ak-predict

name	ak:predict
description	5 expert personas debate proposed changes before implementation. Catches architectural, security, performance, and UX issues early. Use before major features or risky changes.
user-invocable	true
when_to_use	Invoke before high-risk changes that need persona debate.
category	utilities
keywords	["prediction","debate","review","risk"]
argument-hint	<feature description or change proposal> [--files <glob>] [--chain reason|probe]
metadata	{"author":"agentkit","attribution":"Multi-persona prediction pattern adapted from autoresearch by Udit Goenka (MIT)","license":"MIT","version":"1.0.0"}

ak:predict — Multi-Persona Pre-Analysis

Five expert personas independently analyze a proposed change, then debate conflicts to produce a consensus verdict before a single line of code is written.

When to Use

• Before implementing a major or high-risk feature
• Before a significant refactor or architecture change
• Evaluating competing technical approaches
• Stress-testing assumptions in a proposed design

When NOT to Use

• Trivial or low-risk changes (use ak:debug for bugs, ak:plan for already-decided tasks)
• Already-approved work with no open design questions
• Pure dependency upgrades with no API changes

---

The 5 Personas

Persona	Focus	Core Questions
Architect	System design, scalability, coupling	Does this fit the architecture? Will it scale? What new coupling does it introduce?
Security	Attack surface, data protection, auth	What can be abused? Where is data exposed? Are auth boundaries respected?
Performance	Latency, memory, queries, bundle size	What is the latency impact? N+1 queries? Memory leaks? Bundle bloat?
UX	User experience, accessibility, error states	Is this intuitive? What does the error state look like? Accessible on mobile?
Devil's Advocate	Hidden assumptions, simpler alternatives	Why not do nothing? What is the simplest alternative? Which assumption could be wrong?

---

Debate Protocol

1. Read the proposed change/feature description from the argument
2. Read relevant code if file paths are provided (grep for affected areas)
3. Each persona analyzes independently — do not let personas influence each other during this phase
4. Identify agreements — points where all (or 4+) personas align
5. Identify conflicts — points where personas meaningfully disagree
6. Weigh tradeoffs — for each conflict, evaluate which concern has higher impact
7. Produce verdict — GO / CAUTION / STOP with actionable recommendations

---

Output Format

## Prediction Report: [proposal title]

## Verdict: GO | CAUTION | STOP

### Agreements (all personas align)
- [Point 1 — what they all agree on]
- [Point 2]

### Conflicts & Resolutions

| Topic | Architect | Security | Performance | UX | Devil's Advocate | Resolution |
|-------|-----------|----------|-------------|-----|-----------------|------------|
| [Issue] | [View] | [View] | [View] | [View] | [View] | [Recommendation] |

### Risk Summary

| Risk | Severity | Mitigation |
|------|----------|------------|
| [Risk description] | Critical/High/Medium/Low | [Concrete action] |

### Recommendations
1. [Action item — rationale]
2. [Action item — rationale]
3. [Action item — rationale]

---

Verdict Levels

Verdict	Meaning
GO	All personas aligned, no critical risks, proceed with confidence
CAUTION	Concerns exist but are manageable — mitigations identified, proceed carefully
STOP	Critical unresolved issue found — needs redesign or more information before proceeding

STOP Triggers (any one is sufficient)

• Security persona identifies auth bypass or data exposure with no viable mitigation
• Architect identifies fundamental design incompatibility requiring significant rework
• Performance persona identifies unacceptable latency or query explosion with no workaround
• Devil's Advocate exposes a false assumption that invalidates the entire approach

---

Chain Modes

After producing the verdict, predict can chain into a follow-on workflow that always runs as part of a predict session (not as a standalone skill).

Flag	Purpose	When to use
--chain reason	Subjective refinement loop — generate → critique → synthesize → blind judge → repeat until convergence	Verdict is CAUTION with subjective tradeoffs (architecture polish, design coherence)
--chain probe	Requirement interrogation — saturation-driven harvest of missing constraints + assumptions	Verdict is CAUTION or STOP because of "missing constraint" or "unstated assumption" findings

Use the chain-mode summaries below as the supported protocol for this kit.

These chain modes absorb upstream /autoresearch:reason and /autoresearch:probe (uditgoenka/autoresearch, MIT). They're folded into ak:predict — not shipped as standalone skills — because they always chain off a predict invocation. See /ak:autoresearch for the family map.

---

Integration with Other Skills

Workflow Step	Skill	How
Deepen risk scenarios	ak:scenario	Feed Risk Summary rows as feature description
Create implementation plan	ak:plan	Attach Recommendations as constraints to planner
High-risk feature implementation	ak:cook	Reference CAUTION/STOP items as acceptance gates

---

Example Invocations

/ak:predict "Add WebSocket support for real-time notifications"
/ak:predict "Migrate authentication from JWT to session cookies"
/ak:predict "Add multi-tenancy to the database layer"
/ak:predict "Replace REST API with GraphQL" --files src/api/**/*.ts

# Chain modes
/ak:predict "Pick auth library: Passport vs Better Auth" --chain reason
/ak:predict "Move from REST to GraphQL" --chain probe

---

Lineage

Faithful absorption (in scope) of upstream /autoresearch:predict (uditgoenka/autoresearch, MIT). The local version supports the 5-persona debate plus --chain reason (subjective refinement) and --chain probe (requirement interrogation), folding upstream's /autoresearch:reason and /autoresearch:probe sub-commands into chain modes rather than separate skills (closed in #728).

See /ak:autoresearch for the full family map.