Pick a playbook for the demo
Demo 1 of 6
Loading…
~/work — claude-skill
0 / 0
readiness
—
/ 100
findings
0
Auto-loops after each run. The Q&A is pre-recorded — each question shows the
choice CogniDev would pre-select for this codebase, and the reasoning it would log.
What you're watching
A CogniDev playbook is a sandboxed binary that runs the same seven-phase contract whether it's invoked from a workbench, a terminal, or — as you see here — a Claude Skill. The host parses cg-playbooks protocol events over stdio and renders them into the chrome. The playbook itself decides what questions to ask, what artifacts to write, and which transforms are deterministic (codemod) vs. semantic (LLM). Every diff is tagged with which intelligence made the call.
pre-check · analyze
Sense the stack. Build the Merkle index, AST, call graph, hot pockets. No LLM yet — the playbook stands on its own with deterministic parsers.
document · plan
The heaviest Q&A phase. The playbook asks the deep architectural questions — target LTS, build tool, auth, persistence — and the answers shape both the LLM prompt and which artifacts get written under
.cw/.transform · validate · sign-off
Apply changes — deterministic codemods where safe, LLM for residuals. Validate by compile/lint/test. Final read-through against the architectural decisions logged in the document phase.