Lightweight | <context> |
| ai-architect | Yes/No | Critical/Standard/Lightweight | <context> |
| ... | ... | ... | ... |

## Complexity & Effort

**Complexity:** <S / M / L / XL>
**Estimated stories:** <range>
**Recommended delivery approach:** <single sprint / phased / MVP + iterations>

## Risk Flags

- <flag icon> **<risk>** — <impact and recommended mitigation>

## Recommendation

<One of:>
- **Proceed to `/decompose`** — requirement is clear enough for story decomposition
- **Run `/discover` first** — requirement needs more detail before decomposition (list what's missing)
- **Needs human discussion** — <specific concern that requires human judgment>

Step 7: Present Results & Recommend Next Step

Present the assessment summary to the user:

Assessment: <name>
──────────────────
Project type:      <type>
Complexity:        <S/M/L/XL>
Estimated stories: <range>
Risk flags:        <count> (<list key risks>)
Agent activation:  <count> of 12 agents needed

Key architecture decisions:
  • <decision 1>
  • <decision 2>

Recommendation: <proceed to /decompose | run /discover first | needs discussion>

If recommending /decompose, ask:

"Ready to decompose into stories? Run /decompose with this requirement."

If recommending /discover, ask:

"The requirement needs more detail before I can decompose it. Run /discover to elicit the missing information."

If flagging for discussion, explain what needs human input.

Notes

• This skill is intentionally lightweight — it should take 1-2 minutes, not 30 minutes
• The architecture sketch is high-level, NOT a full ADR — ai-architect produces that in DESIGN phase
• Agent activation map tells the orchestrator which agents to dispatch, avoiding unnecessary specialist work
• Complexity estimates use AI context budget sizing (same as story sizing in /decompose)
• Risk flags feed directly into the inversion analysis — they become the starting point for /ai-invert
• If the project type changes mid-execution (discovered during BUILD), re-run /assess to update the agent activation map