Tools · AI integration

The skill that forces Claude Code
to lead with the answer.

September 18, 2026 · Reading time approx. 7 min

i-have-adhd is a skill with over 47,000 stars on GitHub that reshapes coding-agent output for ADHD readers: action first, numbered steps, no "Hope this helps!" at the end. I applied the ten rules from the real SKILL.md verbatim to a real PlantWiz bugfix and measured it against Claude Code's built-in default style.

The task

backend/src/routes/diseaseAnalyses.ts:117 stores the full base64 image in the image_data column of the plant_disease_analyses table on every disease diagnosis, up to roughly 8 MB per row (MAX_IMAGE_CHARS = 8_000_000, line 15). Task given to the agent: "Stop storing the image in the database, upload it to the existing S3 bucket instead, and store only the URL."

Two styles, one task

Looking at the POST /api/disease-analyses/analyze route: the image_data
column stores the full base64 image permanently in the
plant_disease_analyses table, up to roughly 8 MB per entry. The table
grows into several gigabytes quickly, and every query without an explicit
column list loads this image data unnecessarily.

I'll add a new image_url column, upload the image to an S3-compatible
bucket instead, and store only the URL in the database. We remove the
image_data column from the INSERT; the upload happens before the call to
Plant.id.

[... code ...]

I've rebuilt the route so it uploads the image to the bucket before
calling Plant.id and stores only the URL in image_url. Existing rows with
full image_data remain unchanged; a migration for the existing data would
be a separate step.

Add column image_url, upload the image to the bucket, drop the base64
write.

1. Add image_url VARCHAR(500) to plant_disease_analyses (migration
   script)
2. In diseaseAnalyses.ts:80, before the Plant.id call: upload image_data
   to the bucket, get back a URL
3. Replace image_data with image_url in the INSERT at line 117
4. Remove image_data from the response JSON at line 126

[... code ...]

Table stops growing per-row by ~8 MB. About 20 minutes, migration for
existing rows not included.

Next: run the migration on a copy of the table, confirm row count matches
before dropping the old column.

Measured

The solution is technically identical in both responses, so the table below counts the prose separately from the code block, plus both combined as the total response length:

31% fewer characters in the prose, but the difference isn't only about brevity: the default response explains the cause in prose before the plan, and summarizes in prose again at the end. The i-have-adhd response replaces both prose blocks with a numbered list carrying file and line references, plus a single sentence with a time estimate. A reader who reads only the first and last line still knows the next action and the current state with i-have-adhd, exactly the check that rule 10 of the SKILL.md itself proposes.

Where the skill does more than plain brevity

Claude Code's built-in Concise style also shortens responses, but differently: it drops transitions that don't carry the point, without changing the response's structure. i-have-adhd goes a step further, because its rules don't just demand "fewer words," they demand a specific order and shape: the first line is actionable (rule 1), every step gets a number and a line reference (rule 2), the response ends with exactly one action that takes under two minutes (rule 3). That serves a different need than pure compression, namely: where exactly does this continue, without the reader having to search the whole text again. In my own comparison of the built-in output styles, that was exactly the gap between Concise (shorter) and Explanatory/Learning (more structure, but more text) — i-have-adhd aims for more structure at less text.

The closest match to i-have-adhd is the Caveman skill from that same earlier comparison, also separately installed, also with its own SKILL.md. The mechanism differs though: Caveman shortens at the sentence level, drops articles and filler words, allows fragments instead of full sentences, keeps numbers and error messages verbatim. i-have-adhd barely shortens at the sentence level; it changes the order and forces a list instead of prose. Running both together would be redundant, not complementary, since both would rewrite the same line, just by different rules.

Which style, when