Skip to content

gpt53-opencode-playbook.md

Latest commit

 

History

History
49 lines (38 loc) · 2.51 KB

gpt53-opencode-playbook.md

File metadata and controls

49 lines (38 loc) · 2.51 KB

GPT-5.3 + OpenCode Delivery Playbook

Use this one-page runbook to reduce rework when building with GPT-5.3 and OpenCode.

Principles

  • plan before code, verify before PR
  • follow existing repo patterns before introducing new ones
  • make decisions explicit when they affect prompts, model behavior, or workflow
  • leave a clean handoff artifact at session end

Default Flow

  1. Frame scope and definition of done.
  2. Inspect existing patterns and constraints.
  3. Implement in small, testable slices.
  4. Verify with stack-specific checks.
  5. Prepare PR with full context from branch divergence.
  6. Capture session handoff for continuity.

Trigger -> Skill -> Expected Output

Trigger Skill(s) Expected Output
new task request with ambiguity Prompt Wizard, workflow clear task statement with in-scope/out-of-scope and verification checklist
editing unfamiliar code code-search, Pattern Analysis file map, current conventions, and safe insertion points
implementing code changes language-conventions, workflow minimal-diff changes aligned to existing standards
adding or changing tests Testing Strategy test pyramid plan and targeted coverage additions
prompt/model behavior design Prompt Engineering versioned prompt contract (role, constraints, schema, examples)
preventing AI regressions AI Evaluation golden dataset, scoring rubric, and regression thresholds
architecture-impacting choice ADR Template ADR with options, rationale, consequences, and review triggers
git state unclear git-status, git-workflows staged/unstaged snapshot, branch state, and next safe git step
ready to publish work github-workflow PR with why, testing notes, blockers, and CI status summary
session ending or context switch handover concise continuity note with decisions, risks, and ordered next steps

GPT-5.3 Guardrails

  • require structured outputs for machine-consumed steps (JSON schema or strict markdown template)
  • keep prompt versions in repo and tie changes to eval results
  • treat prompt updates like code changes: diff, review, validate, and document
  • fail builds when core eval thresholds regress

PR Ready Checklist

  • scope is explicit (in/out) and unchanged during implementation
  • change matches local conventions and avoids unrelated refactors
  • required checks/tests executed (or risks documented if blocked)
  • PR explains why, not only what
  • handoff note exists when work is incomplete or multi-session