Automating the Full Development Cycle with Claude Code’s /prd, /goal, and /after-goal

Three‑Stage AI‑Assisted Development Workflow

Claude Code provides three slash‑commands— /prd, /goal, and /after-goal —that can be combined into a reusable Skill to automate the full pipeline from requirement definition to code merge.

Stage 1 – /prd : Requirement Decomposition

Install the PRD Skill at https://console.cloud.baidu-int.com/onetool/skills/4793.

Run /prd with the product idea “诊断案例记录与反馈闭环”. Claude Code generates a structured PRD document saved to tasks/prd-self-evolution.md and creates five iCafe cards covering description, acceptance criteria, priority, and dependencies.

The PRD forces clarification of user stories, non‑goals, and card dependencies before any code is written, reducing misunderstanding and missing edge cases.

Stage 2 – /goal : Card‑Based Implementation

Each card is processed individually with /goal. Two representative cards are shown.

Card 46 – Case data model & Markdown I/O library

Claude Code queries the card via icafe-cli card get to obtain description and acceptance criteria.

Project context is read from diag/CLAUDE.md and diag/model/types.go.

Four core functions are generated: WriteCase, ReadCase, ListCases, UpdateCase.

Ten unit tests cover round‑trip serialization, file I/O, list filtering, partial updates, empty results, special characters, and error scenarios.

All checks ( go vet, go build, go test ./diag/...) pass.

Card 47 – Pipeline integration

The exported function claude.BuildPrompt is extracted and a new configuration field PipelineConfig.CasesDir is added to make case‑file output optional. Pipeline.Run is modified to defer writing the case file, handling success and failure paths with golog.Warn without blocking the main flow.

The helper extractSections is fixed to locate content by known section titles ( System Prompt, User Prompt) instead of naïve newline splitting, correctly handling sub‑headings ( ##).

Three new tests verify successful write, failure handling, and behavior when CasesDir is empty.

The /goal command performs code generation and testing; it does not submit code or update cards.

Stage 3 – /after-goal : Automatic Submission, Review, Merge, and Card Closure

Install the after‑goal Skill at https://console.cloud.baidu-int.com/onetool/skills/4797.

After a card’s code is generated, Claude Code automatically invokes the after‑goal Skill.

Step 1 – Commit : git add diag/cases/case.go diag/llm/claude/claude.go … and

git commit -m "baidu-sys-nccl-sys-skills-47 Pipeline 诊断完成后自动写入 Case 文件"

. The commit message starts with the card ID, satisfying iCode’s requirement to bind the commit to a card.

Step 2 – Push to Gerrit : git push origin HEAD:refs/for/master. The refs/for/ prefix creates a review request and returns the CR number.

Step 3 – Review & Merge

icode-cli api get_review_info -n 120869646 -o table
icode-cli api set_review_score -r baidu/sys-nccl/sys-skills -n 120869646 -s 2
icode-cli api submit_review -r baidu/sys-nccl/sys-skills -n 120869646

Step 4 – Update Card Description

icafe-cli card update --space baidu-sys-nccl-sys-skills --sequence 47 \
  --detail "original description<h3>实现总结</h3><ul>…</ul>"

The --detail flag overwrites the description, so the original content must be retained and the implementation summary appended.

Step 5 – Close Card : Determine the correct status name (e.g., “已完成” or “Done”) with icafe-cli card next-statuses --space … --sequence 47, then close the card:

icafe-cli card update --space … --sequence 47 --status 已完成

Key Technical Observations

PRD first : A structured PRD forces clear scope and acceptance criteria, preventing rework caused by AI mis‑interpreting vague requirements.

Appropriate card granularity : Small, well‑defined cards (e.g., separate data‑model and pipeline integration) enable AI to implement each piece correctly without losing focus.

Dependency‑driven order : Implement cards according to their dependency graph (Card 47 depends on Card 46) to avoid conflicts.

CLI over browser : Using CLI tools ( icafe-cli, icode-cli, git) is faster and more reliable than manual browser interactions.

Skill as knowledge carrier : The /after-goal Skill encapsulates the entire post‑implementation process (commit, push, review, merge, card update, closure), turning tacit operational knowledge into reusable automation.

Early bug fixes : Issues discovered during a card’s implementation (e.g., extractSections failing on sub‑headings) are fixed immediately, minimizing downstream cost.

Reference Implementation

The workflow is open‑sourced at https://github.com/smallnest/goal-workflow and documented at https://goal.rpcx.io/.