Agent skill
implement-spec
Implements a spec on a ticket-by-ticket basis.
Install this agent skill to your Project
npx add-skill https://github.com/rstacruz/agentic-toolkit/tree/main/skill/atk/implement-spec
SKILL.md
-
Find the spec:
- Find the spec file. It may be mentioned previously in the conversation, or ask the user if it can't be found.
- Ensure that the spec file is on the disk as an .md file. If not, write it to artefacts/.md first.
-
Evaluate and clarify:
- Evaluate if spec makes sense.
- Check if the spec has structured Tickets:
- If no tickets exist, use the
questiontool to ask the user:- Do they want to use
$spec-implementation-planto break work into tickets? - Or do they have feedback on the spec?
- Do they want to use
- If no tickets exist, use the
- If spec has other ambiguities or open questions, ask the user for clarifications first.
-
Prepare:
- Create an empty progress file (
artefacts/progress.md)
- Create an empty progress file (
-
Identify ticket:
- Read the spec file(s) and progress file
- Pick one ticket — the most important one that's not blocked. It may not be the top-most ticket
- Pick only one ticket, never pick more than one
-
Spawn an agent:
- Prepare the ticket input: the identified ticket (ID and title)
- Prepare the spec path:
artefacts/spec.md(or custom path if provided) - Prepare the progress path:
artefacts/progress.md(or custom path if provided) - Spawn a NEW @general-alpha agent with this prompt:
Load the $implement-spec-subagent skill. Pass these inputs: - {{TICKET}}: [ticket ID and title] - {{SPEC_FILE}}: [spec file path] - {{PROGRESS_FILE}}: [progress file path] - The subagent will validate inputs and execute the ticket workflow
-
Verify commit:
- Verify that the agent created a git commit, create one if it didn't
- Verify that commit message references exactly one ticket ID (e.g., T01, T-02). If multiple found, stop and notify user
-
Assess completeness:
- Check if there are any tickets requiring action after this
- If work remains, repeat step 4
- If 20 iterations reached, create summary in progress.md and notify user
- If there are none, continue to step 9
-
Error handling:
- If agent fails: check for partial work, verify any commits, update progress with error state
- Critical failures (file not found, corrupted spec): stop and notify user
- Non-critical failures: document in progress.md and continue next iteration
-
Post-execution polish:
- Load and run the
$refine-implementationskill - Pass the git range as context (e.g.,
main...HEAD) — let the skill use git commands to retrieve the diff; do not pass the diff directly - This step runs once, after all tickets are complete
- Load and run the
Important reminders:
- Always ask for review after ticket - this greatly impacts build quality
- Only 1 ticket per agent. Do not ask agent to do more than 1 ticket
- Use new agent sessions every iteration (eg, don't reuse session_id)
- Let the
$implement-spec-subagentskill handle all execution details
Recommended Agent Skills
Expand your agent's capabilities with these related and highly-rated skills.
refine-implementation
Use after implementation to simplify and review code. Provide: git range (eg, main...HEAD). Runs simplify + peer review loop until change set is clean.
spec-product-requirements
Gives important guidelines to define product requirements sections (functional requirements, technical requirements, constraints, design considerations, diagrams). Companion to $spec-mode.
coding-practices
Contains important guidelines for software engineering, coding, programming. Includes (but not limited to): - CP1: Functional core, imperative shell - CP2: Operational vs unexpected errors - CP3: Result-oriented interface pattern - CP4: Presentational vs container components - CP5: Log context builder pattern Use this when writing, editing, debugging, planning, or otherwise working with: - Any programming work - UI components in React, Vue, or similar - JavaScript, TypeScript, Rust, or any programming language
spec-tech-design
Gives important guidelines to define technical design sections (call graphs, data models, pseudocode, files, CSS classes, testing strategy). Companion to $spec-mode skill.
review-changes
Use when reviewing code changes against a plan. Provide: plan/spec doc; git range or changed files (eg, branch...HEAD). Returns P1/P2/P3 on alignment, quality, bugs, security.
simplify
Simplifies the implementation
Didn't find tool you were looking for?