| name | ai-pipeline |
| description | Orchestrate the repo-aligned AI engineering pipeline (plan → decisions → eval → improve → cleanup → verify → archive) with explicit approval gates. TRIGGER when user asks to run the full pipeline or go from idea to delivery; DO NOT TRIGGER for small tasks or when user invokes a specific phase. |
AI Pipeline — Repo-Aligned Orchestration
Orchestrate the full “idea → delivery” pipeline while staying strictly compatible with this repository’s workflow gates and artifact conventions.
Hard Compatibility Rules
- Follow
CLAUDE.md lifecycle gates and agents (including Approval Gate and Archive/WAL).
- All runtime artifacts MUST be written under
.claude/runs/task-briefs/<YYYY-MM-DD>_<slug>_*.
- No external state directories, no implicit git/CI automation.
- No infinite loops. Any iterative improvement requires explicit user approval to continue.
- Only delegate to available agent types (e.g., search) and provide self-contained context.
When to Use
- User asks to “run the pipeline / full pipeline”
- Work spans multiple phases: planning + decision records + evaluation + implementation + cleanup + verification
Do not use when the user requests a single phase explicitly (e.g., “just write the plan”).
Pipeline (Repo Lifecycle Mapping)
- Intake:
input-classifier — classify and normalize the raw input before anything else. Emits [Intake] block with Intent/Profile/Scenario/AC-Candidates. Skip only for explicit @shortcuts or single-phase invocations.
- Design:
brainstorming (when creative) → impl-plan or blueprint
- Decisions:
architecture-decision-records
- Evaluation:
eval-harness (define pass/fail and verification commands)
- Approval Gate: if risk is MEDIUM/HIGH, stop after spec and wait for approval
- Implement: implement the plan in-scope only
- Improve (optional):
self-improve (bounded iterations, user-approved)
6.1 External Research (conditional): if self-improve score has plateaued for 2+ iterations, invoke external-research to find new approaches before continuing. External research can only be invoked once per plateau.
- Cleanup:
ai-slop-cleaner (regression-safe)
- Verify:
ac-verify (single pass) or ultraqa (bounded cycles)
- Archive: write WAL fragments + archive the task_brief
- Remember (optional): invoke
remember skill to extract lessons learned into long-term memory. Use when the task surfaced non-obvious project constraints, anti-patterns, or validated approaches.
Orchestration Steps
Step 0 — Intake + Scope
Run input-classifier on the raw input. Use its [Intake] block as the seed for Steps 1–2.
Capture:
- Objective (what changes, what must not)
- File boundary (explicit in-scope paths)
- Acceptance criteria (what evidence means “done”) — from
[Intake].AC-Candidates if available
Step 1 — Create Task Artifacts
Create:
<slug>_task_brief.md<slug>_current_task.md
Step 2 — Run Phases in Order
- Plan → Decisions → Eval → (Approval Gate) → Implement → Verify → Archive
Step 3 — Iteration Policy (if not met)
If acceptance criteria are not met:
- Produce a short delta analysis (what failed, why)
- Propose one next iteration (plan change or fix)
- Ask the user to approve another iteration (no silent looping)
Related Skills
- eval-harness: MUST precede self-improve; defines the objective baseline
- self-improve: bounded improvement loop requiring eval-harness baseline
- remember: optional post-archive knowledge preservation
- external-research: plateau-breaker for stalled self-improve cycles
