AI Work Cycle
Skill File

---
name: ai-work-cycle
description: Unified 5-phase cycle (Plan → Work → Review → Triage → Learn) for both coding and knowledge-work tasks. Auto-detects the work type and routes to specialized sub-pipelines. Survives context compaction via TodoWrite, blocks shipment on P1 findings.
---

# AI Work Cycle

One skill, two paths. This skill is a thin orchestrator — it classifies the work, routes to the right pipeline, and wraps the whole thing in a mandatory review gate plus a Learn phase that compounds.

## The Five Phases

| # | Phase | Code path | Knowledge-work path |
|---|-------|-----------|---------------------|
| 1 | Plan | /yalla-plan (4-agent planning) | Task(knowledge-work:plan) |
| 2 | Work | /yalla build OR inline + TodoWrite | Task(knowledge-work:create) |
| 3 | Review | /yalla-review + Task(qa-validator) | Task(knowledge-work:review) |
| 4 | Triage | Inline P1/P2/P3 consolidation | Task(knowledge-work:triage) |
| 5 | Learn | INSERT INTO memory_knowledge | INSERT INTO memory_knowledge |

Note: /yalla-plan, /yalla-review, and /yalla are the proven code pipelines you already have. The knowledge-work agents are specialized sub-agents. Swap for your own equivalents if you use different tools.

## How to Invoke

/ai-work-cycle [description]
/ai-work-cycle task-1234
/ai-work-cycle mode=code [description]
/ai-work-cycle mode=knowledge [description]

If given a task ID, load context from dev_tasks first.

## Auto-Detect: Code or Knowledge Work?

| Signal | Path |
|--------|------|
| Mentions files, functions, endpoints, tests, UI, database, bugs | Code |
| Mentions strategy, analysis, deck, doc, research, brief, memo | Knowledge |
| Ambiguous | Ask once, then route |

Force path: mode=code or mode=knowledge.

## Phase 1: Plan

Code: Skill /yalla-plan runs a 4-agent adversarial planning team (codebase-analyst, solution-architect, spec-validator, red-team) → plans/[feature].md
Knowledge: Task knowledge-work:plan → brief with research, success criteria, outline

Always: create a comprehensive TodoWrite list before Phase 2. Todos survive context compaction.

## Phase 2: Work

Code: Skill /yalla handles track → plan → work → test → review → ship in one shot. Or implement manually from the plan with TodoWrite checkpoints.
Knowledge: Task knowledge-work:create builds section by section against the brief.

## Phase 3: Review (Mandatory Gate)

Code:
1. Skill /yalla-review — binary pass/fail checks (security, types, tests, conventions)
2. Task qa-validator — browser, database, API, user flows
3. P1 findings BLOCK shipment (SQL injection, XSS, SSRF, data corruption, bypass, failing build, broken user flow)

Knowledge:
1. Task knowledge-work:review runs 12+ specialists in parallel:
   - clarity-maximizer, completeness-validator, structure-architect,
     professional-polish-editor, stakeholder-alignment-checker, impact-analyzer
2. Output: severity-categorized findings

## Phase 4: Triage

Code: Consolidate /yalla-review and qa-validator output into P1/P2/P3 with effort estimates.
Knowledge: Task knowledge-work:triage → prioritized revision list with time estimates.

Pause here. You decide what ships vs. what's deferred.

## Phase 5: Learn (Don't Skip)

Extract 1-3 reusable patterns to memory_knowledge via SQL:

INSERT INTO memory_knowledge (owner_id, type, title, content, tags, source, confidence, decay_score)
VALUES ('you', 'pattern', 'Short title', 'What I learned and when to apply', '["tag"]'::jsonb, 'task-XXXX', 0.85, 1.0);

Capture: non-obvious gotchas, reusable structures, decisions with rationale.
Skip: "fixed typo" — not reusable.

## Override: Ship With Known P1s

I accept risks: [list P1 issues]
Business reason: [why ships now]
Mitigation: [follow-up PR #XXXX]
Acknowledged by: [your name]

Goes into commit message. Audit trail.

## Execution Strategy

- Phases 1 → 2 → 3: run automatically
- Before Phase 4 (Triage): pause for your review
- Phase 5 (Learn): quick, but explicit — do not skip

Say "stop" to interrupt.

## Required Sub-Pipelines

Code path:
- Skill: /yalla-plan (4-agent adversarial planning)
- Skill: /yalla (optional — full autonomous build pipeline)
- Skill: /yalla-review (binary pass/fail review)
- Agent: qa-validator

Knowledge path:
- Agent: knowledge-work:plan, create, review, triage

If any missing, I fall back to inline execution and tell you what's unavailable.