athenai-implement

---

name: athenai-implement description: End-to-end implementation pipeline — design, plan, tasks, implement, test cases, test content. Demo mode: skips unit tests and research deep-dives. Records phase wall times via .specify/scripts/phase-timer.py (per-phase start/end, subagent logs + merge when parallel). Explicit invocation only — never load from context or topic. Optional flags: --no-parallel (run all phases sequentially in main agent, no subagents), --no-test-cases (skip Phase 4 test case generation). disable-model-invocation: true

Athenai Implement — Unified Pipeline

Runs the full implementation pipeline from a single spec.md input, printing a progress report after each phase. Skips non-essential quality gates to maximize visible output speed.

Each phase is defined in a separate skill. The main agent reads the appropriate phase skill for each step.

What This Skill Does

Executes six core phases plus an optional manual asset step (0-assets). By default, Phases 2-4 use parallel subagents to minimize wall-clock time. Pass flags to change this (see Flags below).

Each phase is implemented in a dedicated skill that the agent must read before executing that phase:

• Phase 0 — Read athenai-design skill
• Phase 1-2 — Read athenai-plan skill
• Phase 3 — Read athenai-build skill, then building-blocks (companion read for block JS/CSS conventions; primary phase skill remains athenai-build)
• Phase 4 — Read athenai-testcases skill
• Phase 5 — Read athenai-verify skill

Phase 3 implementation aligns with building-blocks; broader content-first workflow remains under content-driven-development per AGENTS.md.

Default (parallel) execution:

[Input: spec.md validation + org/repo] ← all parallel at start
    ↓
Phase 0 — Design (conditional, Figma) + pre-load Phase 1 context
    ↓
Optional manual step: 0-assets (run `athenai-assets` when binaries are needed)
    ↓
Phase 1 — Plan (includes Tasks for Simple/Standard blocks + SC001 scaffolding)
    ↓
    ├── Subagent A: Phase 3 (Implement) — or Phase 2 → 3 if Complex
    └── Subagent B: Phase 4 (Test Cases) ← runs while A is working
    ↓ (both done)
Phase 5 — Test Content (local draft HTML only; DA upload via /speckit-da-upload)

Phase	Output	Runs	Skill
0 — Design (conditional)	design.md	Sequential	athenai-design
0-assets (optional)	files under drafts/media/ + asset report	Manual/separate	athenai-assets
1 — Plan	plan.md (with Authored HTML Structure + Parsing Logic + Tasks for Simple/Standard), draft HTML	Sequential	athenai-plan
2 — Tasks (Complex only)	plan.md (Tasks section filled)	Subagent A (or main agent if --no-parallel); skipped for Simple/Standard	athenai-plan
3 — Implement	Block JS, CSS	Subagent A (or main agent if --no-parallel)	athenai-build
4 — Test Cases	testcases.csv	Subagent B (or main agent / skipped)	athenai-testcases
5 — Test Content	Local HTML (drafts/{name}.plain.html)	Sequential (waits for active phases)	athenai-verify

Phase 3 companion read: In addition to athenai-build, the implementing agent (Subagent A, Cursor eds-frontend-implementer, or main agent when --no-parallel) must read building-blocks before editing blocks/**.

Flags

These flags are parsed from the invocation command before any phase starts. Both are independent and composable.

Flag	Effect
--no-parallel	Skip subagents entirely. Phases 2 → 3 → 4 run sequentially in the main agent. Eliminates subagent startup overhead at the cost of longer wall-clock time.
--no-test-cases	Skip Phase 4 entirely. No testcases.csv is generated. Useful when test cases already exist or are not needed for this run.

Behaviour matrix:

Invocation	Phase 2+3	Phase 4	Phase 5 waits for
speckit-fast	Subagent A	Subagent B (parallel)	A + B
speckit-fast --no-parallel	Main agent, sequential	Main agent, sequential	Phase 4 complete
speckit-fast --no-test-cases	Subagent A	skipped	A only
speckit-fast --no-parallel --no-test-cases	Main agent, sequential	skipped	Phase 3 complete

What Is Skipped (Demo Mode)

• Unit tests: No npm test, no .test.js file generation, no TS006 tasks
• speckit-analyze: Cross-artifact consistency check removed
• speckit-validate: Full project validation chain removed
• speckit-checklist: Requirements quality gate removed
• speckit-clarify: Spec is assumed clean; no Q&A loop
• speckit-document: Authoring guide generation removed
• Phase 0 research: No block-collection-and-party or docs-search runs unless block type is completely unrecognised

Pre-conditions

The AEM dev server MUST already be running on http://localhost:3000 before invoking this pipeline. The pipeline will not start, stop, or restart any dev server. If the server is not running, abort with:

ERROR: Dev server not reachable at http://localhost:3000.
Start it first: npx -y @adobe/aem-cli up --no-open --html-folder drafts

Verify with: curl -s -o /dev/null -w "%{http_code}" http://localhost:3000 — any 2xx or 3xx response is acceptable. Always use port 3000; never bind to an alternative port.

Input

A single spec.md file.

Flag Parsing (first, before anything else)

Parse the invocation command for flags and store them as pipeline-wide variables:

• --no-parallel found → NO_PARALLEL=true (default: false)
• --no-test-cases found → NO_TEST_CASES=true (default: false)

Flags are case-sensitive and position-independent. Unknown flags are ignored with a warning.

Then run these steps in parallel before any phase starts:

A — Locate spec and validate

1. Run .specify/scripts/bash/check-prerequisites.sh --json from repo root — parse FEATURE_DIR and AVAILABLE_DOCS.
2. Verify FEATURE_DIR/spec.md exists. If missing, abort with:

ERROR: spec.md not found in FEATURE_DIR.
Run /speckit-specify first, or provide the path to an existing spec.md.

3. Validate spec.md has at least these sections (abort and report missing ones before any phase runs):
   • User Stories or User Journeys
   • Acceptance Criteria
   • Content Model or Block Structure

B — Resolve org/repo Run git remote -v, parse origin to extract {org} and {repo}. Store for use in Phase 5 summary links. Do not ask the user unless parsing fails.

C — Initialize execution timer

After A and B succeed, from the repository root, the executing agent must shell out to .specify/scripts/phase-timer.py (Python 3). If A or B fails, do not initialize the timer.

1. RUN_ID — {YYYYMMDD}-{HHMMSS}-{slug}:

   • Date/time: UTC, date -u +%Y%m%d-%H%M%S
   • {slug}: basename of FEATURE_DIR, lowercased; replace any run of non-[a-z0-9] with a single -; trim leading/trailing -; if empty use speckit

2. Persist RUN_ID for the entire pipeline (include it in subagent prompts when using parallel mode).

3. Init (writes the run boundary to the default log — see Execution time tracking):

   python3 .specify/scripts/phase-timer.py init "$RUN_ID"
   

Default log path: .specify/logs/timings.log (override only for subagents via SPECIFY_TIMER_LOG, below).

Once A, B, and C are complete, proceed to D and then Phase 0.

D — Launch draft-content-generator (background)

Immediately after A, B, and C succeed, launch the draft-content-generator subagent in the background using the Task tool (run_in_background: true, model: fast). This runs in parallel with Phase 0 and produces drafts/{blockname}.plain.html early, saving ~30-60s from Phase 1.

Provide the subagent with:

• FEATURE_DIR — the spec directory path
• BLOCK_NAME — extracted from spec.md (the block name from the Scope or Content Model section)

The subagent checks if the file already exists and skips if so. Phase 1 also checks: if drafts/{blockname}.plain.html exists when Phase 1 runs its "Draft Test Content" step, Phase 1 skips draft generation entirely and notes "draft content pre-generated by background agent" in the report.

Do not wait for this subagent before starting Phase 0. Phase 5 verifies the file exists regardless of which phase produced it.

For execution time tracking details, see resources/execution-timing.md in this skill directory.

---

Progress Report Format

After each phase completes, print this markdown section inline:

---
## ✓ Phase N — [Name] complete
- **Outputs**: list files created or updated
- **Key decisions**: notable choices made (e.g. "no design.md, proceeding without Figma context")
- **Warnings**: any non-blocking issues
- **Next**: Phase N+1 — [Name]
---

At the end print a final ## Summary section with all outputs, file paths, execution time report (phase-timer.py report), and any accumulated warnings. If the pipeline aborts after init but before the normal Phase 5 Summary, still run finalize and report, and include the pasted report output in the abort/summary message.

---

Execution time tracking

For complete execution time tracking instructions (phase keys, commands, minimizing dead time, parallel mode, finalize/report), see resources/execution-timing.md in this skill directory.

---

Phase 0 — Design (Conditional)

Read the athenai-design skill and follow its instructions for Phase 0.

If design output references asset binaries that are not yet materialized, run athenai-assets manually as a separate optional 0-assets timed step before Phase 3.

---

Phase 1-2 — Plan and Tasks

Read the athenai-plan skill and follow its instructions for Phases 1 and 2.

Phase 2 is only executed for Complex blocks (9+ tasks); Simple/Standard blocks inline the Tasks section during Phase 1.

---

Execution Split — After Phase 1

Immediately after Phase 1 completes, determine the execution mode based on the flags set during input parsing and the complexity classification.

If Phase 1 inlined the Tasks section (Simple/Standard), skip Phase 2 and proceed directly to Phase 3 (or to the Subagent A launch with Phase 3 only).

If NO_PARALLEL=false (default)

Launch subagents in parallel:

Phase 1 done (Tasks inlined for Simple/Standard)
    ├── Subagent A: Phase 3 (Implement) — Tasks already in plan.md
    │   (or Phase 2 → Phase 3 if Complex)
    └── Subagent B: Phase 4 (Test Cases) ← runs while A is working
         (skip Subagent B entirely if NO_TEST_CASES=true)

Each subagent must receive full input context: FEATURE_DIR, RUN_ID, spec.md contents, plan.md, and design.md (if present). Subagents do not share the parent's memory. Subagent A must verify any referenced drafts/media/* assets exist before implementing styles that depend on them. Subagent A must read athenai-build then building-blocks before editing blocks/** (same read order as Cursor eds-frontend-implementer).

Timer (mandatory): Instruct Subagent A to set SPECIFY_TIMER_LOG=".specify/logs/timings-agent-a.log" before any phase-timer.py start call, and to run start/end for 2-tasks then 3-implement. If Subagent B runs, instruct it to set SPECIFY_TIMER_LOG=".specify/logs/timings-agent-b.log" and run start/end for 4-testcases only. The main agent, after both subagents finish, runs merge (see Execution time tracking) before starting Phase 5 timer.

File conflict: Subagent A updates plan.md (Tasks section) then writes block files; Subagent B writes testcases.csv. No conflict.

Subagent B failure: If Phase 4 fails (e.g. spec contains [NEEDS CLARIFICATION]), warn but do not block Phase 3 → 5 from completing.

Cursor: Subagent A = eds-frontend-implementer (optional path)

When the orchestrator runs in Cursor and delegates Phase 3 via the Task tool or /eds-frontend-implementer, treat that custom subagent as Subagent A (same responsibilities and timings-agent-a.log as generic Subagent A).

• Subagent file: .cursor/agents/eds-frontend-implementer.md (model: fast).
• Context bundle (include in the delegation prompt): FEATURE_DIR, RUN_ID, full spec.md, plan.md, and design.md if present. Subagent A must verify referenced drafts/media/* assets exist before implementing styles that depend on them.
• Mandatory instructions for the delegation prompt:
  1. Read athenai-build fully and execute Phase 3 per that skill.
  2. Read building-blocks before writing block JS/CSS.
  3. Follow Execution time tracking (resources/execution-timing.md): set SPECIFY_TIMER_LOG=".specify/logs/timings-agent-a.log" before any phase-timer.py call; start/end for 2-tasks then 3-implement as already required for Subagent A.

Non-Cursor or environments without this subagent keep using generic “Subagent A” with the same three bullets (read order + timer rules).

If NO_PARALLEL=true

Run all remaining phases sequentially in the main agent — no subagents are launched:

Phase 1 done (Tasks inlined for Simple/Standard)
    → Phase 3 (Implement) — main agent
    → Phase 4 (Test Cases) — main agent (skip if NO_TEST_CASES=true)
    → Phase 5 (Test Content)

Phase 1 done (Tasks stub for Complex)
    → Phase 2 (Tasks) — main agent
    → Phase 3 (Implement) — main agent
    → Phase 4 (Test Cases) — main agent (skip if NO_TEST_CASES=true)
    → Phase 5 (Test Content)

No subagent context-passing is needed; the main agent already holds all context.

Phase 3 reads (sequential mode): Before implementing, the main agent must read athenai-build then building-blocks (same order as Cursor eds-frontend-implementer).

Timer (sequential): On the main log, start/end 2-tasks only if Phase 2 runs (Complex), then start/end 3-implement, then start/end 4-testcases only if NO_TEST_CASES=false.

---

Phase 3 — Implement (Subagent A)

Read the athenai-build skill, then building-blocks, and follow their instructions for Phase 3.

In Cursor, delegate Phase 3 to eds-frontend-implementer when using parallel subagents (see Cursor: Subagent A = eds-frontend-implementer under Execution Split — After Phase 1). In --no-parallel mode, the main agent performs Phase 3 with the same read order.

---

Phase 4 — Test Cases

Read the athenai-testcases skill and follow its instructions for Phase 4.

---

Phase 5 — Test Content

Read the athenai-verify skill and follow its instructions for Phase 5.

---

Key Rules

• Execution time: Follow Execution time tracking for every run that reaches init; finalize + report on success or failure after init.
• Draft .plain.html DOM: Block roots (div.{blockName}) must not be direct children of main next to other section <div>s — decorateSections / decorateBlocks in scripts/aem.js require main > div.section > div > div.{blockName}. Phase 1 must wrap the draft in one outer <div> (or equivalent per-section wrappers). See athenai-plan and athenai-verify.
• design.md is source of truth for HTML/CSS when present; spec.md is source of truth for functional requirements
• CSS: vanilla only, no SCSS/preprocessors, block-scoped selectors, mobile-first file order
• Use @media (width >= Npx) with EDS breakpoints: 600px, 900px, 1200px
• All block JS exports default function decorate(block) { ... }
• Phase 3: Read athenai-build then building-blocks before block edits (Subagent A, Cursor eds-frontend-implementer, or main agent with --no-parallel)
• Lint must pass before Phase 4 — this is the only hard gate in demo mode
• Never modify scripts/aem.js
• Never add !important in CSS