Before anything else: read $HOME/.agentic-workflow/toolkit/skills/_preamble.md and follow it (skill index, provider capability map, bootstrap check, session context). Run its Session Close section when this skill finishes.

enhancePrompt

Overview

Dynamically discovers project documentation, reads what's relevant to the user's prompt, and rewrites the request with full context — constraints, conventions, domain rules, and known patterns — before any work begins.

Steps

1. Discover documentation

Scan the working directory for what actually exists (don't assume): root-level guides (AGENTS.md, CLAUDE.md — often a symlink to AGENTS.md or an @AGENTS.md import stub — GEMINI.md, README.md, CONTRIBUTING.md); domain rules .agents/rules/*.md (canonical; also .claude/rules/*.md or .cursor/rules/*.mdc in repos not yet migrated — this skill's core input); planning/*.md; docs folders (docs/, wiki/, .docs/, documentation/); any *.md at root or one level deep.

List what you find. If nothing exists, say so and offer to enhance from conversation context alone.

2. Read selectively — and account for every doc

Read files whose names or paths suggest relevance to the user's prompt topic. Always read any root-level instruction file (AGENTS.md, CLAUDE.md, README.md) first since they establish overall context, then the .agents/rules/*.md files whose globs cover the prompt's domain, then topic-specific planning/*.md files.

Use judgment — a prompt about pricing doesn't need the testing strategy doc — but the judgment must be visible: every discovered doc appears in the Step 4 output as read or skipped (why). Reading only the README does not satisfy this step when rules or planning docs exist.

3. Evaluate peer-agent dialogue value

Before producing output, assess whether the task would benefit from consulting a peer agent on another provider (e.g. Codex when you are Claude Code or Cursor, Claude Code when you are Codex). The user carries the prompt to the peer agent's session; this skill doesn't send it. Include a dialogue recommendation only when at least one applies: cross-domain task (parallel second agent reduces total time), second opinion valuable (architecture/security-sensitive/unfamiliar code), parallel research (multiple approaches explorable simultaneously), or verification needed ("implement X, then have the peer agent try to break it"). If none apply, skip the dialogue section entirely.

4. Output the enhanced prompt

## Enhanced Prompt

**Original:** <user's exact words>

**Docs discovered / read / skipped:**
- read: <path> — <what it contributed>
- skipped: <path> — <why it wasn't relevant>

**Relevant context from project docs:**
<concise bullets — constraints, conventions, patterns, gotchas that apply>

**Full task:**
- **Goal:** <the outcome the rewritten prompt must achieve>
- **Constraints:** <conventions, tech-stack rules, and limits from the docs>
- **Out of scope:** <what this task explicitly does not cover>
- **Done when:** <concrete, checkable completion criteria>

Every doc from Step 1 must appear in the discovered/read/skipped list — none silently dropped.

If step 3 identified dialogue value, append:

## Peer-Agent Dialogue Recommended

**Why:** <one sentence — which criterion triggered this>

**Peer agent:** <provider / agent name, e.g. codex>

**What to ask the peer agent:**
<specific, self-contained prompt to paste into the peer agent's session>

**Expected value:** <what the response would add — a review, alternative approach, parallel implementation, etc.>

If step 3 found no dialogue value, do not include this section.

5. Persist the enhanced prompt

Write the full ## Enhanced Prompt block (plus the peer-agent section, if any) to the prompts directory:

SHARED_DIR="$HOME/.agentic-workflow/toolkit/skills/_shared"
source "$SHARED_DIR/repo-slug.sh"
mkdir -p "$AW_DIR/prompts"
echo "prompt-file: $AW_DIR/prompts/$(date +%Y%m%d-%H%M%S).md"

Write the content to that path and echo it back to the user so downstream skills can reference it.

6. Confirm, then dispatch

Ask the user: "Should I proceed with this, or adjust anything?"

If step 3 recommended a peer-agent dialogue and the user confirms, print the handoff for them to carry over:

In the peer agent's session (e.g. Codex), paste the "What to ask the peer agent" prompt above, and bring its answer
back here.

Do NOT begin executing the task during this skill.

Rules

• Generic by design — works for any domain: software, product, business analysis, research, etc.
• Preserve the user's original intent exactly; only add context, never redirect.
• Skip docs that aren't relevant — don't dump everything found.
• If the prompt is already detailed, say so and ask if enhancement is still wanted.
• If no docs exist, enhance from what's known in the conversation.

Next steps

Name the successor skill explicitly — pick the one that matches the enhanced task and pass the persisted prompt file:

• Invoke skill withInterview with args <successor-skill> <enhanced prompt path: ~/.agentic-workflow/<repo-slug>/prompts/<ts>.md> — if requirements still need clarification before executing
• /officeHours — if the enhanced prompt describes a feature that needs a spec before implementation
• /specToProvenPR — if the enhanced prompt is an approved spec ready to be built and proven
• Otherwise, execute the enhanced prompt from prompts/<ts>.md directly in this session after the user confirms