# Implement Issue > Validates and implements a GitHub issue against current repository evidence, repeats code review and black-box testing, runs the project's completion gate, and opens a draft pull request. Use when the user runs /implement-issue with an issue number. - Skill: `frostney/implement-issue` (Agent Skill) - Install (CLI): `npx skillmds@latest add frostney/implement-issue` - Raw SKILL.md: https://api.skillmd.com/api/skills/frostney/implement-issue/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - License: Unlicense OR MIT - Author: frostney (https://skillmd.com/u/frostney) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/frostney/implement-issue --- # Implement issue Resolve the issue end to end against the current repository, or establish with evidence that no implementation is needed. ## Gates - Read the issue, project instructions, vision, contribution guidance, Definition of Ready, Definition of Done, relevant domain skills, real project commands, affected code paths, tests, and reproduction before deciding. - Always perform and record web search for current evidence before presenting options. Prefer official and primary sources, reconcile them with the versions in the checkout, and treat remembered links only as search leads. Stop if the search cannot produce current evidence relevant to the decision. - Outside automatic mode, require the registered `grilling` and `render-html` skills plus Python 3. Stop if any dependency is unavailable. Use the actual `grilling` decision loop; do not imitate it with ad-hoc questions. - Before proposing an option, assemble one neutral evidence packet from the issue, repository, reproduction, project contracts, current web research, and shared current-state artifacts. Derive every option from this same packet. - Define one comparison rubric from the issue outcome, constraints, and requirements before scoring. Give every viable option equivalent decision-relevant validation: - for UI/UX differences, show each materially different experience; - for architecture or workflow differences, show each relevant flow; - for interaction-heavy or technical claims, use comparable short-lived prototypes or measurements when static evidence cannot decide them. Equivalent checks do not require equal implementation effort. Label observed facts, proposed behavior, and prototype-only shortcuts. - If an option reveals an evidence gap, run the same relevant check for every affected option. When that is impossible, disclose the unequal evidence and reduce confidence before comparison. - Keep prototypes local and disposable, retain only reviewable captures and findings, and remove them when the grill concludes. Do not deploy or publish them. Preserve or promote a prototype only with explicit user approval; keep approved prototype material outside the selected worktree until its `git-workflow` synchronization gate passes. - Compare two to four genuinely distinct viable options against the declared rubric, then recommend one and wait for the user's choice unless automatic mode applies. Include a compact evidence digest with the most relevant source links, checked project versions, scores, and remaining uncertainty. - Outside automatic mode, use `render-html` to present that comparison as a temporary interactive impact report before the `grilling` decision loop. Include shared structured evidence, the declared rubric, option impacts, pros, cons, scores, uncertainties, and copyable discussion prompts. Keep the recommendation in the closing section after every neutral option. - When current evidence conclusively fails a required prototype or readiness threshold, report that stop without asking whether to bypass the gate. - After option selection or automatic entry, keep the authorized implementation active across status questions, diagnoses, corrections, and failed gates. Return control only when the outcome is complete, safe in-scope progress is exhausted by an external blocker, or continuing requires new authority or a material unresolved decision. A known fix or recommendation is not terminal. - For any code or test change, repeat `/code-review fix-all` and black-box testing against the issue until both pass on the same unchanged implementation, then complete the project gate and `/create-pr`. ## Project definitions Treat the nearest applicable `DEFINITION_OF_READY.md` and `DEFINITION_OF_DONE.md` as canonical. If either is absent after a real search, state that once, carry the gap into the plan and PR, and use only the workflow's built-in checks plus commands the repository actually declares. ## Automatic mode Automatic mode applies only when the original prompt says `automatic` or explicitly requests it. It skips the interactive HTML report and `grilling` loop, but does not waive any other gate. Do not select an option until the shared packet, predeclared rubric, and equivalent checks are complete. If a comparison remains incomplete, report the gap and confidence impact rather than treating the initial preference as validated. A material product, architecture, security, scope, or vision decision disables automatic mode. ## Workflow 1. Fetch the issue, including comments and labels; verify it is open, implementation-ready, and not a PR, duplicate, blocked, or rejected item. 2. Load the applicable project contracts and specialized skills. 3. Trace the full behavior from entry point to symptom and run the cited reproduction or artifact when possible. Search callers, siblings, tests, configuration, and history far enough to identify the real layer. Always perform current web search and reconcile its results with the checkout. 4. If the behavior no longer reproduces, distinguish: - already fixed and covered: recommend closing with the fixing code/commit and test evidence; - fixed without regression coverage: add the missing test only; - shared-root-cause sibling paths still affected: include only those paths. 5. Build the neutral evidence packet and comparison rubric, derive the options, and validate each with equivalent decision-relevant checks. Compare first; recommend only afterward. 6. Outside automatic mode, create a temporary directory outside the worktree. Follow the `render-html` schema and render the option comparison there. Open the report for review. If inline host preview is unavailable, provide its absolute path and continue. Run `grilling` one decision at a time using the report and shared evidence, then wait for the user's choice in chat. Remove the temporary report and its inputs after selection unless the user explicitly asks to keep them. 7. After selection, reuse or create a focused branch/worktree and apply the `git-workflow` remote-default synchronization gate before editing. 8. Implement the smallest complete change at the correct layer. Update tests and docs required by the issue and project contracts. 9. For UI/UX work, render every affected state; capture reviewable before/after evidence; check accessibility, responsive behavior, themes, and design-system consistency; attach the evidence to the PR. 10. Run targeted checks while developing. Fix failures rather than weakening a check. 11. Run `/code-review fix-all` against the requirements, Definition of Done, project conventions, branch diff, and reproducible behavior. Resolve every validated in-scope finding. Stop for a material new decision or unresolved Blocking or Important finding. If `/code-review` is unavailable, perform the same bounded review and fix pass directly. 12. Run `/test-against-spec fix` when it is available. Otherwise perform the same black-box test directly: use explicit issue requirements, avoid source as proof, prefer an exact-revision preview deployment when available, and fall back to the local environment. Record each requirement and its observed result or limitation. 13. If step 11 or 12 changes the implementation or reports incomplete work, continue implementing and restart at step 11. Repeat until code review and behavior testing both pass on the same unchanged implementation. Stop and ask when required behavior remains unverified after exhausting the available environments, unless the user explicitly accepts the limitation or requests a draft PR to obtain the missing preview environment. 14. Run the applicable Definition of Done and repository gate. If fixing a gate failure changes the implementation, return to step 11. Do not duplicate a broad gate inside the behavior-testing step. 15. Use `/create-pr`, include `Closes #`, pass forward the current completion evidence, and report only observed results. 16. Before any earlier return, apply the active-implementation terminal check. Continue with the next safe authorized action when no terminal condition is established.