What the Pre-Send Check in `SKILL.md` Lines 130-141 Validates

The pre-send check is a quality gate that validates assistant responses satisfy strict, ADHD-friendly formatting rules by removing announcements, questions, sidebars, hedging, and idioms, then verifying first and last lines answer "What to do next?" and "What just happened?"

The i-have-adhd skill in the ayghri/i-have-adhd repository implements a specialized response style designed for users with attention differences. According to the source code in skills/i-have-adhd/SKILL.md, the pre-send check enforces this style through systematic content validation. This article explains exactly what this check validates and why each rule matters.

Pre-Send Check Location and Purpose

The pre-send check appears in the Pre-send check section of skills/i-have-adhd/SKILL.md. It operates as the final filter before any assistant response reaches the user, ensuring every output adheres to the skill's core principles: action-first, distraction-free, and concretely worded.

This check transforms verbose, meta-heavy responses into concise, executable instructions. The validation targets five specific patterns for removal, then applies two final verification tests.

Five Elements the Pre-Send Check Removes

1. Opening Announcements

The first sentence must not announce the assistant's intent.

• Forbidden patterns: "Let me...", "I'm going to...", "Here's what I'll do..."
• Why it matters: Meta-talk diverts attention from the actual task. Users need immediate access to actionable content without processing the assistant's internal state.

2. Closing Questions and Recaps

The last sentence must not ask follow-up questions or summarize completed work.

• Forbidden patterns: "Anything else?", "Does that help?", "So we've just..."
• Why it matters: Dangling prompts create decision fatigue and fragment the user's workflow. The response should end with clear forward momentum.

3. "By the Way" Sidebars

Any clause introduced with "by the way" or similar tangential markers must be stripped.

• Forbidden patterns: "By the way, you could also...", "Incidentally...", "As an aside..."
• Why it matters: Tangential information consumes limited working memory and disrupts task continuity.

4. Hedging Adverbs Without Substance

Words conveying false uncertainty must be removed unless genuine ambiguity exists.

• Forbidden patterns: perhaps, might, could possibly (when used as filler)
• Why it matters: Pretend uncertainty undermines confidence. The check preserves legitimate qualifiers but eliminates habitual hedging that obscures clear instructions.

5. Idioms and Figurative Phrases

All metaphorical language must convert to literal actions.

• Forbidden patterns: "circle back", "get the ball rolling", "on the same page", "low-hanging fruit"
• Why it matters: Abstract phrases require additional cognitive processing. Concrete language delivers unambiguous meaning.

Two Final Verification Tests

After removing the five element types, the check validates that a reader skimming only the first and last lines can answer:

1. "What to do next?" — The opening line must contain the immediate concrete action.
2. "What just happened?" — The closing line must confirm the completed step and point forward.

Failure on either test blocks the response until reformatted.

Before and After: A Practical Example

The following demonstrates the transformation applied by the pre-send check as implemented in the i-have-adhd skill:

Before check (invalid):

Let me explain the next steps. First, open `src/auth.ts`. Then, add the token logic. By the way, you might want to run the tests later. Anything else?

After check (validated):

Open `src/auth.ts` and add the token logic.
Step 1 of 2 done: file opened. Next: run `npm test` to verify.

Changes applied:

• Removed opening announcement: "Let me explain the next steps."
• Deleted "by the way" sidebar about tests
• Stripped hedging adverb "might"
• Eliminated closing question "Anything else?"
• Added explicit step numbering and forward pointer in final line

Where This Validation Lives in the Source

File	Relevance
skills/i-have-adhd/SKILL.md	Defines the complete pre-send check specification including all five removal rules and two verification tests

The pre-send check exists solely within this skill manifest. No additional code files implement this logic—the validation rules are declarative guidelines that constrain response generation.

Summary

• The pre-send check in SKILL.md validates assistant responses through five removal rules and two verification tests
• It strips opening announcements, closing questions, tangential sidebars, empty hedging, and idioms
• It confirms first-line actionability and last-line closure for skim readers
• This validation ensures ADHD-optimized responses: concrete, compact, and immediately actionable

Frequently Asked Questions

What happens if a response fails the pre-send check?

The response does not reach the user. The system regenerates or reformats content until it satisfies all validation rules. The check serves as a hard gate, not a suggestion.

Why target specifically "by the way" rather than all digressions?

"By the way" functions as a predictable linguistic marker for off-topic inserts. The skill targets this pattern because it reliably signals working-memory disruptions that derail task-focused attention, making it efficient to catch and remove automatically.

Can the pre-send check be disabled or customized?

The SKILL.md file defines fixed rules for this skill. As implemented in ayghri/i-have-adhd, the validation is integral to the skill's identity. Users seeking different response styles would activate alternative skills rather than modify this check.

How does this differ from standard prompt engineering?

Standard prompts request formatting; the pre-send check enforces it through structural validation. Rather than hoping the model complies, the check programmatically verifies compliance using explicit, testable criteria applied to every output.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}