coding-agents-prompting-flow

<description_and_purpose> Orchestrates prompt authoring/adaptation via discover -> extract+intake -> blueprint -> for_each_prompt_loop(draft -> hardening -> edit) -> simulate -> validate. Every phase logs status in coding-agents-prompting-flow-state.md, Prompt Brief carries through all phases, and final output traces to request intent. </description_and_purpose>

<workflow_phases>

1. Orchestrator must trust the system and skills; coordinate only sequence, artifacts, state, and approvals.
2. Load only references needed for the current phase.
3. Execute phases sequentially, do not skip!

1. All Rosetta prep steps MUST be FULLY completed
2. USE SKILL load-project-context, orchestration, hitl
3. MUST ALWAYS use todo tasks ledger, ASAP. Phases are sequential. Independent tasks can run in parallel.
4. MUST just-in-time load/execute/update each phase's: instructions, definitions, skills, state file.
5. Orchestrator and subagents MUST USE SKILL coding-agents-prompt-authoring.
6. No rush, take your time, MUST FOLLOW WORKFLOW ENTIRELY, no skipping.
7. This workflow MUST be used with Fable, Opus, GPT-5.5+ class models => IF NOT - DEMAND USER TO SWITCH MODEL.
8. Workflow state MUST be saved to agents/TEMP/<FEATURE>/coding-agents-prompting-flow-state file.

1. Discover project-local context, relevant prompt-family artifacts, and required references for this request.
2. Input: request + optional existing prompt. Output: Discovery Notes + Reference Set.
3. Update coding-agents-prompting-flow-state.md.
4. HITL when discovered context conflicts with user intent or critical references are missing.

<extract_intake step="2" subagent="prompt-engineer" role="Intent extractor" subagent_required_model="Claude Opus 5, GPT-5.6 Sol, Gemini 3.7 Flash">

1. Extract requirements from source prompt when present and intake clarifications from user.
2. Input: request + optional existing prompt + Discovery Notes + Reference Set. Output: Prompt Brief + Open Questions.
3. Update coding-agents-prompting-flow-state.md.
4. HITL explicit approval required for Prompt Brief.

</extract_intake>

1. Design blueprint: structure, actors, contracts, and boundaries for target prompt set.
2. Input: approved Prompt Brief. Output: Blueprint.
3. Update coding-agents-prompting-flow-state.md.
4. HITL approval when architecture or tradeoffs are ambiguous.

<for_each_prompt_loop step="4" subagent="prompt-engineer" role="Prompt Author" subagent_required_model="Claude Opus 5, GPT-5.6 Sol, Gemini 3.7 Flash">

1. Scope: draft target prompts. Subagents must draft one individual prompt file or surgical changes to one file at-a-time: it must think thoroughly about each file first, then it provides the full structure of the file with specific grounded points each section should contain or change, then it validates with schemas requirements/suggestions, and only then it can generate actual file. Repeat for each file. ALSO draft means fully ready for review prompt, just not reviewed and approved! Aim newly created files to have small extensive coverage, clear concerns, while each rules is very small
2. Input: approved Prompt Brief + Blueprint. Output: Draft Prompt Set + optional change-log.md in FEATURE PLAN folder.
3. Update coding-agents-prompting-flow-state.md.
4. HITL when loop stalls, conflicts appear, or intent becomes unclear.

</for_each_prompt_loop>

<for_each_prompt_loop step="5" subagent="prompt-engineer" role="Prompt reviewer and hardening" subagent_required_model="Claude Opus 5, GPT-5.6 Sol, Gemini 3.7 Flash">

1. Run loop for each target prompt: hardening -> edit until pass criteria or HITL decision.
2. This is automated review by subagent, this is not HITL review!
3. Input: approved Prompt Brief + Blueprint + Draft Prompt Set + prompt family information. Output: Prompt Set + optional change-log.md in FEATURE PLAN folder.
4. Update coding-agents-prompting-flow-state.md.
5. HITL when loop stalls, conflicts appear, or intent becomes unclear.