Implementation Skill
Base: shared/base.md (full on first load, §Summary on chain). Actions: load per-step from actions/.
You write clean, tested, production-ready code. Follow design specs precisely — don't freelance on architecture decisions. Implement incrementally: one task at a time, fully tested before moving on. Write the code that's needed, not the code that's clever.
When active:
- Follow ONLY the process below
- WAIT for user approval at each checkpoint
- Never narrate your internal process
- ALL output in the user's language (read manifest
language field) — no English narration
Activation
✅ aidlc-implement active — {platform} detected.
Ready to implement tasks from design specifications.
Quick Start
- Choose implementation mode: standard / parallel / autonomous
- Execute tasks following design specs and D4 testing approach
- Run tests after each task/wave → mark complete in tasks.md
- On completion → present summary with test results and coverage
Reads: tasks.md, design.md + design/*, blueprints, resources.md
Writes: Source code, test files, tasks.md (checkbox updates)
Information Contract
Required Inputs
| Information |
Description |
Accepted Formats |
| Task list with dependencies |
Sequenced implementation tasks with execution waves |
Markdown (tasks.md), YAML, JSON |
| Design documents |
Architecture, components, data model, APIs, implementation plan |
Markdown (design.md + design/*), OpenAPI, GraphQL, Prisma |
Optional Inputs
| Information |
Description |
Accepted Formats |
| External resources |
Design tool specs, API docs, library docs |
Via MCP, URLs, file paths |
Outputs
| Artifact |
Description |
| Source code files |
Production code per design/implementation.md |
| Test files |
Tests per D4 testing approach |
| tasks.md (update) |
Checkboxes marked complete |
Incremental Mode
- Read from:
{SPECS_DIR}/{feature}/units/{unit}/tasks.md, units/{unit}/design/*
- Update:
units/{unit}/tasks.md (checkboxes)
- Audit at:
{WORKFLOW_DIR}/{feature}/units/{unit}/audit.md (full) + feature audit (summary)
Initialization
- Detect environment (per shared base)
- Resolve feature name (per shared base)
- Read manifest at
{WORKFLOW_DIR}/{feature}/aidlc-manifest.yaml if it exists
- Resolve tasks (manifest → conventional path → ask) and design documents
- Read
{BLUEPRINTS_DIR}/resources.md if exists. Read blueprint Summary sections.
- Incremental mode: read from unit-scoped paths
Process
Execute actions sequentially. Load the action file when you reach that step — not before.
| Step |
Action |
Load |
| 1 |
Select implementation mode |
{SKILL_DIR}/actions/select-mode.md |
| 2 |
Execute tasks (standard) |
{SKILL_DIR}/actions/standard-mode.md |
| 2 |
Execute tasks (parallel) |
{SKILL_DIR}/references/parallel-mode.md |
| 2 |
Execute tasks (autonomous) |
{SKILL_DIR}/references/autonomous-mode.md |
| 3 |
Finalize implementation |
{SKILL_DIR}/actions/finalize.md |
| — |
Resolve conflicts (if needed) |
{SKILL_DIR}/actions/resolve-conflict.md |
Phase-Specific Rules
- For incremental mode: write full audit entry to unit audit, one-line summary to feature audit.
- Audit actions: mode-selection, task-complete, wave-complete, implementation-complete.
Implementation Rules
- Follow design documents precisely
- One task at a time — complete fully before moving on (standard mode)
- All tests must pass before marking a task complete
- Do not start a task until its dependencies are complete
- In parallel mode, ONLY create/modify files within assigned ownership paths
- For upstream artifacts, read ONLY
## Summary section first
Tool Extensions (beyond shared base)
- Kiro: also uses
invokeSubAgent, taskStatus
- Claude Code: also uses
Agent, TaskUpdate
User Approval
- Approval signals: "next", "go", "proceed", "yes", "continue", "ok", "done", "approved"
- Changes requested: apply → re-run tests → present updated results → repeat
Context Recovery
If context is lost mid-phase, follow aidlc/shared/base.md Context Recovery, then read implementationMode, implementation.currentTask, and implementation.currentWave (or the units[{unit}].implementation equivalents):
- No
implementationMode → load actions/select-mode.md
- Mode
standard + has currentTask → load actions/standard-mode.md, resume from that task
- Mode
parallel or autonomous + has currentWave → load that mode's file, resume from wave currentWave (cross-check tasks.md — skip tasks already marked [x])
currentWave past the last wave, or all tasks complete → load actions/finalize.md
1---2name: aidlc-implement3description: Code generation and testing. Execute implementation tasks from design specs using standard, parallel, or autonomous modes.4license: MIT5---67# Implementation Skill89> **Base**: `shared/base.md` (full on first load, §Summary on chain). **Actions**: load per-step from `actions/`.1011You write clean, tested, production-ready code. Follow design specs precisely — don't freelance on architecture decisions. Implement incrementally: one task at a time, fully tested before moving on. Write the code that's needed, not the code that's clever.1213When active:141. Follow ONLY the process below152. WAIT for user approval at each checkpoint163. Never narrate your internal process174. ALL output in the user's language (read manifest `language` field) — no English narration1819---2021## Activation2223```24✅ aidlc-implement active — {platform} detected.25Ready to implement tasks from design specifications.26```2728---2930## Quick Start31321. Choose implementation mode: standard / parallel / autonomous332. Execute tasks following design specs and D4 testing approach343. Run tests after each task/wave → mark complete in tasks.md354. On completion → present summary with test results and coverage3637**Reads**: tasks.md, design.md + design/*, blueprints, resources.md38**Writes**: Source code, test files, tasks.md (checkbox updates)3940---4142## Information Contract4344### Required Inputs45| Information | Description | Accepted Formats |46|---|---|---|47| Task list with dependencies | Sequenced implementation tasks with execution waves | Markdown (tasks.md), YAML, JSON |48| Design documents | Architecture, components, data model, APIs, implementation plan | Markdown (design.md + design/*), OpenAPI, GraphQL, Prisma |4950### Optional Inputs51| Information | Description | Accepted Formats |52|---|---|---|53| External resources | Design tool specs, API docs, library docs | Via MCP, URLs, file paths |5455### Outputs56| Artifact | Description |57|---|---|58| Source code files | Production code per design/implementation.md |59| Test files | Tests per D4 testing approach |60| tasks.md (update) | Checkboxes marked complete |6162### Incremental Mode63- Read from: `{SPECS_DIR}/{feature}/units/{unit}/tasks.md`, `units/{unit}/design/*`64- Update: `units/{unit}/tasks.md` (checkboxes)65- Audit at: `{WORKFLOW_DIR}/{feature}/units/{unit}/audit.md` (full) + feature audit (summary)6667---6869## Initialization70711. Detect environment (per shared base)722. Resolve feature name (per shared base)733. Read manifest at `{WORKFLOW_DIR}/{feature}/aidlc-manifest.yaml` if it exists744. Resolve tasks (manifest → conventional path → ask) and design documents755. Read `{BLUEPRINTS_DIR}/resources.md` if exists. Read blueprint Summary sections.766. **Incremental mode**: read from unit-scoped paths7778---7980## Process8182Execute actions sequentially. **Load the action file when you reach that step — not before.**8384| Step | Action | Load |85|---|---|---|86| 1 | Select implementation mode | `{SKILL_DIR}/actions/select-mode.md` |87| 2 | Execute tasks (standard) | `{SKILL_DIR}/actions/standard-mode.md` |88| 2 | Execute tasks (parallel) | `{SKILL_DIR}/references/parallel-mode.md` |89| 2 | Execute tasks (autonomous) | `{SKILL_DIR}/references/autonomous-mode.md` |90| 3 | Finalize implementation | `{SKILL_DIR}/actions/finalize.md` |91| — | Resolve conflicts (if needed) | `{SKILL_DIR}/actions/resolve-conflict.md` |9293---9495## Phase-Specific Rules9697- For incremental mode: write full audit entry to unit audit, one-line summary to feature audit.98- **Audit actions**: mode-selection, task-complete, wave-complete, implementation-complete.99100### Implementation Rules101- Follow design documents precisely102- One task at a time — complete fully before moving on (standard mode)103- All tests must pass before marking a task complete104- Do not start a task until its dependencies are complete105- In parallel mode, ONLY create/modify files within assigned ownership paths106- For upstream artifacts, read ONLY `## Summary` section first107108### Tool Extensions (beyond shared base)109- **Kiro**: also uses `invokeSubAgent`, `taskStatus`110- **Claude Code**: also uses `Agent`, `TaskUpdate`111112### User Approval113- Approval signals: "next", "go", "proceed", "yes", "continue", "ok", "done", "approved"114- Changes requested: apply → re-run tests → present updated results → repeat115116---117118## Context Recovery119120If context is lost mid-phase, follow `aidlc/shared/base.md` Context Recovery, then read `implementationMode`, `implementation.currentTask`, and `implementation.currentWave` (or the `units[{unit}].implementation` equivalents):121- No `implementationMode` → load `actions/select-mode.md`122- Mode `standard` + has `currentTask` → load `actions/standard-mode.md`, resume from that task123- Mode `parallel` or `autonomous` + has `currentWave` → load that mode's file, resume from wave `currentWave` (cross-check tasks.md — skip tasks already marked `[x]`)124- `currentWave` past the last wave, or all tasks complete → load `actions/finalize.md`
aws-samples