A Claude Code plugin for answering headache & migraine questions that refuses to make things up.
Animated demo (loops). Not seeing motion — e.g. in email? Here's the static screenshot.
Headache Helper answers your questions only from a curated local wiki of peer-reviewed papers, clinical guidelines, and neurology reference texts that you approve — with inline citations, source-conflict resolution, and a hard no-fabrication rule. If the wiki doesn't cover something, it says so and offers to research it. It never answers from the model's general knowledge.
It ships two skills:
/headache-helper— asks the wiki and answers (read-only, safe)./headache-ingest— fetches candidate sources from a trusted whitelist, drafts them, and adds them to the wiki only after you approve (the write path, gated).
⚠️ Not medical advice. This is a research and knowledge-management tool. Its answers are summaries of the sources you curated — not clinical guidance. Treatment decisions belong with a qualified clinician. You are responsible for the sources you ingest and how you use the output.
Large language models are useful but they hallucinate — and residual hallucination persists even in clinically-tuned models (a published example: an RLHF-tuned diagnostic model still produced incorrect outputs ~2% of the time). For anything health-related, a plausible-but-wrong answer is worse than no answer.
Headache Helper's premise is simple: grounding, not model quality, is what makes an answer trustworthy. So it inverts the usual flow — instead of asking the model what it "knows," it asks a curated corpus of real, cited, maintainer-approved sources and forbids the model from going beyond them.
- Grounded-only answers. Every claim traces to an approved source file, cited inline as
[title, year]. No source, no claim. - Never fabricates or hedges. If the wiki doesn't cover it, the skill says "the wiki doesn't cover this yet" and offers to ingest it — it won't guess, and it won't soften an ungrounded statement into "might/possibly/some sources suggest."
- Tiered trust whitelist. Only guidelines/classification (Tier 1), peer-reviewed literature (Tier 2), and reference texts/registries (Tier 3) can enter. Preprints, blogs, marketing, and social media are hard-blocked.
- Recency-wins-within-tier conflict resolution. When sources disagree, the higher evidence tier wins first; among equals, the newer one wins — and the answer says when it applied this.
- Human approval gate. Fetched sources land in
pending/and never enter the answerable wiki without your explicit approval. - Re-validation safeguard. Every answer is logged with the exact sources it used. When you ingest new material, the plugin re-checks past answers and flags any that a newer/higher-tier source has changed — so guidance can't quietly drift.
- Personal headache history. On first run the plugin sets up
HEADACHE-HISTORY.md— your private journal of questions asked, treatments tried, and attack entries. Answers use it as context about you (never as medical evidence), and it stays on your computer. - Running bibliography. A
CITATIONS.mdof every source, DOI-linked and grouped by topic, is kept current automatically. - Rejection audit trail. Every source the gate refuses is logged with a reason, so you have a durable record of what was excluded and why.
- Standing disclaimer on every answer.
you ask ─────────────► /headache-helper ──► reads headache-wiki/ ──► cited answer
│ │
└──── logs Q + sources ──► ANSWER-LOG.md ◄┘
you request a topic ──► /headache-ingest ──► whitelist gate ──► pending/ ──► YOU APPROVE ──► wiki/
│ │
REJECTED.md ◄─ off-whitelist INDEX.md + CITATIONS.md
+ re-validate ANSWER-LOG
The wiki is plain Markdown — one file per source with a metadata header (tier, year, source_type, doi, topics, approved_by, …) plus a fixed-format summary body. Because the metadata is structured, conflict resolution is a mechanical lookup rather than a judgment call. See knowledge/DESIGN.md for the full data contract.
New to all of this? This section assumes you've never installed a plugin — or Claude Code itself — before. Follow it top to bottom and you'll be fine. It takes about ten minutes.
- Claude Code is a free tool from Anthropic that lets you talk to Claude (an AI assistant) from your computer's terminal — the app that lets you type commands to your computer. You'll install it once.
- A plugin adds new abilities to Claude Code. Headache Helper is a plugin.
- A skill is one of those abilities. This plugin adds two: one that answers questions (
/headache-helper) and one that adds new research to your library (/headache-ingest). You run a skill by typing its name with a slash in front of it, like/headache-helper. - A wiki here just means a folder of research summaries on your own computer. Nothing is shared online.
You do not need to know how to code. You'll copy and paste a few commands.
- Mac: Press
Cmd + Space, typeTerminal, press Enter. A window with a text prompt opens. - Windows: Open the Start menu, type
PowerShell, press Enter. - Linux: Open your Terminal app.
Anything shown in a grey code box below is meant to be typed (or pasted) into this window, then run by pressing Enter.
If you've never used Claude Code, follow the official installation guide here: claude.com/claude-code. It walks you through installing it and signing in with your Anthropic account.
To check whether it's already installed, type this and press Enter:
claude --version- If you see a version number (like
1.2.3), you're good — go to Step 3. - If you see
command not found, Claude Code isn't installed yet. Use the guide linked above, then come back.
In your terminal, type:
claudePress Enter. The prompt changes — you're now inside Claude Code. From here on, the commands that start with a slash (/) are typed at this Claude Code prompt, not the plain terminal.
How to tell the two apart: commands starting with
/(like/plugin) go to the Claude Code prompt. Commands likecporgit(no slash) go to a plain terminal. When in doubt, this README labels each one.
At the Claude Code prompt, type these two lines, pressing Enter after each:
/plugin marketplace add GenkiTaco/headache-helper
/plugin install headache-helper
The first line tells Claude Code where to find this plugin (my public GitHub repo). The second line installs it. If it asks you to confirm, say yes. That's it — the plugin is installed.
Prefer to do it by hand? You can instead clone the repo in a plain terminal with
git clone https://github.com/GenkiTaco/headache-helper.gitand copy the two folders insideskills/into your~/.claude/skills/directory. The marketplace method above is easier and is recommended.
Good news — the plugin now does this for you. The first time you run
/headache-helperor/headache-ingestin a folder with no wiki, the skill offers to createheadache-wiki/from the shipped template and then interviews you to set up your personalHEADACHE-HISTORY.md— a private file that tracks your questions, your treatments, and your experiences over time (it stays on your computer, and it's context for answers, never a medical source). You can simply skip ahead to Usage and let the first run walk you through it.The manual steps below do the same thing by hand — kept for reference, and for when you want to control exactly where everything goes.
The plugin comes with an empty starter library in a folder called knowledge/. You need to make your own copy of it — named headache-wiki/ — inside the folder where you want to keep your research. This is where your approved sources will live.
Decide where you want your wiki to live — for most people a new folder in Documents is perfect. In a plain terminal (not the Claude Code prompt), create one and move into it. For example:
mkdir ~/Documents/headaches
cd ~/Documents/headachesmkdir makes the folder; cd ("change directory") moves you into it. From now on, this is your project folder — the place Claude Code will look for your wiki.
The marketplace installer keeps its copy of the plugin under ~/.claude/plugins/cache/. Copy the starter library from there into your project folder with this one command:
cp -R ~/.claude/plugins/cache/headache-helper/headache-helper/*/knowledge/ ./headache-wiki/cp -R means "copy a whole folder and everything inside it." The * in the middle just stands in for the version number, so this keeps working after updates. This creates a headache-wiki/ folder right where you are.
If that command can't find the folder (you see "No such file or directory"), you installed manually instead. Just find the
knowledge/folder inside wherever you cloned the repo, and copy it — e.g.cp -R /path/to/headache-helper/knowledge/ ./headache-wiki/.
You should now have this inside your project folder:
headache-wiki/
├── DESIGN.md # the rulebook for how sources are stored (don't need to read it)
├── _TEMPLATE.md # the blank form each source is filled into
├── INDEX.md # a table of every source (kept up to date for you)
├── CITATIONS.md # a running bibliography (kept up to date for you)
├── ANSWER-LOG.md # the system's record of answers, for re-validation
├── HEADACHE-HISTORY.md # YOUR file: your questions, treatments, and experiences
├── wiki/ # your approved sources live here (starts with one EXAMPLE file)
└── pending/
└── REJECTED.md # a log of sources that were rejected, and why
The wiki/ folder ships with a single placeholder called EXAMPLE-entry.md so you can see the format. Delete it so it doesn't clutter real answers:
rm headache-wiki/wiki/EXAMPLE-entry.mdYour wiki now starts empty and ready. You'll fill it up in the next section.
Where does the wiki have to be? The skills automatically look for a folder named
headache-wiki/in whatever project folder you're in when you start Claude Code. As long as you launchedclaudefrom the folder that containsheadache-wiki/, it just works. If you keep your wiki somewhere else, simply tell the skill the path when you run it (e.g. "use the wiki at ~/Documents/headaches/headache-wiki").
Every time you want to use it: open a plain terminal,
cdinto your project folder (the one that containsheadache-wiki/), and runclaudeto start Claude Code. Then type the slash-commands below at the Claude Code prompt. For example:cd ~/Documents/headaches claude
Add sources to a topic (nothing enters the wiki without your OK):
/headache-ingest preventive treatment of chronic migraine
The skill queries whitelisted databases (e.g., PubMed), rejects anything off-whitelist (logging why), drafts candidate entries into pending/, and shows you a summary table. You reply with which to approve; approved sources move into wiki/ and the index, citations, and answer-log update automatically.
Ask a question:
/headache-helper What does the evidence say about CGRP antibodies for chronic migraine prevention?
A grounded answer, every claim cited to a wiki file, with the standing disclaimer — or a plain "the wiki doesn't cover this yet" if it doesn't.
| Tier | What's allowed | Examples |
|---|---|---|
| 1 | Clinical guidelines & classification | ICHD-3, AHS, AAN, EHF, EAN, NICE |
| 2 | Peer-reviewed literature | Cochrane, PubMed/MEDLINE; Cephalalgia, Headache, J Headache Pain, Neurology, Lancet Neurology, JAMA Neurology, Brain, Nat Rev Dis Primers/Neurology |
| 3 | Reference texts & registry | Neurology textbooks, UpToDate (manual); ClinicalTrials.gov (status only) |
Hard-blocked: bioRxiv/medRxiv preprints, general web/blogs/news/Wikipedia, forums/social media, supplement & clinic marketing, anything generated from the model's training, and topically-adjacent-but-wrong-type sources. Every rejection is logged.
Edit the whitelist in knowledge/DESIGN.md (§4) and the ingest skill to match your own trust boundary.
- No source, no claim. The model may not exceed the wiki.
- Whitelist at the door. Trust is decided before a source is ever summarized.
- Recency-wins-within-tier. Conflicts resolve mechanically and transparently.
- Human approval gate. You decide what counts as authoritative.
- Log and re-validate. New evidence surfaces any past answer it changes.
Full rationale and the complete data contract: knowledge/DESIGN.md.
- Other headache types / topics: just
/headache-ingestthem — the wiki grows one approved source at a time and organizes bytopics. - A different domain entirely: the pattern (tiered whitelist → approval gate → grounded-only Q&A → re-validation) is domain-agnostic. Swap the whitelist journals/guidelines in
DESIGN.mdand the ingest skill, and you have a grounded assistant for any evidence-based field.
This repository ships only the framework — the skills, the design, and an empty wiki. It does not include anyone's curated source corpus. When you ingest sources, you create local summaries with citations for your own reference; respect the copyright and terms of the works you summarize, and cite them properly.
MIT — see LICENSE.
Built with Claude Code.