Ai engineering workflow
Skill Jared13PG/ai-engineering-workflow/.agents/skills/ai-engineering-workflow
Description: A structured engineering workflow skill for complex AI-assisted coding, coursework, data analysis, and research projects.
npx -y skills add Jared13PG/ai-engineering-workflow --skill ai-engineering-workflowAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 17 days oldThe repository was created 17 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Use this skill for complex coding, coursework, data analysis, notebook, research, or multi-file projects that need structured issue creation, context analysis, task breakdown, atomic execution, verification, status tracking, and final closure.
SKILL.md
11.1 KB, ~2.3k tokens by cl100k_base, as published. Nobody here has run it
ai-engineering-workflow
Purpose
Use this Skill to manage complex project work through a controlled engineering workflow instead of jumping directly into implementation.
issue-create → issue-breakdown → issue-execute → issue-update → issue-status → issue-close
The workflow helps Codex understand context, control scope, execute bounded tasks, verify honestly, synchronize documentation, and produce a clear final handoff.
When To Use
Use this Skill when the user invokes:
$ai-engineering-workflow
It is appropriate for:
- coursework and research projects;
- multi-file coding or AI engineering work;
- data analysis and notebook projects;
- projects with starter code, datasets, PDFs, templates, or grading rules;
- requests to create an issue, plan tasks, implement one task, revise the plan, report progress, or close a project.
Do not force the full workflow onto a simple one-step task unless the user explicitly invokes the Skill.
Core Rules
1. Understand Before Implementing
Before editing files, identify:
- user goal and project background;
- available files and current repository state;
- expected deliverables;
- constraints and acceptance criteria;
- missing or unclear information.
Load context in layers when available: project instructions, repository guidance, current issue documents, architecture documents, relevant source files, and tests.
Do not silently invent missing requirements. Record uncertainty and its impact.
2. Keep Scope Controlled
Only perform work that is requested, required by the active issue, required by the selected task, or necessary for verification.
Suggest optional improvements separately. Do not implement them automatically.
3. Execute Atomic Tasks
Every implementation task should define:
- task ID and objective;
- scope and related files;
- dependencies;
- acceptance criteria;
- verification method;
- status.
During issue-execute, complete only one selected task or one clearly bounded task group.
4. Verify Honestly
Use available checks such as tests, scripts, notebook execution, output inspection, Markdown checks, file-structure checks, and requirement matching.
Never claim a check passed when it was not run. When verification cannot be performed, record:
Verification not run
and explain why.
5. Keep Documents Synchronized
After meaningful work, update the relevant workflow documents:
issue.md
status.md
docs/system-analysis.md
docs/architecture.md
docs/task-breakdown.md
close-report.md
Documentation must reflect the actual project state.
6. Protect User Work
When working in a Git repository:
- inspect the current branch and working tree before editing;
- avoid overwriting user changes;
- do not commit secrets, credentials, private datasets, or private assignment answers;
- do not publish before local testing and privacy checks.
Mode Selection
Supported conceptual modes:
issue-create
issue-breakdown
issue-execute
issue-update
issue-status
issue-close
These are Skill modes, not operating-system commands. Natural-language requests may invoke them.
When the user does not name a mode, choose the safest next step from the current project state:
- new complex project →
issue-create; - context exists but no executable plan →
issue-breakdown; - a bounded task is ready →
issue-execute; - requirements, files, scope, or blockers changed →
issue-update; - progress summary requested →
issue-status; - work is ready for handoff →
issue-close.
Task Status Vocabulary
Use these values consistently for individual tasks:
todo
in_progress
blocked
done
done_unverified
skipped
needs_review
todo: not started;in_progress: currently being worked on;blocked: cannot proceed because required information, files, or dependencies are missing;done: completed and verified;done_unverified: appears complete, but verification could not be fully run;skipped: intentionally omitted, with the reason recorded;needs_review: requires human confirmation.
Use these values only for the overall issue at closure:
complete
partially_complete
blocked
canceled
Do not mix task status and issue closure status.
Mode Behavior
issue-create
Goal: establish project context before implementation.
Required actions:
- Read the request and inspect available files.
- Identify goal, scope, deliverables, constraints, risks, and acceptance criteria.
- Create the issue memory directory and initial workflow documents.
- Record missing information and assumptions.
- Do not implement unless the user explicitly requests combined planning and implementation.
Expected files:
memories/YYYYMM/CHANGEID/issue.md
memories/YYYYMM/CHANGEID/status.md
memories/YYYYMM/CHANGEID/docs/system-analysis.md
memories/YYYYMM/CHANGEID/docs/architecture.md
memories/YYYYMM/CHANGEID/docs/task-breakdown.md
issue-breakdown
Goal: convert established context into atomic executable tasks.
Required actions:
- Read the issue, system analysis, architecture, status, and relevant project instructions.
- Break the work into phases and atomic tasks.
- Add dependencies, related files, acceptance criteria, and verification methods.
- Update
docs/task-breakdown.mdandstatus.md. - Do not implement code.
issue-execute
Goal: complete one selected task without scope expansion.
Required actions:
- Identify the selected task and confirm its dependencies.
- Inspect relevant files before editing.
- Modify only the files necessary for that task.
- Run available verification.
- Update task status and
status.md. - Report changed files, verification results, blockers, and the next step.
issue-update
Goal: synchronize the workflow after requirements, files, blockers, assumptions, or scope change.
Required actions:
- Identify and classify the change.
- Update all affected issue, analysis, architecture, task, risk, and status documents.
- Revise dependencies or acceptance criteria when needed.
- Mark invalidated work or verification honestly.
- Do not continue implementation using outdated assumptions.
- Do not implement code unless explicitly requested.
issue-status
Goal: report the current state without implementation.
Include:
done
in_progress
todo
blocked
done_unverified
needs_review
Verification Notes
Next Step
Do not modify project code during a status-only request.
issue-close
Goal: create an honest final handoff.
Required actions:
- Check task completion and deliverables.
- Review changed and created files.
- Review verification results.
- Record unresolved risks, limitations, or manual checks.
- Create or update
close-report.md. - Mark the issue
complete,partially_complete,blocked, orcanceled.
Do not claim full completion unless all required tasks are complete and verified.
Project Memory Structure
Create this structure during issue-create:
memories/
└── YYYYMM/
└── CHANGEID/
├── issue.md
├── status.md
└── docs/
├── system-analysis.md
├── architecture.md
└── task-breakdown.md
Add close-report.md during issue-close.
Use English filenames by default unless the user explicitly requests otherwise.
Templates
Load reusable templates relative to this Skill directory from:
assets/templates/
Mapping:
assets/templates/issue-template.md → issue.md
assets/templates/system-analysis-template.md → docs/system-analysis.md
assets/templates/architecture-template.md → docs/architecture.md
assets/templates/task-breakdown-template.md → docs/task-breakdown.md
assets/templates/status-template.md → status.md
assets/templates/close-report-template.md → close-report.md
If a template is unavailable, create the document manually using the same structure and note the missing template.
Reference Loading Rules
Load detailed guidance relative to this Skill directory only when relevant:
- for overall workflow selection or an unclear project state, read
references/workflow-overview.md; - for
issue-create, readreferences/issue-create-guide.md; - for
issue-breakdown, readreferences/issue-breakdown-guide.md; - for
issue-execute, readreferences/issue-execute-guide.md; - for
issue-updateorissue-status, readreferences/issue-update-status-guide.md; - for
issue-close, readreferences/issue-close-guide.md; - before final completion, release, or when verification is important, read
references/quality-control.md.
Do not load every reference automatically when one focused guide is sufficient.
Quality Gates
Before marking work complete, check:
- Requirement Gate: goals, deliverables, constraints, and uncertainties are clear.
- Scope Gate: no unrelated work was added.
- Implementation Gate: only necessary files were changed.
- Verification Gate: checks were run or skipped with an explanation.
- Documentation Gate: issue, task, and status documents match reality.
- Git Gate: no secrets, private data, or unrelated changes are included.
- Handoff Gate: the user knows what changed, what remains, and what to do next.
Forbidden Actions
Do not:
- jump directly into implementation for a complex task without context;
- invent requirements, files, datasets, test results, or command output;
- implement unrelated features or broad refactors;
- overwrite user changes silently;
- mark work verified when checks were not run;
- publish secrets or private data;
- use inconsistent workflow filenames or status labels.
Progress Response Format
Use a compact structure such as:
Done
- ...
Files Created or Updated
- ...
Verification
- ...
Current Status
- done: ...
- todo: ...
- blocked: ...
- done_unverified: ...
Next Step
- ...
When blocked or unverified, state the reason explicitly.
Example Invocations
Use $ai-engineering-workflow in issue-create mode. Analyze the project and create the initial workflow documents, but do not implement yet.
Continue in issue-breakdown mode and create atomic tasks with dependencies, acceptance criteria, and verification steps.
Continue in issue-execute mode and complete Task 1.1 only.
Use issue-update mode. The input file and requirement changed; revise the workflow documents before further implementation.
Use issue-status mode to summarize completed, pending, blocked, and unverified work.
Use issue-close mode to create the final handoff report.
Final Rule
Always prefer:
understand → document → break down → execute one task → verify → update → close
over:
guess → code everything → skip checks → claim completion
What ships with it: 13 files
53.5 KB alongside SKILL.md
assets/
references/
- issue-breakdown-guide.md4.8 KB
- issue-close-guide.md4.5 KB
- issue-create-guide.md5.6 KB
- issue-execute-guide.md6.0 KB
- issue-update-status-guide.md5.8 KB
- quality-control.md6.8 KB
- workflow-overview.md5.9 KB