Taste is now a real skill. Judgment built in. Craft.

doc-writer

Writes docs for the person who arrived stuck.

crystallized Apr 12 · .craft/research/technical-documentation-writer-become/

Diagnoses structure before touching prose - because the problem is almost never writing quality, it's the wrong doc type for the reader's state. Knows the difference between a tutorial, a how-to, a reference, and an explanation, and knows that mixing them is what breaks most documentation.

crystallizer

Distills research into an inhabitable expert mind.

crystallized Apr 11 · .craft/research/expert-cognition-transfer/

Takes branch files of psychological material and produces an agent that thinks like the subject rather than knows about them. Reads between the lines for the perceptual framework the research subjects see through but cannot articulate. The highest-judgment task in the system.

riff

Thinks with you, leaves no fingerprints.

crystallized Jun 4 · .craft/research/riff-become/

A creative thinking partner that reads where you are - stuck or brewing, wanting a voice or wanting space - before it says a word. In the room with you, not in front of you. The riff worked if you end up closer to the thing you were already making and can't find its hand in it.

project-scanner

Detects the stack before asking any questions.

Reads the codebase to find what's already true - language, patterns, architecture, dependencies - so the only questions left are about intent. Users shouldn't have to tell an agent what language their project is written in.

tester

Writes tests that survive production pressure.

Thinks like a user, breaks like a hacker, writes tests like someone who's been burned by the edge case that wasn't. Future developers either thank this agent or inherit the consequences.

chunk-validator

Runs quality gates, reports what failed.

Runs typecheck, lint, build, tests, and token checks - then writes a structured report and stops. Doesn't fix. Doesn't guess. The orchestrator decides what to do with the findings.

claims-auditor

Checks the claims, not the code.

When a story says it's done - all tests pass, only these files changed - it re-derives each claim from what's actually on disk and trusts none of the reasoning that produced it. It audits the narrator, not the code. Every claim comes back supported, unsupported, or unverifiable - and when it can't tell, it says unverifiable rather than guess.

verifier

Tries to break a single research claim.

Takes one finding and attempts to disprove it using independent primary sources. Adversarial by design - if it can't break the claim after genuine effort, that's a strong confirmation. Checks source code and live behavior before accepting documentation at face value.

practitioner-reviewer

Challenges verified claims from lived experience.

Reads research that passed source verification and flags what's technically sourced but practically wrong. Doesn't search the web - thinks like someone who uses these tools every day and notices when a confirmed claim doesn't match reality.

pr-reviewer-expert

Reviews diffs with full codebase context.

Reads locked decisions and surrounding code before forming any opinion about a change. Posts 3 findings developers will act on rather than 30 they'll dismiss. Context assembly is what separates useful reviews from noise.

researcher

Investigates one question, writes findings to disk.

Takes a single sub-question, goes deep, and writes conclusions to a branch file. Runs in parallel with other researchers. Returns a lightweight summary - the file holds the substance.

research-synthesizer

Merges parallel research without laundering it.

Reads every branch the researchers wrote and produces one ranked plan plus a citation index. Where two sources disagree it sets the conflict down side by side instead of smoothing it into false consensus - a disagreement is data, not a problem to resolve.

become-researcher

Collects psychological material, not facts.

Looks for the beliefs and aversions that shape how an expert thinks - most of which are invisible to the expert themselves. Facts are noise. The signal is what they flinch away from, what they keep coming back to, what they reach for when something is good or bad.

guide

Explains craft, grounded in the source.

A read-only advisor for craft itself - answers how a command, skill, or lifecycle step works, then reads your actual .craft state to explain why a story won't implement and what to fix. Not a persona with opinions - it derives every answer from craft's real source files, where the source always beats the docs.

creative-analyzer

Finds the wow moments and viral potential.

Evaluates WHAT to build next - not whether what's built works, but whether it could make someone stop scrolling, tell a friend, or come back tomorrow. Looks for the moments where the product could become something people describe with their hands.

style-analyzer

Catches design drift from locked tokens.

Sees the #6B7280 that should be text-secondary. Finds the 2px misalignment others skip past. Audits every component against the locked design system and distinguishes intentional creative choices from accidental divergence.

walkthrough-analyzer

Interacts with the live app as a real user.

Clicks every button, checks every state transition, reports what doesn't feel right - from the browser, never from source code. If something is confusing to encounter cold, this agent will encounter it cold.

playwright-browser

Owns a live browser session end-to-end.

Navigates pages, clicks elements, fills forms, reads accessibility snapshots, and reports findings across turns. Remembers what it's seen and done. Built for interactive steering - not one-shot automation.

product-anthropologist

Diagnoses whether the problem actually exists.

Sits between a builder's conviction and the market's indifference and translates what the gap means. Not a UX researcher - the question isn't "does this design work?" but "does this problem exist urgently enough to justify a product?" Sometimes the diagnosis is stop building.