VibeCoding Start
Use this Skill to give Codex a lightweight engineering system from the first project session through release and observation. The desired result is a project that a later session can understand, review, verify, accept, and safely continue without replaying the entire chat.
First action
Read the project's AGENTS.md, then follow this Hot Context order:
docs/INDEX.mddocs/CURRENT.mddocs/PRODUCT.mddocs/PRD.mddocs/ACCEPTANCE.md
Read docs/CODEMAP.md, active ADRs, plans, worklogs, standards, and local raw evidence only when the task needs them. If the six-document skeleton is missing, use the companion vibecoding-project-knowledge Skill's INIT mode before implementation.
Non-negotiable project boundary
- Every project receives the same minimum skeleton:
AGENTS.md,docs/INDEX.md,PRODUCT.md,PRD.md,ACCEPTANCE.md,CURRENT.md, andCODEMAP.md. - Complexity changes document depth and gate strength; it does not remove the skeleton.
PRODUCT.mdexplains why,PRD.mdexplains what, andACCEPTANCE.mdexplains how to prove it. Do not silently merge these responsibilities.- Update
PRD.mdwhen the session changes a requirement, feature, non-goal, constraint, or meaningful user scenario. - Keep raw sessions, logs, failed attempts, investigations, and evidence local by default under
.project-memory/. Promote only redacted, intentionally curated examples todocs/examples/. - Prefer existing project capabilities, the standard library, official SDKs/APIs/tools, mature open-source projects, packages, adapters, or composition before custom infrastructure.
- Do not build a memory database, vector database, search engine, testing framework, deployment framework, CI service, or observability backend as part of this workflow.
Engineering gates
Use the same vocabulary for every project while scaling the depth to risk:
G0 Scope
→ G1 Intent
→ G2 Reuse
→ G3 Plan
→ G4 Build
→ G5 Adversarial Review
→ G6 Verification
→ G7 Human Acceptance
→ G8 Release
→ G9 Observation
G0 Scope
Classify the work as application/product, integration/adapter, or infrastructure/security/runtime. The closer it is to infrastructure, the stronger the reuse, review, and recovery evidence must be.
G1 Intent
Make the chain explicit:
PRODUCT (Why) → PRD (What) → ACCEPTANCE (How to prove)
Record the problem, users, current and desired states, functional requirements, non-goals, constraints, acceptance, and risks before implementation.
G2 Reuse
For any non-trivial general capability, search before building and decide in this order:
USE → ADAPT → COMPOSE → BUILD → STOP
If BUILD is selected, record searched alternatives, rejected reuse paths, maintenance responsibility, and the owner's acceptance method in an ADR. A model's ability to write code is not evidence that custom code is the right choice.
G3 Plan
Break complex work into small milestones with a verification point after each milestone. Stop and re-plan when scope, architecture, or risk changes materially.
G4 Build
Keep implementation small and reversible. Use the existing toolchain and avoid speculative abstractions, dependencies, and infrastructure.
G5 Adversarial Review
An independent reviewer examines PRODUCT.md, PRD.md, ACCEPTANCE.md, active decisions, relevant structure, the diff, and tests. Look for omissions, wrong assumptions, regressions, duplicated capability, over-design, security/data risks, test gaps, and rollback failures. The author must not be the final reviewer for an important change.
G6 Verification
Cover normal, error, boundary, invalid-input, regression, and recovery paths in proportion to risk. Report machine evidence and an owner-readable explanation; a green check proves only the invariant that it actually checks.
G7 Human Acceptance
The owner confirms that the result is the intended product and that its observable behavior is understandable. Automated checks and independent review do not replace this step.
G8 Release
Record the version, last-known-good state, release notes, backup, rollback method, migration risk, and exact release target before real use.
G9 Observation
After release, observe health and user evidence. Continue only when healthy; stop and roll back or roll forward when evidence shows a problem. Feed the result back into tests, worklog, ADR, product intent, acceptance, or a postmortem as appropriate.
Knowledge closeout
At the end of every substantive session, use the companion vibecoding-project-knowledge Skill's UPDATE mode:
- Preserve raw session details and large output in the local
.project-memory/directory. - Curate
CURRENT.mdand the relevant active documents. - Update
CODEMAP.mdwhen structure changes. - Update
PRD.mdwhen requirements, features, non-goals, constraints, or user scenarios change. - Record material choices in an ADR.
- Update indexes and the worklog.
- Run the relevant deterministic audits.
Do not claim completion without machine evidence, a human-readable explanation, and an owner acceptance method.
References
- Read Standard v1.3 for the normative rules and release boundary.
- Read the companion
vibecoding-project-knowledgeSkill forINIT,UPDATE, orAUDITprocedures.