SR Method EN
How to talk to Codex
Development, framing, strict resume, Fact Gate and domain skill prompts.
How to talk to Codex
A good prompt for a code AI agent should specify the verifiable objective, constraints, sources to read, expected checks and SR memory to update.
Good development prompt
Project: /path/to/repo.
Verifiable objective: fix [precise problem].
Constraints: do not change dependencies, do not create migrations, do not touch secrets.
Before coding: read CURRENT_STATE, the relevant lot, RepoMap and real candidate files.
After correction: run relevant tests, update SR memory and give user E2E tests.Good framing prompt
I want to frame a new feature: [description].
Do not code.
Turn this request into a proposed SR lot with objective, scope, out of scope, acceptance criteria, sources to read, risks, checks and user E2E tests.Good strict resume prompt
Strict SR resume.
Project: /path/to/repo.
Read docs/codex/tasks/YYYY-MM-DD_slug/NEXT_SESSION_PROMPT.md and associated contracts.
Summarize state, do not code before validation.Prompt when Codex assumes too much
Apply Fact Gate: verify in code or SR sources before concluding. If you do not have proof, mark the element as an unverified assumption.Domain skill generation prompt
Turn the plan below into Codex skills ready to be placed in project-skills/.
Respect this format: frontmatter name/description, objective, triggers, domain vocabulary, non-negotiable rules, sources of truth, expected patterns, anti-patterns, closure checklist.
The content must be operational for a Codex agent, not written as a pedagogical article.
Skills plan:
[paste 13_DOMAIN_SKILLS_PLAN.yaml]With practice, requests can become simpler. The method gives Codex enough context to understand shorter intents.