feat(respond): skill to enforce complete implementation

[paulnsorensen]

Summary

• New /respond skill that enforces complete implementation with zero deferrals, shortcuts, or evasion
• Based on analysis of 269 conversation logs (8651+ assistant messages) from the tern project
• Identifies and bans 7 recurring evasion patterns: "leave it for now", "I'll skip", "out of scope", "can be done later", "deferred", "would need to", "you can add"
• Includes self-check protocol — Claude scans own output for evasion language before presenting

Companion (local-only, not in PR)

Three hookify rules created locally (gitignored):

• warn-deferred-stop — completion gate on stop event
• warn-placeholder-code — catches TODO/FIXME/unimplemented!() in written code
• warn-ellipsis-code — catches // ... and "rest is similar" in written code

Test plan

• Invoke /respond on a multi-step implementation task
• Verify skill triggers on "complete this", "finish this", "no shortcuts"
• Verify evasion language in output triggers self-check
• Verify hookify rules warn on placeholder code and ellipsis comments

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR introduces a new respond skill for Claude Code that enforces complete, zero-deferral implementation behavior. It is the "antidote" to Claude's tendency to skip, defer, or hand-wave instead of writing actual code — built from an analysis of 269 real conversation logs. The skill defines 7 evasion patterns to avoid, a set of rules, and a mandatory self-check protocol before output is presented.

Changes:

• New claude/skills/respond/SKILL.md defining the skill's frontmatter, evasion pattern taxonomy, rules of engagement, self-check protocol, and usage examples.

[Copilot]

Every other user-invokable skill in this codebase has a corresponding command file in claude/commands/ that wires the slash-command to the skill (e.g., claude/commands/diff.md → claude/commands/worktree.md). The respond skill describes /respond as its primary invocation mechanism (line 134), but no claude/commands/respond.md file exists in this PR.

Without a command file, /respond will not be discoverable via /agents and cannot be invoked as a slash command. A minimal command file following the established pattern (e.g., claude/commands/diff.md) should be added alongside this skill.

[paulnsorensen]

Fixed in 82d18f5 — added claude/commands/respond.md following the same pattern as diff.md and other commands.

[Copilot]

The self-check scan list (line 115) bans "separately" as an evasion word, but Rule 6 (line 101–104) explicitly recommends asking: "Should I stub the interface and implement it, or do you want to handle auth separately?" The word "separately" appears in the prescribed example output that Claude should produce when correctly following this skill.

This same issue affects the word "handle" in "handle separately" (banned at line 115 via the "separately" ban). The Exception clause doesn't cover user-facing questions, so the self-check would incorrectly block the exact phrasing the rules recommend using.

[paulnsorensen]

Good catch — fixed in 82d18f5. Broadened the Exception clause to explicitly cover clarifying questions posed to the user, so the Rule 6 example ("Should I stub the interface and implement it, or do you want to handle auth separately?") is no longer flagged by the self-check.

[Copilot]

The claude/README.md skills table (lines 122–134) lists every skill in claude/skills/ by name and purpose. The respond skill is not listed there, leaving the skills registry out of sync with the actual directory contents. The README.md is the canonical reference for "what skills exist and what they do" — it should be updated to include this new entry.

[paulnsorensen]

Fixed in 82d18f5 — added respond to the skills table in claude/README.md. Also backfilled de-slop and tdd-assertions which were missing from the same table.

[Copilot]

The Skill Delegation table in claude/CLAUDE.md (lines 124–136) enumerates every skill and its canonical invocation. Skills de-slop and tdd-assertions are listed there, meaning Claude knows to invoke them proactively. The respond skill is not listed, so Claude won't know to use it when the described trigger conditions fire (user says "complete this", "finish this", expresses frustration about incomplete work, etc.). The table should be updated to include respond.

[paulnsorensen]

Fixed in 82d18f5 — added respond to both the Key Skills table (under Cleanup) and the Skill Delegation table in claude/CLAUDE.md.

[Copilot]

The Self-Check Protocol (line 121) bans the word "stub" from any output, but Rule 6 (line 103) explicitly suggests using "stub" as a legitimate communication pattern when asking the user about scope: "Should I stub the interface and implement it, or do you want to handle auth separately?"

The Exception clause at lines 128–130 covers quoting existing code and analysis of the codebase, but does not cover this third category: asking the user a clarifying question about how to proceed. As written, the self-check would flag the phrase from Rule 6's own example as a violation, creating an internal contradiction.

The exception should be broadened to also cover clarifying questions posed to the user, or the word "stub" should be removed from the banned-word list and replaced with more targeted patterns like "placeholder stub" or "empty stub" that only catch unimplemented placeholders rather than legitimate interface-first design discussions.

[paulnsorensen]

Fixed in 82d18f5 — same fix as the "separately" comment. Broadened the Exception clause to cover clarifying questions posed to the user, which is the third category you identified. The banned-word list stays as-is (catches implementation evasion) but the exception now explicitly permits using these words when asking the user how to proceed.