implement — build a change in an existing repo
Polymorphic on the input. Plans before it acts — the default flow stops after the plan and waits for a go-ahead before it touches a file. Writes the smallest correct change in the repo's own idiom, with tests for new behavior, validated by the repo's own tooling.
The full operating contract lives in this skill folder — read these as you need them:
| Aspect | File |
|---|---|
| How you implement (voice, discipline, edit format) | persona.md |
| The phased process + Workflow orchestration | workflow.md |
| Hard rules + refusals + safety | rules.md |
| Input routing (Jira / issue / doc / freeform) | dispatch.md |
Quick start
- Read
dispatch.mdand classify the input. Resolve the requirement:- Jira URL /
KEY-123→ fetch the issue via the Atlassian MCP (summary, description, acceptance criteria). - GitHub issue URL /
#N→gh issue view <url> --json title,body,labels,comments. (GitHub access is always theghCLI — assumegh auth loginis done.) - Confluence / Slack URL → fetch via the Atlassian / Slack MCP (or the
context-gathereragent for a mix). - freeform prose → take the text as the spec; ask the one question that unblocks it.
- Jira URL /
- Read
persona.md— adopt the smallest-correct-change implementer stance. - Run the workflow in
workflow.md. Gather context → present 2-4 approaches → confirm the plan → execute (fan out via the Workflow tool for multi-file work) → validate at every checkpoint → report. - Validate with the repo's own typecheck + lint + narrow tests after each checkpoint. A failing gate stops you (
rules.md).
Workflow is the default for real changes
"Always have a workflow." A change worth making gets the plan→act Workflow in workflow.md: gather → advise → confirm → execute → self-review → validate. For a multi-file change, fan out — spawn implementer to write the code, test-engineer to add tests, and code-reviewer / security-auditor to self-review the diff before you report. Skip the Workflow only for a one-file, single-concern edit — and say so.
Modes
- default (
--planthen confirm then--act) — gather, present approaches, write the plan, stop and confirm, then implement + validate + report. The plan→act gate is always on unless the user opts out. --plan— stop after the plan. Produce the approach menu and the file-level change list; touch nothing.--act— skip the confirm gate and implement the recommended approach directly (still validates, still gated on shared-state writes).-i— walk each fork with the user (scope, approach, test strategy) before proceeding; cap 3 questions.--deep— use a stronger reasoning profile; auto-select for cross-module, migration-shaped, or security-sensitive changes.