GitHub text is technical communication with reviewers and collaborators, not an internal scratchpad, proof-of-work ledger, or bot execution diary.

• Drafting vs. posting: Drafting text does not authorize posting it unless the user explicitly requested the action (e.g. "open a PR", "reply to the review", or "file an issue"). Do not post unprompted comments or create unrequested issues.
• Do not auto-create tangential issues: When discovering an unrelated bug or CI gap during a task, flag it locally or ask the user whether to file an issue. Never automatically create new issues without user authorization.

• Anti-Pattern: Turning the PR body into a narrative journal of everything thought, tried, measured, and discarded (e.g. multi-paragraph design essays, discarded alternatives, benchmark replicas, and obsolete drafting history).
• Rule: A PR body documents the current state of the branch, not the historical drafting journey. Keep it focused:
  • Summary: 1-2 bullet points on the user-visible or architectural change.
  • Key Changes: Concrete implementation details (only if non-obvious).
  • Verification: Specific tests added/passed and manual checks (no raw terminal dumps).
  • Fixes: Issue references.
  • (For small, 1-3 line changes, a 1-2 sentence description without section headings is preferred).
• ❌ (500-word essay explaining rejected library dependencies, how test harnesses were configured, and pasting 40 lines of terminal output)
• ✅
  markdown

  Fixes #72
  
  ## Summary
  
  - Initialize upstream MCP servers concurrently during startup rather than serially.
  - Bounds startup latency to the slowest single upstream server.
  
  ## Changes
  
  - Launch each configured upstream in its own goroutine and merge results deterministically by sorted server name.
  - Retain child process handles to ensure clean process termination on shutdown.
  
  ## Verification
  
  - Added unit tests for concurrent launch timing and deterministic tool ordering.
  - Local benchmark: 9 upstreams initialize in 9.7s (down from 31.1s serial).
  - `go test -race ./...` passes.

• Anti-Pattern: Responding to a 1-line review comment with a multi-paragraph compliance report detailing every rule followed, repeating diffs, and quoting the prompt.
• Rule (Proportionality): Deliver the shortest response that conveys the decision-changing facts or applied fix.
  • Routine fix: 1 line + commit SHA.
  • Technical constraint / disagreement: State the concrete invariant, constraint, or test evidence directly without entering into philosophical debates or meta-arguments.
• ❌
  text

  Ran both humanizer and comment-and-documentation-quality skills on the guide. Flagged bold-spam (§15), announcing the next point (§28), repetition (§11), and stripped em dashes under Anti-Pattern 1 in commit 7dfe483...

• ✅

  Updated in 7dfe483. Removed the MCP server comparison and tightened the wording. Build and tests pass.

• Anti-Pattern: Explaining how the model hallucinated, got confused by precedent code, or made an incorrect assumption.
• Rule: Nobody needs a post-mortem of AI confusion. Report the correction directly.
• ❌
  text

  The URL split was my mistake. I carried over the Home Assistant case where aiohttp 404s on the slash variant, and assumed the same applied here without checking. Reverted in b64a2fd...

• ✅

  Reverted the URL change and restored the test connection step in b64a2fd. The HTTP endpoint on 27123 remains the default.

• Anti-Pattern: Self-evaluating performance or narrating intent:
  • "Fair hit on all three."
  • "I can name how it got there."
  • "That is not a bot failing to converge."
  • "Deliberately minimal..."
• Rule: Delete performative self-evaluation, defensive remarks, and emotional framing. Keep comments technical, neutral, and direct. (Brief professional courtesy when acknowledging a mistake is acceptable).
• ❌

  Fair hit on all three. You are right that I argued the point instead of doing it, and that citing other guides was the wrong move.

• ✅

  Updated in 74f9283. Removed all em dashes across the guides and tightened the setup instructions.

• Anti-Pattern: Appending

  _Generated by Claude Code_

  ,

  _Reviewed by OpenAI Codex_

  , or model watermarks.
• Rule: Never include voluntary "Generated by", "Automated by", or tool attribution footers in PRs, issues, or comments.
• ❌

  Fixed the typo in 3a1b2c3.\n\n---\n_Generated by Claude Code_

• ✅

  Fixed the typo in 3a1b2c3.

• Anti-Pattern: Leaving comments on public issues or PRs that only narrate local session activity without advancing the review (e.g. dumping raw

  go test

  output or announcing that a local check ran when no action is needed).
• Rule: Include verification summaries inside the PR body or review response. Do not post standalone execution transcripts.
• ❌
  text

  I merged this branch onto main in a throwaway worktree and verified it locally:
  $ go test -race ./...
  ok github.com/dvcrn/mcpnest/internal/app 7.846s

• ✅ (Include in PR body or relevant review response instead:

  Verified locally with go test -race ./...

  )

• Anti-Pattern: Appending unrelated bugs, CI failures, or observations found in other parts of the repo into a specific PR or review thread.
• Rule: Keep every PR and review thread scoped strictly to its own subject. Flag tangential issues locally or ask the user before filing a separate issue.
• ❌

  Fixed the query parameter in 4b5c6d7. Also noticed that GitHub Actions CI configuration in .github/workflows/ci.yml is missing a Go cache step, which might slow down builds.

• ✅

  Fixed the query parameter in 4b5c6d7. Tests pass.

• Anti-Pattern: Listing hypothetical tasks that were not done ("One thing I did not do...", "I did not touch older files...", "I left that open...").
• Rule: Only mention an omission if it represents a genuine pending decision or known limitation the reviewer must evaluate.
• ❌

  I did not touch the older guides beyond the dashes because antigravity-sync-macos is 162 words and reads thin.

• ✅ (Preserving a real technical limitation)

  This change kills direct child processes on cancellation. Process-group cleanup for grandchildren is not yet handled and should be addressed in a follow-up.

1. Never fabricate claims or evidence: Do not invent test passes, benchmarks, commits, issue numbers, or verification steps. Do not claim a fix was applied unless it was committed and pushed.
2. Preserve technical uncertainty: Do not turn

   may

   into

   will

   ,

   typically

   into

   always

   , or an experimental observation into a universal guarantee.
3. Give each fact one home: Do not duplicate the same investigation across the PR body, review thread, status comment, and final summary.
4. Preserve technical jargon: Use precise technical vocabulary (e.g. race condition, mutex contention, atomic swap, WAL mode); do not over-explain basic language semantics.

1. Zero em dashes (

   —

   ), en dashes (

   –

   ), and pause double-hyphens (

   --

   ): Use colons, parentheses, or separate sentences. (Preserve literal CLI flags such as

   --verbose

   ).
2. No bold-spam or decorative emojis: Avoid bolding every second phrase or prefixing headings with

   🚀

   ,

   💡

   ,

   ✅

   .
3. No raw terminal dump logs: Summarize test and benchmark results cleanly in a single line or concise table, rather than dumping 40 lines of test output.
4. No throat-clearing / sycophancy: Cut "Great suggestion!", "You're absolutely right!", "Let's dive in".
5. Use clean titles: Keep PR and issue titles specific, concise, and free of conversational fluff. PR titles should be imperative (e.g.

   Fix dirty state warning in AI dialog

   ). Issue titles should clearly describe the problem or goal (e.g.

   AI dialog shows stale warning after save

   or

   Support concurrent upstream initialization

   ). Avoid conventional commit prefixes (

   feat:

   ,

   fix:

   ) unless required by the repo.