Claude-powered · multi-agent · auditable
Six specialist agents chained into four workflows. Research becomes a cited evidence ledger; a red team tries to kill the idea; the spec traces every requirement back to evidence; and a quality gate with blocking authority checks whether anyone made anything up.
The chain
Every agent has one job, one output contract, and one thing it is not allowed to do. The prompts live in /agents as reviewable markdown — a prompt change shows up as a diff in a pull request, not as a mystery regression.
Research Synthesizer
claude-sonnet-5 · temp 0.2 · → evidence_brief
Turns a fuzzy problem statement into a cited Evidence Ledger — claims, sources, confidence, and explicit contradictions.
Assumption Challenger
claude-opus-5 · temp 0.7 · → challenge_report
Adversarial review. Extracts unstated assumptions, ranks them by blast radius, and designs the cheapest test that could falsify each one.
Spec Writer
claude-opus-5 · temp 0.3 · → prd
Writes a build-ready PRD where every requirement traces to evidence and every metric has a baseline, a target, and an instrument.
Prototyper
claude-opus-5 · temp 0.6 · → prototype
Turns the PRD's primary flow into a self-contained clickable HTML prototype plus the acceptance criteria that make it testable.
Output Auditor
claude-opus-5 · temp 0.1 · → audit_report
The quality gate. Scores artifacts against a published rubric, verifies every citation against the ledger, and can block the run.
Decision Archivist
claude-sonnet-5 · temp 0.2 · → adr
Writes the durable record — what was decided, what was known at the time, what was rejected, and the date it gets revisited.
The full chain. Takes an ambiguous problem statement and returns a build-ready PRD that has survived adversarial review, plus the decision record that explains it.
Recorded run: Why do 60% of our SMB accounts never add a second user?
🧭 Discovery → Spec
New problem space, low internal knowledge, decision has not been made yet.
Press Replay to watch the recorded run, or add your own inputs and run it live.
Design decisions
Chaining LLM calls is easy. Making the output trustworthy enough that someone would stake a quarter on it is the actual work. These are the four choices that did most of it.
Grounding
The Research Synthesizer writes an Evidence Ledger of atomic claims with IDs, source tiers and confidence grades. Every downstream agent cites those IDs. A claim with no ID is not a claim — it is an open assumption, and it gets logged as one.
Tradeoff — Costs output quality on open-ended questions, because agents refuse to fill gaps with plausible-sounding filler. That refusal is the feature.
Adversarial
Reviewing a finished PRD produces line edits. Running the challenge first changes what gets written: in the recorded run it moved the North Star metric off second-seat rate, because the pre-mortem predicted that exact number would rise while retention stayed flat.
Tradeoff — A stage that can return DO NOT BUILD and halt the chain. Sometimes the deliverable is the reason not to proceed.
Quality gate
Two rubrics — a 5-dimension quality score with a pass threshold on every dimension, and six binary hallucination checks. It quotes the offending sentence, names the defect, and hands it back. It never fixes the artifact itself.
Tradeoff — An agent that improves its own input hides the failure rate. Separating detection from repair keeps the error signal visible.
Loop control
On a BLOCK, only the failing stage re-runs — never the whole chain. After two failed repairs the run stops and escalates with the specific disagreement. Agents iterating on their own output converge on confident nonsense.
Tradeoff — Some runs end unresolved. An unresolved run that says why beats a resolved one that quietly lowered its standards.
Outputs
Every file below was produced by the agent whose prompt sits in /agents. They are in version control so you can read the outputs without running anything — and so the demo works with no API key.
Evidence Brief
EB-001 — Second-seat activation
10 graded claims, 2 logged conflicts, and 4 things it refused to establish.
Challenge Report
CR-001 — Red team
8 assumptions scored by blast radius; 3 ship blockers with numeric kill criteria.
PRD
PRD-001 — Second-seat activation
8 requirements, each tracing to an evidence or assumption ID. One baseline left UNKNOWN.
Prototype
PROTO-001 — First-run invite
Self-contained HTML with a state switcher: 3 of 5 states are failure states.
Audit Report
AR-001 — Quality gate
Citation-by-citation integrity pass. Caught 2 drifted claims and an invented baseline.
Decision Record
ADR-001 — Seat pricing
Reversibility class, rejected alternatives, and the dissent preserved verbatim.
Stack
Models
Claude Opus 5 for adversarial and authoring work, Claude Sonnet 5 for retrieval and archival. Routing is declared per agent in frontmatter.
Retrieval
Perplexity Sonar for live research, server-side. Degrades to disciplined-uncertainty mode when unavailable rather than failing the run.
Orchestration
A sequential async generator emitting typed events over SSE. Stages are sequential by design — the auditor must see what the writer actually wrote.
Source of truth
Agent prompts are markdown in /agents; workflows are YAML. A build script generates the TypeScript registry and fails the build on a broken reference.
App
Next.js App Router on Vercel, React 19, Tailwind v4. One streaming code path shared by live and recorded runs.
Evals
Rubrics in /evals as versioned YAML, scored against golden cases by a standalone harness — the same rubrics the in-run auditor uses.