Profiler

Audience intelligence for enterprise work. Open a profile to see how a person prefers work framed, reviewed, and communicated. Drop in an artifact and get sharp, specific recommendations on how to land it.

A working prototype of the app spec'd in PLAN.md.

Quick start

pnpm install
pnpm dev

Open http://localhost:3000.

To enable real LLM-powered analysis, set an ANTHROPIC_API_KEY in .env.local:

echo "ANTHROPIC_API_KEY=sk-ant-..." > .env.local
pnpm dev

Without the key, the app runs in mock mode — analysis is generated deterministically from the selected people + objectives. Output is real and profile-grounded; just less nuanced than Claude. Mock mode is clearly labelled in the UI.

What's built

• Home — entry CTAs, featured profiles, recent analyses, saved audiences.
• People directory — 8 hand-authored profiles, searchable + tag-filterable.
• Person profile — overview, communication & presentation preferences, decision triggers, predictable objections, dos/don'ts, example guidance, inline "Analyze artifact" and "Add to audience" actions.
• Objectives library — 7 objectives with success criteria, common risks, and recommended framing. Multi-select to feed the audience builder.
• Audience builder — multi-select people and objectives, save and load named audiences, live preview of the audience composition.
• Artifact analyzer — paste or upload markdown/text, pick audience, pick (optional) objectives, generate recommendations. Three sample artifacts ship with the prototype for instant demo.
• Results — fit score gauge, executive summary, audience read, key risks (each tied to a named person or objective), recommended framing, tactical edits with before/after, narrative structure, emphasize/avoid lists, meeting/readout approach for multi-person audiences, revised artifact (copyable + downloadable markdown).

LLM design

• Model: claude-sonnet-4-6 by default.
• Structured output: forced tool use (submit_recommendation). The recommendation schema lives in lib/llm/schema.ts.
• Prompt caching: the system prompt and the full reference library (every person and objective) are cached. Only the selected subset plus the artifact varies per request, so repeat analyses in a session are fast and cheap.
• Persona: a senior chief-of-staff / design strategist. Banned hedge phrases, mandatory specificity, risks must be tied to named audience members or objectives, before/after rewrites required when prose is worth rewriting.

See lib/llm/prompts.ts and lib/llm/analyze.ts.

Project layout

app/
  page.tsx                  Home dashboard
  people/                   Directory + [personId] profile
  objectives/               Library
  audience/                 Builder
  analyze/                  Artifact analyzer (uses ?personIds=…&objectiveIds=…&strategy=1)
  results/[resultId]        Recommendation view
  actions.ts                Server action wrapper for runAnalysis
components/
  layout/                   Sidebar, Topbar
  ui/                       Button, Card, Badge, Input, Avatar (mini shadcn-style)
  people/                   PersonCard
  audience/                 AddToAudience
  analyzer/                 AnalyzeForm
lib/
  data/
    people.ts               8 typed profiles
    objectives.ts           7 typed objectives
    sample-artifacts.ts     3 demo artifacts
  llm/
    prompts.ts              System prompt + serializers
    schema.ts               JSON-Schema for forced tool use
    analyze.ts              Orchestrator (Anthropic SDK + caching, with mock fallback)
    mock.ts                 Deterministic recommendation builder
  store.ts                  Zustand store (audiences, recents, results)
  types.ts                  Person / Objective / Artifact / RecommendationResult
  utils.ts                  cn(), initials(), id()
scripts/
  smoke-analyze.ts          Mock recommendation smoke test

Persistence

Pure client-side: Zustand + localStorage. No database. Generated results live in the same store keyed by resultId, so links survive refresh on the machine that generated them but don't transfer across browsers. Swap to SQLite + Drizzle if persistence across users matters.

Known limitations

• File parsing: text and markdown only. PDF and DOCX support is staged in the plan; the UI surfaces the limit with a "paste instead" fallback.
• No auth: anyone with the URL can use the app.
• No streaming: results return all at once. Acceptable for prototype latency; the model call is ~3–8 seconds depending on artifact length.
• Mock mode quality: deterministic, profile-grounded, and useful for demos — but it's pattern-driven, not insight-driven. Always set ANTHROPIC_API_KEY for real feedback.

Scripts

pnpm dev                          # dev server
pnpm build && pnpm start          # production build
pnpm exec tsc --noEmit            # type-check
pnpm exec tsx scripts/smoke-analyze.ts   # mock recommendation smoke test

Status

Phase 0–7 from PLAN.md are complete. Full walking skeleton with all spec pages, real LLM integration, and a graceful mock fallback. Ready for a demo; ready to extend.