Menu
Starts working on a task using the TRIPLE protocol, as the impl agent.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
| name | triple-impl |
| description | Starts working on a task using the TRIPLE protocol, as the impl agent. |
| disable-model-invocation | true |
| argument-hint | Extra feedback or instructions for the impl agent. Optional. |
Review the project local TASK.md, and follow the triple.md protocol to start work.
You are the impl agent. Surface any concerns per the protocol.
$ARGUMENTS
Create and maintain GitHub issues from user reports and screenshots, fast and without over-investigating. CREATE: use when the user says 'log a GitHub issue', 'file a bug', 'raise an issue', 'open an issue', 'log this on GitHub', or pastes/attaches a screenshot of something broken and asks to report it — captures the symptom, works out steps to reproduce where possible, uploads any screenshot/video and embeds it, applies the GitHub issue authoring standard, and files immediately. UPDATE: use when the user says 'update the issue', 'add a comment to #N', 'attach this to the issue', 'edit the issue body', 'relabel', 'close/reopen the issue' — adds comments with media, edits the body, upgrades the repro provenance badge, relabels, and closes/reopens. It does NOT root-cause the code or prescribe a fix — that's for triage.
Stage or unstage specific hunks from a changed file in a git repo without interactive prompts. Use this skill whenever you need to partially stage or unstage hunks — for example, committing only the bug fix hunk but not the refactor, or unstaging something that was staged by mistake. Common phrases: 'commit only the X changes', 'stage that hunk', 'don't include the Y part', 'unstage the refactor'. Pass args as: first line = file path, remaining lines = plain-language description of which changes to (un)stage. Example args: "src/foo.ts\nstage only the import changes, leave the function rename".
Scan git diffs and stage only the changes that are self-evidently correct and intentional in context — changes a reviewer would have nothing to deliberate on. Covers lint fixes, formatting, typos, dead code, observability, uniform sweeps, AND self-evidently complete additions like new dev-harness pages wired up via import + menu entry + new file, testid sweeps, mount-marker additions, AGENTS.md pointers. Leaves anything requiring reviewer judgment unstaged: API contract changes, secret rotations, schema requirement changes, algorithm changes, config values the reviewer must weigh, code smells. Use this skill when the user says 'stage the obvious stuff', 'stage trivial changes', 'stage the safe diffs', 'cherry-pick the lint fixes', 'stage what's obviously correct', or wants to triage a messy working tree before committing.
Run the full unit/integration test suite (typically vitest) and end-to-end tests (typically Playwright), then auto-fix failures iteratively using subagents. Use this skill when the user says 'run all tests', 'run the test suite', 'run e2e', 'fix all tests', 'get tests green', 'run vitest', 'run playwright', or any variation of running + fixing tests across a monorepo. Supports --env dev (default) and --env staging for the e2e target. Also trigger when the user asks to 'check everything works' or 'make sure nothing is broken'.
Finalises a reviewed task — creates PR and merges using the TRIPLE protocol, as the lead agent. Trigger when the user says "finalise the task", "create the PR and merge", or invokes /triple-finalise after the review agent has approved.
Run the project's code quality pipeline: oxfmt (format), oxlint --fix (lint auto-fix), and typecheck — then fix the trivially correct issues that remain. Use this skill whenever the user says /tidy, asks to clean up lint/format/type errors, mentions oxfmt/oxlint/typecheck, or wants to tidy code before committing. Also use when the user pastes compiler or lint errors and wants them cleaned up mechanically. If the user asks for a report only (e.g. '/tidy --report', 'just check', 'what's broken'), run the pipeline but report issues without fixing.
