Engineering discipline
for AI harnesses.
Eight skills that mitigate the failure modes that show up once you're mostly programming in English: silent assumptions, sycophancy, overcomplication, half-finished migrations, tautological tests, and confidence without calibration.
The default load for any project.
/deliberate teaches your agent nine principles for how to write code - when to plan, when to push back, when to stop. Everything else in this library builds on top of it.
Nine principles for how an agent should write code: think before coding, push back when warranted, plan inline, simplicity first, surgical changes, goal-driven execution, know when to stop, calibrate confidence, maintain context.
Silent assumptions, sycophancy, overcomplication, scope creep, orthogonal edits, runaway loops, confidence without calibration.
Seven specialized skills.
Load these alongside the foundation when the task calls for them. Each is a standalone file - take what you need, skip what you don't.
Think
2 skillsBefore you touch a keyboard. Understand intent, shape the solution, pick where it lives.
Build
2 skillsTurning intent into durable code. Tests that earn their keep. Migrations that revert.
Investigate
2 skillsFinding what's wrong and why. Debugging with named causes, reviews that catch omissions.
Respond
1 skillWhen production is on fire. Stabilize, preserve evidence, write the honest postmortem.
Drop it in. Reference it. Ship.
Every skill is plain markdown. No tool-specific syntax, no build step, no lock-in.
npx skills add angad729/deliberateInstalls all eight skills. Works with Cursor, Claude Code, Codex, Gemini, and 40+ other agents via vercel-labs/skills.
Prefer to control what lands in your repo? Clone the repo and reference skills from CLAUDE.md, .cursorrules, or paste into any system prompt. Plain markdown.
The models will keep writing code. How they write it is still up to us.
In late 2025, Andrej Karpathy posted a thread on what changed - and what didn't - as agent coding crossed a coherence threshold. The diagnosis was sharp:
"The models make wrong assumptions on your behalf and just run along with them without checking. They don't manage their confusion, don't seek clarifications, don't surface inconsistencies, don't present tradeoffs, don't push back when they should."
Those are the failure modes Deliberate pushes against. Credit for the diagnosis is Karpathy's. This library is the practical extrapolation into files you can drop into a project.