Building
I’m building an agent
“I have an idea but no clear bar for good vs bad output.”
Output: .lignos/canvas.md + constitution — no Studio install required.
Then govern · steal an example →
Before you ship another agent on vibes
Build agents that do what you actually want — faster.
Most agents drift from your intent the moment they hit real inputs. Answer four questions — get a system prompt your agent actually follows, a judge prompt to verify it, and your first eval run. All in one session, no setup.
The reveal: define what “good” means once, then carry it from V0 → eval → production.
Where are you right now?
Pick the symptom — not a product tour. Each door ships a concrete artifact in <30 minutes.
Evals
I’m setting up evals
“Braintrust is empty — who writes the judge?”
Output: judge prompt + scenario seeds from your anti-pattern field.
Export demo →Also: LangSmith
Production
Something’s off in prod
“Evals are green but the product still feels wrong.”
Output: drift alerts that cite your words — not generic monitoring.
Lignos Studio — coming soon · See what's coming →
The continuity moment: the anti-pattern you wrote in Cursor becomes your blocking eval case, your prod drift alert, and your system prompt guardrail. You didn’t configure three tools — you defined judgment once. This is your eval V0 — a foundation from requirements. The seeds are synthetic starters; real outputs make it sharper.
Labs skill chain
/lignos-canvas
/lignos-eval
/lignos-govern
ship
/lignos-scope (optional)
Full skills overview · Claude Code commands in
.claude/commands/