Steering Specs Generator

Steering Specs Generator

Conducts context-aware interviews to extract tacit engineering knowledge and generate agent-readable steerings. Format: Intent (Why) → Rules (What) → Practices (How) → Meta.

Supports predefined packs (8 areas) and custom topics (user-specified).

Flow overview: See flow-diagram.md for visual representation.

Prerequisites

• pack-reference.md — Topic areas and questions
• steering-template.md — Output format
• Access to an "ask user" tool and a "run subagent task" tool

Tooling Compatibility (Claude Code ↔ Codex CLI)

This skill was originally authored for Claude Code tool names (e.g. AskUserQuestion, Task). To keep it portable, treat these as capabilities and map them to your runtime:

• Ask user (blocking input)
  • Claude Code: AskUserQuestion
  • Codex CLI: request_user_input
  • If choice options aren't supported by your tool, present choices in text and ask for an index/label.
• Run subagents / parallel work
  • Claude Code: Task (+ its output/wait mechanism)
  • Codex CLI: spawn_agent + wait (+ optionally send_input to clarify) + close_agent when done
• Repo scanning / file IO
  • Claude Code: Read / Write / Glob
  • Codex CLI: exec_command for search/listing, apply_patch for edits (or your runtime’s native file tools)

In the rest of this doc:

• ASK_USER(...) means "use your environment’s ask-user tool".
• "Task agent" means "spawn a subagent and wait for its output".

Mode Selection

Interview Flow

Step 0: Check for Existing Sessions

Before configuring paths, check if sessions already exist in the repo:

1. Scan common session directories:

   • .sessions/
   • sessions/
   • docs/sessions/

2. If sessions found, present ASK_USER(...):

questions:
  - question: "Found existing session(s). Continue or start fresh?"
    header: "Session"
    options:
      - label: "Continue {sessionId}"
        description: "Resume incomplete session ({N} of {M} packs done)"
      - label: "Start new session"
        description: "Create fresh session with new ID"

3. If continuing existing session:

   • Set sessionsPath to parent directory of found session
   • Set sessionId to selected session name
   • Scan {sessionsPath}{sessionId}/ for completed pack files
   • Completed pack detection: A pack is complete if:
     • File {packId}.md exists AND
     • Contains ## Interview section with at least one ### Q entry
   • Store completedPacks[] list (pack IDs to skip)
   • Read explore-docs-conventions.md and explore-repo-context.md paths if they exist
   • Skip to Step 3 (Pack Interview Loop), filtering out completed packs

4. If starting new or no sessions found → Continue to Step 0a

Step 0a: Configure Output Paths

Ask user to confirm paths using ASK_USER(...):

questions:
  - question: "Where should steering files be saved?"
    header: "Steerings"
    options: ["./steerings/", "docs/steerings/", ".memory-bank/steerings/", "Custom"]
  - question: "Where should session files be saved?"
    header: "Sessions"
    options: ["./sessions/", ".sessions/", "docs/sessions/", "Custom"]
  - question: "Where should action items be saved?"
    header: "Backlog"
    options: ["./backlog/", "Same as steerings parent", ".backlog/", "Custom"]

Store: steeringsPath, sessionsPath, backlogPath. Create directories if needed.

Defaults: steerings/, sessions/, backlog/

Step 0b: Generate sessionId - short words id

Step 1: Define Topics

Custom topic detected (patterns: "steerings for X", "document X conventions"):

• If clear → Generate packId, packName, packType: "custom", customTopicDescription
• If broad → Clarify with ASK_USER(...) (aspects, level, scope)

No custom topic → Present 8 predefined packs as multi-select:

Step 1a: Choose Interview Mode

Ask user to select interview mode using ASK_USER(...):

questions:
  - question: "Interview mode preference?"
    header: "Mode"
    options:
      - label: "Interactive (default)"
        description: "Answer questions one-by-one with discussion"
      - label: "Fast mode"
        description: "Answer all questions at once, faster execution"

Store: interviewMode ("interactive" or "fast")

Step 2: Discovery (Parallel Explore)

Run TWO Explore agents in parallel. Each writes report to {sessionsPath}{sessionId} and returns path.

Explore #1 - Docs & Conventions:

Analyze repository for: steering files, CONVENTIONS.md, ARCHITECTURE.md,
CLAUDE.md, README conventions, eslint/prettier/tsconfig.

OUTPUT: Write to `{sessionsPath}{sessionId}/explore-docs-conventions.md`, return path.

Explore #2 - Repo Context:

Analyze: project purpose, tech stack, directory structure, main modules, patterns.

OUTPUT: Write to `{sessionsPath}{sessionId}/explore-repo-context.md`, return path.

Capture paths: docsConventionsReportPath, repoContextReportPath

Step 3: Pack Interview Loop

Filter packs: If completedPacks[] exists (from session continuation), exclude those pack IDs from processing.

Show progress when continuing:

Continuing session: {sessionId}
✅ Completed: {completedPacks.join(', ')}
⏳ Remaining: {remainingPacks.join(', ')}

When running packs in parallel:

• Launch all pack agents as background tasks
• Capture all task IDs
• Wait for all to complete (Claude: Task output/wait; Codex: wait)
• Display progress as packs finish

Spawn a Task agent per pack - run all in parallel for faster execution. Claude Code: use run_in_background: true for each Task call, then wait for all to complete.

Codex CLI: spawn_agent for each pack, capture agent IDs, then wait for completion (optionally streaming progress as each finishes).

subagent_type: "general-purpose"
description: "Interview for {packName}"
prompt: |
  Conduct interview for a single pack and save results.

  ## Pack Info
  - Pack ID: {packId}
  - Pack Name: {packName}
  - Pack Type: {packType}  # "predefined" or "custom"
  - Custom Description: {customTopicDescription}  # only if custom

  ## Context Files
  - Pack Reference: pack-reference.md
  - Repo Context: {repoContextReportPath}
  - Docs & Conventions: {docsConventionsReportPath}

  ## Output
  - Path: {sessionsPath}{sessionId}/{packId}.md

  ## Instructions

  ### 1. Read Context
  Read pack-reference.md to get question themes for this pack.

  Read repoContextReportPath and docsConventionsReportPath for grounding.
  These reports contain ALL necessary findings - do NOT run additional explores.

  ### 2. Generate Questions
  Question count:
  - Predefined: 5
  - Custom (narrow): 3-4
  - Custom (medium): 5
  - Custom (broad): 6-7

  Guidelines:
  - Ground ONLY in the provided explore reports + existing docs
  - Reference actual code: "I see X in Y file..." (from reports)
  - Ask about conventions, not roadmap
  - Offer 4 options (A/B/C/D)
  - Mark one as "⭐ Recommended" at end of description

  Pattern:
  Q: I found {finding}. How should {convention question}?
  A) {Option} — {rationale}
  B) {Option} — {rationale} ⭐ Recommended
  C) {Option} — {rationale}
  D) {Option} — {rationale}

  ### 3. Conduct Interview
  If interviewMode is "fast":
    - Present all questions in a single markdown code block
    - User responds with all answers at once (e.g., "A, B, A, C, B")
  If interviewMode is "interactive":
    - Present via `ASK_USER(...)` (max 4 questions per call if your runtime supports batching)

  ### 4. Classify Responses
  For each response, classify as:
  - CONVENTION: Timeless, future-focused ("When implementing X, do Y")
  - ACTION_ITEM: Temporal, fixes current state ("Replace X", "Fix X")

  ### 5. Extract and Save Results
  Extract rules immediately (not full Q&A) and write to {sessionsPath}{sessionId}/{packId}.md:
  - CONVENTION items → "## Conventions" section
  - ACTION_ITEM items → "## Action Items" section with severity
  - Optionally preserve raw Q&A in collapsed "## Raw Interview" section

Session structure:

{sessionsPath}
└── {sessionId}/
    ├── codebase-topology-ownership.md
    ├── architecture-design-invariants.md
    ├── {custom-pack-id}.md
    └── ...

Pack file format:

# {Pack Name}
**Pack ID:** {id}

## Conventions
- Q1: {extracted rule}
- Q2: {extracted rule}

## Action Items
- Q3: {action item with severity}

## Raw Interview (optional, for reference)
Preserve Q&A in collapsed detail if needed for debugging.

Step 4: Await Pack Interviews

Wait for all pack interview agents to complete. Each writes its results to {sessionsPath}{sessionId}/{packId}.md with classifications already included.

Note: Pack agents use existing explore reports - no additional explores are run.

Step 5: Generate Outputs

Delegate to general-purpose subagent (use the strongest available model in your runtime):

subagent_type: "general-purpose"
description: "Generate steerings and action items"
prompt: |
  Generate steerings AND action items from interview sessions.

  Session directory: {sessionsPath}{sessionId}/
  (contains one .md file per pack with classifications)

  Docs report: {docsConventionsReportPath}
  Repo report: {repoContextReportPath}
  Template: steering-template.md

  Output paths: {steeringsPath}, {backlogPath}

  Instructions:
  1. Read all pack files from session directory
  2. Extract CONVENTION items → generate steering files
  3. Extract ACTION_ITEM items → generate backlog file
  4. Generate index.md for steerings

Steering filenames:

Rule style:

• ✅ "When implementing X, do Y" / "Use X for Y" / "New features require X"
• ❌ "Proactively refactor X" / "Continue using X" / "Add X to existing code"

Action items file: {backlogPath}steering-specs-action-items.md

Severity: 🔴 CRITICAL (data loss, security) → 🟡 HIGH → 🟢 MEDIUM → 🔵 LOW → ⏸️ DEFERRED

Step 6: Present Results

Generated steerings:
- {steeringsPath}*.md (N rules, M practices each)
- {steeringsPath}index.md

Action Items: {backlogPath}steering-specs-action-items.md
- 🔴 Critical: N | 🟡 High: N | 🟢 Medium: N

Session: {sessionsPath}{sessionId}/
- {packId-1}.md
- {packId-2}.md
- ...

Review Mode (Step R)

Transform existing steerings to standard format.

R1: Locate Steerings

Auto-detect or ask: steerings/, docs/steerings/, .steerings/

R2: Analyze with Explore

subagent_type: "Explore"
prompt: |
  Analyze steering files in {steeringsPath}:
  - Structure: Intent → Rules → Practices → Meta?
  - Rules: numbered, prescriptive, no metadata?
  - Code examples: 5-15 lines?
  - File size: <200 lines?

  Report compliant vs non-compliant files with specific issues.

R3: Present Summary

✅ Compliant: file1.md, file2.md
⚠️ Need transformation:
- file3.md: Missing Intent, verbose rules
- file4.md: Code examples too long

R4: Ask Action

Options: Transform all | Transform specific | Backup and transform | Plan only | Skip

R5: Transform

Delegate to general-purpose subagent with transformation guidelines from steering-template.md.

R6: Present Results

Show changes per file, regenerate index.md.

Quick Reference

Activation keywords: steerings, tacit knowledge, interview, conventions, "steerings for X", "continue steerings"

Session continuation: Auto-detects existing sessions in .sessions/, sessions/, docs/sessions/. Offers to resume incomplete sessions, skipping completed packs.

8 Predefined Packs: Topology, Architecture, Domain, Quality, Testing, Risk, Security, Delivery

Output format: Intent (1 sentence) → Rules (numbered) → Practices (heading + explanation) → Meta

Files generated:

• {steeringsPath}*.md — Steering files
• {steeringsPath}index.md — Table of contents
• {backlogPath}steering-specs-action-items.md — Action items
• {sessionsPath}{sessionId}/{packId}.md — Interview responses per pack
• {sessionsPath}explore-*.md — Discovery reports

Reference files:

• flow-diagram.md — Visual flow
• pack-reference.md — Pack definitions and Explore prompts
• steering-template.md — Output format validation