core

name	core
description	prd.json schema and task system - auto-loaded when prd.json exists
allowed-tools	Read, Write, Edit, TaskCreate, TaskUpdate, TaskList, Grep, Glob
model	opus
user-invocable	false
disable-model-invocation	true

Hybrid Task System

Sprint Summary

!node -e "try{const p=require('./prd.json');const sp=p.sprints?p.sprints[p.sprints.length-1]:p;const s=Object.values(sp.stories||p.stories||{});const name=sp.id||sp.name||p.sprint||'unknown';const done=s.filter(x=>x.passes===true).length;const pending=s.filter(x=>x.passes===null||x.passes===false).length;const deferred=s.filter(x=>x.passes==='deferred').length;console.log('Sprint:',name,'| Done:',done,'| Pending:',pending,'| Deferred:',deferred,'| Total:',s.length)}catch(e){console.log('No prd.json')}"

For large prd.json (100+ stories), use Grep to find specific stories. For typical sizes (<50 stories), reading the full file is fine with 1M context.

When to Sprint

• 5+ related tasks — create a sprint in prd.json
• < 5 tasks or single fixes — work directly, no sprint or stories needed
• Design/creative work — iterate freely, skip planning overhead
• Quick fixes — just fix, verify, done

Sprints are for tracking, not for ceremony. If the work is small, skip the overhead.

Two Layers

Layer	Tool	Purpose
Long-term	prd.json	Sprint history, resolutions (Git-tracked)
Short-term	Native Tasks	Active work (session only)

Work with native Tasks during session, batch-update prd.json at end.

prd.json Story Schema

{
  "id": "S26-001",
  "title": "Fix tooltip clipping",
  "priority": 1,
  "passes": null,
  "realness": null,
  "type": "fix",
  "category": "components",
  "notes": "",
  "resolution": ""
}

Field	Values
passes	null (pending), true (done), false (failed), "deferred"
realness	0-100 (optional, see below). null = not rated yet
type	fix, feature, refactor, qa, perf
priority	0=critical, 1=high, 2=medium, 3=low
resolution	HOW it was fixed (learning)

Realness Scale (0-100)

passes: true is binary and hides the difference between "wired up" and "works in production." Fill in realness when closing non-trivial stories so the sprint summary reflects reality.

Score	Meaning
20	Stubbed — UI exists, no backend
40	Wired — frontend and backend connected, happy path works in dev
60	Functional — handles the obvious edge cases, one real end-to-end test
80	Production-ready — error handling, empty/loading/error states, observability
100	Battle-tested — used by real users, edge cases caught and fixed

Rules:

• Bug fixes default to 80 (fixing a real issue is usually production-ready).
• Features require a manual rating — don't auto-assign 100 just because passes: true.
• When in doubt, pick the lower number.
• Sprint summaries report the AVERAGE realness of closed stories, not just the pass count.

Resolution Learning

When completing bug fixes, document HOW:

[PATTERN]: [SPECIFIC FIX]

Examples:

• null-check: Added optional chaining at line 45
• missing-import: Added import for DateRange
• type-mismatch: Changed Record<string, T> to Partial<Record<K, T>>
• overflow: Added max-h + overflow-auto

Context Tips

With 1M context, aggressive token saving is unnecessary. Prefer clarity over brevity.

Action	Do This
Check status	Read prd.json header or use dynamic context injection
Start task	Grep specific story
Track progress	Native TaskUpdate
Complete work	Batch edit prd.json at session end

Archive Trigger

When ANY of these are true, run archive before starting new work:

• 4+ total sprints exist in prd.json
• prd.json > 500 lines
• prd.json > 50KB
• Starting a new sprint while previous sprint's stories are all complete

Do not ask — just archive. Archive keeps only last 3 sprints active. Completed stories move to .claude/archives/prd-archive-YYYY-MM.json.