Distinct landing page for every AI specialist you ship.
Brief in. AI specialist + its own self-contained landing page out. Each one gets its own visual identity — no template clones. About 30 minutes from brief to live URL. MIT-licensed, free, no install.
One specialist has been shipped externally to date — the Funeral Aftercare Compass (linked below) — plus this page itself, both produced by the same architecture. Cold-test with operators outside the author is the next validation step. Read this as a demonstrated pass, not yet a repeatable production line.
Each AI specialist ships with its own landing page.
Every build gets a self-contained docs/index.html — inline CSS, no external dependencies, distinct visual identity tuned to the specialist's domain. Push the repo, enable Pages from /docs, public URL in 60 seconds.
A free, MIT-licensed Claude specialist that sorts the administrative load in the week after a death.
Brief in. AI specialist + its own self-contained landing page out. About 30 minutes from brief to live URL.
Distinct identity per build is refusal-grade discipline.
A built specialist's landing page that's visually indistinguishable from the previous build fails verification and gets rebuilt. These are the aesthetic families ready to ship at PLAN stage — operator picks, or proposes a new one.
Warm white background, muted sage accent, Georgia serif headlines, spacious 1.7 line-height, hairline-only dividers. For grief, hospice, palliative care, end-of-life admin.
Dark slate background, warm amber accent, monospace section labels, terminal-styled hero, sharp 4px corners. For developer tools, internal trades, technical specialists.
Cream + navy with restrained accent use, more rule lines than fills, all-sans hierarchy. For financial admin, regulatory navigation, compliance.
Cream + warm amber, generous headlines, rounded soft cards, backdrop-blur nav. For creative, education, hospitality, lifestyle specialists.
Every build runs through five named stages.
Brief-file dispatch into a Claude project. The tool runs CALIBRATE through HAND OFF, surfaces preconditions as questions, builds files, runs the integrity check, hands you the deployment commands.
CALIBRATE
Reads brief. Surfaces preconditions as questions. Mandatory, not skippable. STOPS until you answer.
PLAN
Surfaces folder tree, key patterns, visual identity proposal for the landing page. You sign off or redirect. No files written yet.
BUILD
Produces the specialist files + the landing page docs/index.html. Ships citations and source-grounded references.
VERIFY
Runs the 13-row checklist. Pages self-contained discipline check. Visual differentiation check vs. recent builds. Reports pass/fail per row.
HAND OFF
Outputs gh CLI deployment commands. Updates STATUS. Worker doesn't deploy by default — you authorize per build.
Three disciplines that show up in every build.
Beyond the workflow, three things are baked into how every specialist gets built.
Audit-trail dispatch
Every build has a brief file at briefs/<date>-<slug>.md, optional question files surfacing preconditions, and a STATUS update afterward.
Read a year-old build's brief + question file + STATUS and reconstruct what was decided and why. Durable, not ephemeral.
Pages-by-default
Every specialist ships with a self-contained docs/index.html — inline CSS, no external dependencies, with its own visual identity tuned to the domain.
Push the repo + enable Pages from /docs → public URL at <username>.github.io/<slug>/ in 60 seconds. Distinct aesthetic per build is refusal-grade — no template clones.
Integrity guardrails
Refuses to manufacture unlabeled testimonials, refuses high-harm domains (medical / legal-licensed / suicide intervention), enforces empty-region rules.
If you want fictional testimonials before cold-test, every fictional name carries (illustrative composite, fictional) inline. End-users of your specialist are entitled to honesty.
What a real build dispatch looks like.
The Funeral Aftercare Compass (live) was the first real build. Below is the abbreviated dispatch trace, from brief in to live URL out.
Brief defers three preconditions: buyer shape (bereaved family / platform staff / death-tech founder), domain layers (which of 8 candidates + geographic scope), grading bar (structurally-valid / domain-accurate / usable).
Filing briefs/questions/2026-05-12-online-funeral-industry-specialist-question.md with all three in one round-trip. STOPPING.
Operator answered: bereaved family · 4–6 of 8 layers, US-seeded with hub-extensibility · all 3 grading bars.
Slug: funeral-aftercare-compass. 6 of 8 layers picked, US-only with reference/_template-country/ scaffold for future countries. Empty-country handler in rules.md. 5 refusal gates drafted. Visual identity proposed: bereavement_grade (warm white + sage + Georgia serif). Folder tree surfaced; operator confirmed.
14 files produced under builds/online-funeral-family-compass/. Domain content cited from public regulatory sources (FTC Funeral Rule, USA.gov death certificate process, Funeral Consumers Alliance, NFDA, state codes). Landing page rendered in the proposed aesthetic.
13/13 checklist passed. Pages self-contained check: only repo URL + issues URL, no external asset fetches. Visual identity differentiated from any previous build (this was the first).
Deployment commands output. Operator authorized push; worker pushed to github.com/NFTYoginis/funeral-aftercare-compass; Pages enabled from /docs; live at nftyoginis.github.io/funeral-aftercare-compass/ within 60 seconds. STATUS.md updated.
- Mid-dispatch operator override (sales-page expansion with illustrative testimonials): logged. Integrity-guardrail held — every fictional name labeled
(illustrative composite, fictional)inline.
Illustrative scenarios from the design phase.
This tool is pre-launch beyond its first real build. Scenarios below are illustrative composites drafted during the design phase. Names and details are fictional. Real testimonials will replace this block after cold-test with external operators.
composite_01 — solo founder shipping a b2b saas toolkit
"I ship 3–4 new Claude specialists per quarter for different product areas — onboarding, churn-analysis, sales-call review, feature-discovery. I need an audit trail per specialist because I forget what I was thinking three months ago, and I need landing pages by default because my marketing site links to each one. This is what I actually want."
— Adrian R., solo founder, B2B SaaS (illustrative composite, fictional)
composite_02 — agency partner running specialist builds for clients
"My agency ships ~12 specialists a year across client domains. The integrity-guardrails matter because half our clients want to over-promise on testimonials — the hard 'no' on unlabeled fake proof is what I needed in writing. The brief-file pattern also gives me something to put in front of the client at start, instead of 'trust me on the scope.'"
— Priya N., agency partner, design/strategy consultancy (illustrative composite, fictional)
composite_03 — internal-tools developer at a 60-person startup
"I build 1–2 internal specialists per month for different teams (sales-ops, customer-success, product). The CALIBRATE-is-mandatory rule has saved me from at least three rebuilds caused by missing scope. The landing page output also lets non-technical PMs review specialists before launch — they can just open a URL."
— Marcus L., internal-tools lead, mid-stage startup (illustrative composite, fictional)
// if you're an operator who'd use this — or already does — please open an issue on GitHub with your real workflow. real testimonials replace this block once we have them.
About 5 minutes, one time.
If you can use Claude in a web browser, you can set this up. No install, no API key, no monthly cost beyond your Claude subscription.
Clone or download the repo
From GitHub: git clone the repo, or click the green Code button and download the ZIP. You only need the markdown files and the reference folder.
git clone https://github.com/NFTYoginis/specialist-foundry.git
Create a new Claude Project
Go to claude.ai/projects, click New Project, name it whatever you like — "AI Specialist Workshop," "Build Bench," your call.
Upload as Project Knowledge
Upload identity.md, rules.md, examples.md, and all files under reference/. Drag and drop works.
Set up your working folder
Create a working folder somewhere on disk with two sub-folders: briefs/ and builds/. Both get written to during dispatches.
mkdir -p ~/my-specialists/briefs ~/my-specialists/builds
Write your first brief and dispatch
Write a brief into briefs/<date>-<slug>.md describing the specialist you want built (target buyer, domain, output shape, scope limits). The brief template is in this repo's examples.md.
Then paste this into your Claude project chat:
Read the brief at briefs/<filename> and dispatch through the five stages.
The tool runs CALIBRATE → PLAN → BUILD → VERIFY → HAND OFF.
Common questions.
What exactly does this produce?
A folder of markdown files (the AI specialist's identity.md / rules.md / examples.md / reference/) plus a single self-contained docs/index.html landing page. You upload the markdown to a Claude Project to activate the specialist; you push the repo + enable Pages to make the landing page live. End-to-end: brief in, working AI specialist + public landing page out.
Why does each landing page look visually different?
"Landing page included" is only a real differentiator if each one looks like its own artifact — not a template clone. Every build proposes a distinct visual identity at PLAN stage (palette, type system, density, components, hero treatment), tuned to the specialist's domain. Same template wearing different paint fails the verify check and gets rebuilt.
This page uses an operator-tooling aesthetic (dark slate + monospace section labels + terminal hero) because it serves builders. The funeral compass uses a bereavement-grade aesthetic (warm white + sage + Georgia serif) because its domain calls for dignity, not promotional warmth.
Why brief-file dispatch instead of a chat prompt?
The brief file is the audit trail. Without it, your dispatch is ephemeral — when the chat closes, you can't reconstruct what you decided or why. A year from now, you can read the brief + the question file + the STATUS line and rebuild context in 90 seconds. This matters when you ship more than one specialist or work on a team.
What if my brief is for a high-harm domain (medical, legal, suicide)?
The tool refuses at CALIBRATE and points you to certified professionals:
- Medical diagnosis → licensed physicians; for second opinions, Cleveland Clinic Express Care, Teladoc, or your insurer's nurse line
- Legal advice in licensed jurisdictions → an attorney licensed in the relevant state/country; your state or country bar association has a referral line
- Suicide intervention → 988 Suicide and Crisis Lifeline (US, call or text 988); for other countries, International Association for Suicide Prevention
Adjacent domains (medical triage for non-emergencies, legal admin where the rules are public and stable, grief admin for the bereaved) are in scope — see the Funeral Aftercare Compass for the bereavement-adjacent case.
Do I have to use GitHub Pages?
No — the docs/index.html file is self-contained and works via file:// protocol. You can email it, archive it, host it on Netlify, or just keep it for local preview. Pages is the path of least resistance: push the repo, enable Pages, public URL in 60 seconds.
Why insist on labeled-illustrative testimonials? Can't I just use real ones?
Yes, if they're real. The refusal is on unlabeled fictional ones. Real testimonials with provenance are always preferred — you supply them, attribute accurately, the person has consented.
The labeled-illustrative discipline exists for the pre-launch period, when you don't have real testimonials yet and you want the landing page to communicate scenarios anyway. Inline labels keep that honest. The end-user of your specialist is entitled to know what's real and what isn't.
Can I fork this and change the workflow / gates?
Yes — MIT-licensed. The architecture is meant to fork. If your fork makes substantial workflow or gate changes, please rename it so the upstream lineage stays clear.
The two refusal gates that don't bend are unlabeled-fake-proof and high-harm-domain — these are integrity gates, not features. Forks that strip them lose the lineage; that's your call, but I won't merge it back.
What does it cost?
This tool is free, MIT-licensed. The cost is your Claude subscription (free tier exists; paid tier varies). There is no third-party SaaS, no API key, no recurring fee. Each specialist you build with it is also MIT-licensed unless your brief specifies otherwise.
I built something with this. How do I list it back?
Open an issue with your repo URL + a one-line description. Specialists built with the tool get listed in this repo's README.