gjames-development-workflow

name: gjames-development-workflow description: "Use for disciplined AI-assisted software development workflows: planning, coding, debugging, code review, testing, deployment, loops, tools, model routing, and automation across Claude and Codex. Applies when implementing features, fixing bugs, reviewing PRs, triaging CI/logs, setting up agent loops, or deciding how to route work between helper, coding, planning, and escalation models."

G.James Development Workflow

Use this skill to run software work through a repeatable engineering loop instead of one-shot prompting.

Default Lifecycle

Use this outer lifecycle for non-trivial work:

Discover -> Plan -> Discuss -> Build -> Review -> Test -> Deploy -> Observe -> Learn

Use this inner agent loop inside each phase:

Inspect -> Decide -> Act -> Verify -> Record

Phase Rules

1. Discover before asking when repo facts can answer the question.
2. Plan before mutation unless the task is clearly tiny.
3. Discuss only decisions that materially affect scope, behavior, risk, cost, model/tool choice, or tradeoffs.
4. Build in small reversible changes; prefer existing patterns.
5. Review before declaring implementation complete.
6. Test with the narrowest useful check first, then broaden based on risk.
7. Deploy only from verified source; prefer reversible rollout paths.
8. Observe after release or handoff.
9. Learn from repeated failures — convert mistakes into rules, references, scripts, hooks, or tests.

See references/loops-and-tools.md for verification guidance and tool order.

G.James-Specific Content

This skill contains generic workflow guidance. Over time, add G.James-specific content here and to the reference docs:

• repo conventions, test/build commands, and CI system names
• deployment targets, staging/canary procedures, and rollback policies
• org patterns for schema changes, migrations, permission audits, and security reviews
• recurring review findings and their mechanical checks

Capture new patterns via the Learn phase — convert repeated mistakes into skill rules, automated checks, or scripts.

Tool Policy

• Use search/read tools first: rg, file reads, configs, manifests, docs, schemas, logs, diffs.
• Use planning and discussion tools when they reduce risk or lock material decisions.
• Use mutation tools only after the path is clear: patches, generators, migrations, formatters.
• Use verification tools after meaningful changes: tests, builds, lint, smoke tests, screenshots, data checks.
• Use external tools only when needed: GitHub, browser, deployment, spreadsheets/docs, MCP/connectors.
• Gate dangerous tools: deletes, resets, production writes, broad migrations, deploys, and permission changes.

Reference Loading

Read only the reference needed for the current task:

• references/model-routing.md: choose Claude/Codex model lanes, reasoning effort, and escalation rules.
• references/loops-and-tools.md: design goal loops, watch loops, triage loops, learning loops, and tool gates.
• references/automation-lanes.md: design durable intake, implementation, review, fixer, deployment-watch, and learning automations.

Completion Standard

Finish with:

1. What changed or what was decided.
2. What verification ran.
3. What could not be verified, if anything.
4. Any residual risk or follow-up that materially affects the user.