
spec-driven-translation
by JonahSussman
Experimentation with CLI agents replacing the modern SWE zeitgeist.
SKILL.md
name: spec-driven-translation description: | Spec-driven source translation & modernization workflow for codebases. Use when asked to migrate/translate a project (e.g., framework/SDK/language upgrade), or when the user wants a structured, externally-validatable implementation loop with PRD + plans + checklists and iterative validation/remediation.
Spec-Driven Translation
This Skill executes a documentation-first, validation-driven workflow to translate or modernize a codebase. It turns the existing source into a Product Requirement Document (PRD), derives a Project Specification, Implementation Plan, and Task Execution Checklist, then iteratively implements the work in externally-validatable steps with a remediation loop until success.
High-level flow (from the provided diagram):
- Ingest -> PRD
- Plan translation
- Docs to guide AI (Project Spec, Implementation Plan, Task Checklist)
- Iterative task loop (choose task -> implement step -> run validation -> remediate w/ retry budget)
- Translated Source Code
Preconditions & Setup
- Identify the source tree and the target environment (runtime, language, framework versions, build tools).
- Ensure we have authority to read/write files and run the project’s build/test commands.
- If needed, ask to install or enable required tools (build system, linters, test runner, package managers).
Artifacts We Will Create
templates/PRD.md-> filled asProduct_Requirement_Document.mdtemplates/Project-Specification.md->Project_Specification.mdtemplates/Implementation-Plan.md->Implementation_Plan.mdtemplates/Task-Execution-Checklist.md->Task_Execution_Checklist.md
Keep these docs up-to-date as living sources of truth. They guide all subsequent steps.
Step-by-Step Instructions
1. Ingest -> PRD
- Read the existing codebase (entry points, modules, configs, CI, tests).
- Extract intent, behaviors, constraints, SLAs, non-functionals, and integration points.
- Populate and save
Product_Requirement_Document.mdfromtemplates/PRD.md. - Call out unknowns, risks, and assumptions explicitly.
2. Plan Translation
- Define target state (language/framework versions, platform changes, build/test strategy, deployment).
- Fill in
Project_Specification.mdgrounding every requirement in the PRD. - Draft
Implementation_Plan.md(milestones, work streams, sequencing, acceptance criteria). - Create
Task_Execution_Checklist.md(repeatable per-task DoD & safety checks).
3. Choose-Implement-Validate Loop
For each iteration:
-
Choose next task
- Use rubric (below) to pick a task that yields an externally-verifiable delta (e.g., unit test passes, binary builds, CLI works, endpoint responds).
- Update the checklist with the task ID and success conditions.
# Pick-Next-Task Rubric Score each candidate task 0–3 on: - External verifiability (clear test/build signal) - Risk isolation (blast radius if wrong) - Unblocks others (dependency breaker) - Knowledge gain (reduces uncertainty) - Effort (fit in one iteration) Choose the highest total; tie-break on verifiability. -
Implement a minimal, verifiable step
- Make small changes scoped to one verifiable outcome.
- Prefer creating/updating tests that encode the desired behavior.
-
Validation
- Run validation.
- Treat builds/tests/lints/type-checks/formatters as the external oracle of success.
-
Remediation (bounded retries)
- On failure, capture logs; fix the smallest issue first; re-run validation.
- Limit to a local retry budget (e.g., 3 attempts) before escalating:
- Update docs with new facts found.
- Re-plan or split the task if necessary.
-
Commit
- When validation passes, commit with an evidence-rich message (what/why, tests affected, acceptance criteria met).
- Append outcomes/notes to the Implementation Plan.
Repeat until the milestone is complete, then proceed to the next milestone.
4. Exit Criteria
- All tasks in the current milestone have green validation.
- The translated artifact(s) function under the target environment with documented acceptance evidence.
- PRD, Spec, Plan, and Checklist are updated to reflect reality.
Pick-Next-Task Rubric (summary)
Prefer tasks that:
- Increase test coverage or convert implicit behavior into explicit tests.
- Deliver a user-visible or CI-visible signal (a test or build turns green).
- Reduce coupling or blockers for later tasks.
- Are reversible and low-risk when uncertain.
Full rubric in scripts/pick-next-task.md.
Notes
- Keep docs concise but always actionable. Treat them as contracts with the validator.
- When in doubt, shrink the task until it’s externally verifiable within a single iteration.
- Prefer to keep the human in the loop, stopping for confirmation on assumptions or changes to the plan.
スコア
総合スコア
リポジトリの品質指標に基づく評価
SKILL.mdファイルが含まれている
ライセンスが設定されている
100文字以上の説明がある
GitHub Stars 100以上
3ヶ月以内に更新がある
10回以上フォークされている
オープンIssueが50未満
プログラミング言語が設定されている
1つ以上のタグが設定されている
レビュー
レビュー機能は近日公開予定です