bootstrapping-design-docs

For the rest of this document {DESIGN_DOC_SKILL_NAME} will refer to the name of the skill we will make that will later be used to generate design docs for this project. Default: creating-design-doc. Once you create the base design doc template and skill, you must ask questions to the user to help clarify anything you are unsure of. This is a joint user-agent process: ask targeted questions when unsure; do not finalize templates without user review.

<file_list>

• design-docs/README.md - How design docs are written, reviewed, and used for delegated execution.
• design-docs/active/ - Directory for active design documents.
• design-docs/agents/ - Directory for agent-executable XML subtasks.
• .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/SKILL.md - The generated ${DESIGN_DOC_SKILL_NAME} skill.
• .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/templates/README.md - Lists available templates and when to use them.
• .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/templates/base.md - Meta "base" design doc template (guardrails + structure).
• .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/templates/<variant>.md - Project-specific templates (e.g., etl-pipeline.md, feature.md). </file_list>

Note: Templates are bundled with the skill (not in design-docs/) per Agent Skills progressive disclosure pattern.

<output_format> Design docs are written in markdown - not XML. Markdown is easier for humans to read/edit and for agents to parse during collaboration. Agent-executable subtasks use XML in a separate file under design-docs/agents/. </output_format>

<agent_name_resolution> Replace {AGENT_NAME} with the agent's config directory:

• .aider/ - Aider
• .claude/ - Claude Code
• .codex/ - OpenAI Codex CLI
• .copilot/ - GitHub Copilot
• .cursor/ - Cursor

If the repo has an existing agent config directory, use it. If multiple exist, prefer the current agent's. If none exist, create one. </agent_name_resolution>

<generation_steps>

1. Create folder structure (if missing):

   • design-docs/
   • design-docs/active/
   • design-docs/agents/
   • .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/
   • .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/templates/

2. Write design doc skill and bundled templates:

   • design-docs/README.md (workflow + approvals)
   • .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/SKILL.md (the skill)
   • .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/templates/README.md (template index)
   • .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/templates/base.md (meta template with guardrails)
   • .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/templates/skill.md (skill creation template)
   • .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/templates/feature.md (feature template)

3. Run Adaptive Questioning Protocol:

   • Produce "Detected Defaults Summary"
   • Ask triggered questions
   • Stop for answers before proceeding

4. Idempotency:

   • Never overwrite existing templates without timestamped backup
   • Add new variants to .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/templates/README.md </generation_steps>

<questioning_protocol> Ask only questions that materially influence template contents.

<question_tags> Every question must be tagged:

• [blocking] - must be answered to generate correct template or enforce safety
• [important] - significantly improves template quality / reduces future rework
• [optional] - preference-level; safe defaults exist </question_tags>

<question_budget>

• Max 10 questions total.
• Each question includes a one-line "why" tying to a template section, guardrail, or convention. </question_budget>

<triggered_questions> Ask only what applies based on scan:

• Migrations tooling detected -> ask DB reset/backfill/rollback/testing policy
• HTTP/WS dependencies detected -> ask record/replay + fixture storage constraints
• Monorepo structure detected -> ask where templates live and how commands differ
• No tests/CI detected -> ask what canonical verification commands should be </triggered_questions>

<template_delta_preview> Before writing files, show:

• Which variants will be created
• Repo-specific defaults that will be baked in
• Any tightened guardrails derived from project norms
• Any unresolved [blocking] items

Stop and ask for [blocking] answers before proceeding. </template_delta_preview> </questioning_protocol>

<common_questions> Ask only what you cannot infer from the scan: <question_examples>

1. [blocking] Template variants desired

   • "Which templates do you want generated? (e.g., feature, etl_http_ws, db_migration, refactor)"
   • Why: determines templates bundled with skill.

2. [blocking] Test integrity policy

   • "What is your rule on mocks? (only mock boundaries; prefer record/replay; never mock core transforms)"
   • Why: populates guardrails + testing strategy.

3. [blocking] Data capture constraints

   • "Can we store captured payloads under tests/fixtures/? Any size/licensing/PII constraints?"
   • Why: populates test-data acquisition plan.

4. [blocking] Destructive actions boundary

   • "What DB reset patterns are allowed in tests? (transactions/temp schema/containers). Forbidden actions?"
   • Why: prevents unsafe "delete DB to pass tests" behavior.

5. [important] Type discipline

   • "For Python projects, what is your stance on Optional/None? (forbid unless justified; require invalid states unrepresentable)"
   • Why: shapes interface-contract requirements.

6. [important] Review workflow

   • "Stop at Template Delta Preview for approval, or write files and you review diffs?"
   • Why: controls collaboration checkpoint. </question_examples> </common_questions>

<base_template_sections> The base template MUST include these sections (in order):

1. Identity and lifecycle
2. Context
3. Problem
4. Goals / Non-goals
5. Requirements (FRs / NFRs)
6. Constraints and invariants (include DB + test integrity)
7. Proposed design (architecture, contracts, schemas, idempotency/retries)
8. Project common utilities that will be used.
9. Project common utilities that will be created.
10. Interface contracts (explicit signatures; Optional/None justification required)
11. Alternatives considered (>=2)
12. Test data acquisition plan (mandatory for external I/O / ETL)
13. Testing and verification strategy (commands + CI gates + test integrity rules)
14. Rollout / migration / ops (flags, backfills, observability, runbook)
15. Open questions / follow-ups
16. Subtasks (human-readable checklist for review)

Subtasks Section Format:

## Subtasks

### T1: [Task Title]
- **Summary**: One sentence objective
- **Scope**: What's in / what's out
- **Acceptance**: Binary pass/fail criteria
- **Status**: [ ] Not started / [~] In progress / [x] Complete

### T2: [Task Title]
...

</base_template_sections>

• Reuse-first rule (anti-duplication): search existing utilities; record decision per task.
• No destructive shortcuts: never delete dev/prod data to pass tests; destructive actions require confirmation.
• Test integrity (anti-mock-cheating): mock boundaries only; prefer record/replay; do not change tests without spec change approval.
• Signature discipline (anti-vagueness): explicit types/invariants; Optional/None requires justification and handling strategy.
• Alternatives requirement: evaluate >=2 alternatives + "do nothing".
• Uncertainty protocol: ask [blocking] questions when unsure; do not proceed.

<completion_criteria>

• design-docs/ exists with README + active/ + agents/ subdirectories
• name of ${DESIGN_DOC_SKILL_NAME} confirmed with user.
• ${DESIGN_DOC_SKILL_NAME} skill exists at .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/
• Templates bundled at .{AGENT_NAME}/skills/{DESIGN_DOC_SKILL_NAME}/templates/
• base.md includes guardrails and human-readable subtasks section format
• At least one project-specific template variant exists with repo-specific test commands
• Skill workflow: human-readable subtasks in main doc -> user approval -> XML in design-docs/agents/ </completion_criteria>

<decision_points>

• No existing config directory -> create one for current agent
• Multiple templates requested -> generate each as separate variant file
• User wants immediate write vs preview -> honor preference from question 6 </decision_points>

<failure_modes>

• Scan finds nothing -> ask user for stack/commands explicitly
• Conflicting conventions -> note in template as "resolve before use"
• Overwriting existing templates -> create backup with timestamp first </failure_modes>