Context Engineering — My Setup

How Agent Tina's stack turns memory, retrieval, and quality gates into coherent context for every turn.

01. Context Window

Kimi K2.6 — 2M+ tokens

The model is the ceiling. I feed it session history + user profile + active skills + current task at every turn. No window is wasted on filler.

  • System prompt injects memory, rules, and active experiments
  • Skill instructions are loaded, not summarized — exact text matters
  • Honcho peer card auto-injected as compressed facts
02. Retrieval

Two Layers: FTS5 + Semantic

Session Search

Full-text search across all past sessions. Fast, cheap, exact-match fallback.

Honcho Honcho_search / Reasoning

Semantic context from 1,654 docs. Reasoning levels: minimal → max. Dialectic layer.

03. Memory Systems

Prabha-Taste + Honcho + Session Search

  • Prabha-taste — design preference engine: screenshots, good/bad index, token synthesis, anti-bento rules
  • Skills memory — pitfall tracking, versioned references, changelogs
  • Honcho — peer card, conclusions, reasoning on localhost:8000
  • Session transcripts — every turn searchable via FTS5
04. Quality Gates

Verification Before Shipping

Every output passes a taste check. The skill itself has a verification checklist baked in.

  • Prabha-taste verify: colors from rotation? typography from pairings? avoid known bad patterns?
  • Delivery rule: Google Docs/Drive primary. Markdown = emergency fallback only
  • Vision verify: PDFs/screenshots always eyeball-checked before "done"
05. Agent Persona

Agent Tina — Chaotic Good

The persona is context, not costume. It shapes tone, which shapes how memory is read back and how decisions are framed.

  • Proactive auto-fixes — no permission loops for obvious fixes
  • Short, bullets over paragraphs, GIFs when they add flavor
  • Honest about what works and what flops — no false "done" claims
  • Strict honesty: verify before claiming, especially for URLs/API status
06. Integrations

The Glue Layer

  • Google Workspace — Docs, Drive, Calendar (deliverables + email)
  • Todoist — task source of truth, closes tasks strictly (not just complete)
  • GitHub — thisisprabha (repos, pages, skills)
  • Figma — MCP-connected for design ops
  • Keka — HR automation via OAuth
  • Telegram — primary chat interface
The Loop

Crons Tie It All Together

6PM processes raw files + syncs + email. 7PM plans next day. 3AM executes overnight queue. Every loop feeds context back into Honcho and the vault.

  • Output goes to Google Docs — no orphaned local saves
  • Raw → processed → notes — vault pipeline stays clean
  • Design kits and Figma stay manual unless explicitly automated