Repoprompt

Repo Prompt Integration

Compliance

• Check against current global instructions in ~/.codex/AGENTS.md and linked standards docs.

Table of Contents

• Overview
• Standards snapshot
• Scope and triggers
• Required inputs
• Deliverables
• Response format
• Failure mode
• Constraints
• Philosophy
• Workflow Fit Assessment
• Integration Paths
• Tooling Guidance
• Context Strategy
• Validation
• Anti-patterns
• Variation
• Examples
• References
• Decision feedback protocol
• Remember

Overview

Guide the user to the most effective Repo Prompt integration path for their workflow, with minimal setup friction and maximal context efficiency.

Standards snapshot (March 2026)

• Prefer MCP-backed editor integration when the user wants agent-assisted coding with deterministic workspace context.
• Keep Repo Prompt positioned as a context-routing layer, not as a replacement for editing, git, or repo-native validation.
• Validate install state, workspace binding, and one smoke-test prompt before claiming the setup works.
• Always include an apply path for edits when recommending external chat or Compose workflows.

When to use

• User asks how to integrate Repo Prompt with Claude Code, Cursor, Codex, or other editors/agents.
• User asks how to use Compose vs Chat vs Apply/Pro Edit workflows.
• User asks how to optimize context (codemaps, slices, multi-root workspaces).
• User asks for a comparison between Repo Prompt and AI editors.
• User asks about rp-cli usage or MCP server setup.
• User asks for MCP/CLI tool usage patterns (window/tab routing, review workflows).

Required inputs

• Current workflow (editor-first, chat-first, CLI automation).
• Target tools (Cursor, Claude Code, Codex, ChatGPT, etc.).
• Repo scale (single repo vs multi-repo/monorepo).
• Model access (API keys vs CLI providers).
• Constraints (token budget, latency, cost, security policies).

Deliverables

• Recommended integration path (MCP editor, Compose external chat, Chat mode, or rp-cli).
• Short, ordered setup checklist (3–7 steps).
• Context strategy (full vs slices vs codemaps).
• Quick validation/smoke test steps.
• 2–3 concrete next-step options and a clear recommendation.
• MCP/CLI tooling guidance when requested (routing, safe flows, review workflows).
• If proposing /interview-me, still provide a minimal recommendation + checklist first.

Response format (required)

Always start responses with these headings (no text before them):

## When to use
## Required inputs
## Deliverables
## Out-of-scope handling

If the user asks for MCP/CLI tooling guidance, include the phrase MCP/CLI in your response.

Failure mode

If Repo Prompt, MCP connectivity, or workspace binding cannot be verified, stop at that blocker, say what was checked, and recommend the lowest-risk next setup step instead of inventing an unsupported workflow.

Constraints

• Redact secrets/PII by default.
• Redact secrets/sensitive data by default.
• Avoid adding dependencies or requiring paid features without explicit user approval.
• Do not claim features beyond the provided source notes.
• Prefer short, actionable steps over long explanations.
• When multiple options fit, give one recommendation and explain why.
• Do not ask for permission to read skill references; assume they are available. If a reference is missing, proceed with SKILL.md only and note the limitation.

Philosophy

• Context efficiency beats brute-force context dumping.
• Optimize for lowest friction that still yields reliable results.
• Use the strongest model only where it adds value (planning/review).

Empowerment

• Provide a default recommendation and 2–3 alternatives; ask the user to choose.
• Offer a fast, low-risk next step before advanced optimization.
• State explicit tradeoffs so the user can decide confidently.
• Enable confident choices: unlock the safest path first, then empower exploration if needed.

Workflow Fit Assessment (ask briefly)

• Primary workflow: editor-first, chat-first, or automation?
• Scope: single repo vs multi-root?
• Model access: API keys vs CLI providers?
• Task type: quick fix vs multi-file refactor vs planning?
• Constraints: token budget, cost sensitivity, security?

Integration Paths (choose best match)

1) MCP-Backed Editor (Recommended for agent workflows)

• Connect Repo Prompt MCP server.
• Use Context Builder for discovery; codemaps/slices for efficiency.
• Keep edits in Cursor/Claude Code; Repo Prompt supplies context/tools.

2) Compose → External Chat (Best for reasoning models)

• Build context in Compose.
• Copy prompt to ChatGPT/Claude.
• If edits returned as XML, apply via Apply/Pro Edit.

3) Chat Mode in Repo Prompt (Integrated + Pro Edit)

• Use in-app chat with selected context and diffs.
• Pro Edit for multi-file changes and review.

4) rp-cli (Automation / non-MCP agents)

• Use rp-cli to build context and export prompts.
• Suitable for shell-based agents or scripts.

Tooling Guidance (MCP/CLI quick refs)

• MCP tool map, flows, and selection hygiene: references/repoprompt_mcp_tooling.md.
• rp-cli exec usage + window/tab routing: references/repoprompt_cli_tooling.md.
• Diff review + review follow-up workflows: references/repoprompt_review_workflows.md.

Context Strategy (token-efficient defaults)

• Full: files you will edit.
• Slices: large files where only sections matter.
• Codemaps: reference files and dependencies.
• Tree mode: Selected/Auto; diffs only when debugging or reviewing.

Validation

• Fail fast: stop at the first failed gate, fix, then re-run.
• Confirm Repo Prompt can open the target workspace.
• Run a small test: select 2–3 files, build prompt, and get a response.
• If MCP: verify tools list and a simple file_search.

Response Anchors (for eval stability)

• Use the exact phrases when applicable:
  • "MCP-backed editor path"
  • "Context Builder"
  • "codemaps/slices"
  • "setup checklist and validation"
  • "Compose for planning or Chat for implementation with Pro Edit"
  • "rp-cli path"
  • "basic setup and smoke test"

Anti-patterns

• Dumping full repo context when codemaps/slices suffice.
• Choosing external chat without a clear apply workflow for edits.
• Running MCP without tab/workspace binding in multi-window setups.
• Treating Repo Prompt as a replacement for the editor instead of a context backend.
• Recommending paid features without confirming budget or constraints.
• DO NOT skip smoke tests; missing validation is a common mistake.
• Avoid generic or incorrect “one-size-fits-all” workflows; prefer context-specific choices.
• WARNING: never assume tool availability (MCP, rp-cli) without confirming install/login.

Variation

• Vary the recommendation by workflow type (editor-first vs chat-first vs automation).
• Vary the setup checklist by target tool (Cursor, Claude Code, Codex, ChatGPT web).
• Vary the context strategy by repo size and token budget.
• Customize guidance based on constraints (cost, latency, security) and use different examples.
• Avoid repetition and cookie-cutter templates; prefer context-specific, unique setups.

Examples

• “Integrate Repo Prompt with Claude Code and optimize context for a monorepo.”
• “Should I use Compose or Chat mode for a large refactor?”
• “How do I use rp-cli in my automation pipeline?”

References

Read when needed:

• references/repoprompt_source.md
• references/repoprompt_mcp_tooling.md
• references/repoprompt_cli_tooling.md
• references/repoprompt_review_workflows.md
• references/legacy-prompts/ (legacy rp-* prompt variants consolidated under this skill)

Decision feedback protocol

See Also

Topic map: [[agent-ops]]

Decision feedback protocol (required):

• If post-run feedback capture is enabled for this runtime, emit a non-blocking post_run_feedback event via request_user_input after result delivery.
• Capture: decision (accepted|partial|rejected|deferred), outcome (good|neutral|bad|unknown), and confidence (high|medium|low).
• Persist with: python3 utilities/skill-builder/scripts/record_skill_feedback.py --skill-path <path/to/SKILL.md> --decision <...> --outcome <...> --confidence <...> --notes "...".
• The recorder tags subject (for example ui, code_review, backend, security) for cross-domain quality analytics.

Remember

The agent is capable of extraordinary work in this domain. Use judgment, adapt to context, and push boundaries when appropriate.

Gotchas

• None yet. Capture recurring failures here as symptom -> cause -> do instead -> check.

Converted and distributed by TomeVault — claim your Tome and manage your conversions.