A community platform for discovering, refining and deploying high-quality AI prompts. Scored across six independent dimensions, versioned like code, and tuned per model.
by @promptsmith · catches the bugs your reviewer misses
No Llama variant yet.
Be the first to port it and earn contribution credit.
Port this prompt[ 01 / 12 ] The problem
Today it lives in Reddit threads, Discord servers and GitHub gists. No quality control, no version tracking, no model specificity.
No trusted central repository. Users rely on word-of-mouth and aggregators with no scoring, freshness signal, or model specificity.
Prompts are shared as finished artifacts. No version history, no collaborative iteration, no way to track how a prompt improved.
Upvotes don’t capture what matters. A prompt can be popular but unclear, or effective on one model and brittle on another.
No shared workspace to test, benchmark and compare outputs before publishing. Solo iteration is the default mode.
Core insight
The ceiling of any AI model is set by the quality of the prompts it receives.
Everyone is racing to swap models. The cheaper win is upstream: write the instruction properly, prove it works, and keep the version that did. Promptdec addresses this at the source.
[ 02 / 12 ] Scoring system
A single upvote count is too coarse to describe prompt quality. Every dimension is rated separately, and all six stay visible, so you can weight what matters for your use case.
Does it consistently produce the desired output type?
How original is the framing or the technique?
Easy to understand, adapt and re-use?
Token cost and time to reach a good result.
Does it encourage factual, grounded output?
Avoids bias, toxicity and harmful output patterns.
Honest by default
Below 10 ratings per dimension we show raw vote counts, never a normalised 0 to 10 score. A score built on two votes is a false signal.
Temporal score decay
Scores fade unless re-validated. A freshness ring runs green when recently confirmed and red when stale. When a provider ships a new model, every prompt tagged to it enters the re-validation queue automatically.
Polarizing, not averaged
High rating variance gets a Polarizing flag instead of a mediocre middle score. Those prompts are often the most interesting: genuinely novel, or subtly broken.
[ 03 / 12 ] Per-model variants
GPT-4o responds to direct instruction. Claude performs better with role-framed context. Midjourney needs a different grammar entirely. Variants share one identity and nothing else.
No Llama variant yet.
Be the first to port it and earn contribution credit.
Port this promptSwitch a model. Body, scores and validation all change with it.
Variant rules
| Attribute | Scope |
|---|---|
| Title & description | Shared |
| Category & taxonomy tags | Shared |
| Original author & attribution | Shared |
| Prompt body text | Per-variant |
| Version history & diffs | Per-variant |
| Scoring dimensions | Per-variant |
| Community validation | Per-variant |
| Output benchmarks | Per-variant |
| Freshness signal | Per-variant |
Model Migration Assistant. Save a GPT-4o prompt while you mostly run Claude and Promptdec surfaces the community-tested port, or invites you to write it. You’re never left with a prompt that silently underperforms.
In the product
[ 04 / 12 ] The platform
Differentiator
Curated sequences of linked prompts for multi-step work. A Journey is not a collection. Each step consumes the previous output.
Ordered nodes with declared input → output relationships.
Full version history on every prompt. Forks are first-class, not copies. The lineage view shows the original, every suggested improvement, what was accepted, and the canonical version today.
Git-inspired. Zero git jargon.
A sandboxed staging environment. Compare outputs side by side, submit alternative phrasings as candidates, and let the author choose what ships. A PR review process, for prompts.
Pro tier · live inference via your own API key.
Parameterised slots with declared types keep the engineer’s structure intact while dropping the barrier for everyone else.
[TOPIC][TONE: formal|casual][FORMAT: bullets|prose]
Categories are specific, not broad. Semantic tagging reads intent, style and expected output, so search surfaces contextually similar prompts rather than keyword matches.
Users submit real AI outputs into a monitored thread on every prompt. Outputs are collected, discussed and rated across the six dimensions, so scores stay tied to actual results rather than vibes.
Flagging + community moderation ship before launch.
Also on the roadmap
[ 05 / 12 ] Inside the product
Same object, five surfaces: found, forked, diffed, proven, credited. Nothing here is a mockup made for a landing page. These are the build screens.
Discovery
Ten categories, sixty-two niches, three levels deep. Rising niches surface what the community is pushing this week, before it ever reaches a leaderboard.
Lineage
Every fork keeps a line back to the original. Attribution is automatic, the lineage panel shows who improved what, and the author is notified rather than erased.
Evolution
Compare any two versions, unified or split. The score line moves with the edits, so you can see which rewrite actually earned the +0.5 and which one just churned words.
Evidence
Attach the run and the code that produced it. Comments carry ratings, moderation is visible in the open, and the leaderboard rewards the people sharing results.
Reputation
Prompts published, deploys they earned, average score received and the categories you actually work in. Depth and consistency on one page, instead of a post count.
[ 06 / 12 ] Adaptive UX
Complexity always exists in the data layer. What changes is how much of it is surfaced by default. Advanced users don’t get more features, they get less hidden. Mode is a display preference, never a paywall.
| Feature | Explorer | Creator | Engineer |
|---|---|---|---|
| Aggregate quality score | Visible | Visible | Visible |
| Individual scoring dimensions | Hidden | Visible | Visible |
| Model variant tabs | Active only | All tabs | All tabs |
| Version history timeline | Hidden | Collapsed | Expanded |
| Fork & remix controls | Hidden | Visible | Visible |
| Scoring lenses | Hidden | Hidden | Visible |
| Robustness score | Hidden | Detail page | On cards |
| Prompt Labs | Hidden | Accessible | Prominent |
| API key connector | Hidden | In settings | Prominent |
| Keyboard shortcuts | Not shown | Not shown | Documented |
| Contribution heatmap | Hidden | Shown | Shown |
| Submission form | Single field | Slot builder | Slots + linting |
Behaviour-based promotion · submit 3 prompts, fork one, or open a version history and Promptdec offers the next mode. One prompt per threshold event, no repeated popups.
[ 07 / 12 ] Who it’s for
Wants ready-to-use prompts that just work. Discovers by search and browse. Values zero-friction copy-paste and clear output expectations.
Lands in ExplorerWriters, artists and developers using AI daily. Cares about model-specific nuance and wants to build reputation through quality contributions.
Lands in CreatorProfessional focus. Needs version control, benchmark comparison, team workspaces and API access to wire Promptdec into an existing pipeline.
Lands in Engineerv1 focus is Types A and B. Type C capabilities (API access and team workspaces) are a v2 priority. Onboarding is not designed around them, or casual activation suffers.
[ 08 / 12 ] Connectivity
Promptdec is not another chat interface. The connectors sit alongside the tools you already use. The value is making your library reachable from wherever you already are.
Surface compatibility
| Surface | Friction | Ship |
|---|---|---|
| Promptdec web app | Zero, stays on page | v1 |
| ChatGPT | One click | v1 |
| Claude.ai | One click | v1 |
| Gemini | One click | v2 |
| Third-party apps | REST API | v3 |
Connect your OpenAI, Anthropic or Gemini key and fire any saved prompt straight at the model, streamed back inline. No tab switching, no clipboard.
v1 · P0A Promptdec panel injected into ChatGPT and Claude.ai. One click inserts the prompt text with the model-correct variant already selected.
v1 beta · P1Favourites sync in real time across the web app, the extension and anything wired in through the API. One organised library, everywhere.
v1 · P1A floating command palette over your saved prompts from anywhere in the browser. Filter by model, category or free text. Enter to insert.
v2 · P2[ 09 / 12 ] Contribution
Depth and consistency, not volume of posts. Every profile carries the full picture: what you shipped, how it scored, how fairly you rate others.
Thank You economy
Send a note when a prompt genuinely helped. Authors with the most meaningful acknowledgements get the weekly digest, not just the most upvoted. Rewards mentorship over virality.
Prompt Rescue
A queue of high-potential prompts that are underperforming: good structure, low rating count, few views. Adopt one, test it, resubmit, and earn co-author credit when it crosses the threshold.
Ghost Mode
Submit anonymously. Your name is revealed automatically once the prompt clears a rating threshold, so a first contribution never costs you reputation.
[ 10 / 12 ] Roadmap
Auth, submission, three-level taxonomy, multi-dimensional scoring, model tagging, full-text search, public profiles.
200 seeded prompts at launchVersion history, forking, comment threads, output benchmarking, Journeys, templates, semantic search. Pro tier with Labs.
2,500 registered usersLive LLM testing in Labs (OpenAI + Anthropic), Chrome extension beta, expert curated collections.
6,000 registered usersTeam workspaces, public REST API, image model integrations, marketplace for premium curated packs.
10,000 registered users[ 11 / 12 ] Pricing
The free tier has to be genuinely useful. That is what builds the network effect the scores depend on.
Tier 1
$0forever
Tier 2 · most popular
$12/ month · $99 yearly
Tier 3
$49/ month per workspace
Where we are · pre-launch targets, not results
Our north star is the weekly active prompt-using user: someone who either submits a prompt or copies and deploys one. Nothing here has happened yet. These are the numbers we are building toward, and we would rather publish the target than dress up a launch that has not shipped.
[ 12 / 12 ] Questions
drag, scroll or use the arrows
An upvote is one bit of information from someone who may never have run the prompt. A Promptdec score is six independent dimensions, each rated separately, and we refuse to show a normalised number until a dimension has at least ten ratings. Below that you see raw counts, because a 9.4 built on two votes is a lie with a decimal point.
ScoringEvery prompt tagged to that provider enters the re-validation queue automatically, and its freshness ring starts ageing in public. Variants are independent, so a prompt can be confirmed on Claude and stale on GPT at the same time, which is the honest answer rather than an inconvenience to hide.
FreshnessNo. A fork keeps a permanent line back to the original: your handle stays on the lineage, you get notified, and the fork page shows what changed and why. Rescue an underperforming prompt and you earn co-author credit on it once it crosses the rating threshold.
AttributionUsable. Browsing, search, voting on all six dimensions, ten submissions a month and a public contribution profile are free forever. We gate the power tools (semantic search, Labs, Journeys, version control) and never the content. The network effect the scores depend on cannot be built behind a paywall.
PricingOnly when you ask. Prompt Labs runs live inference through your API key, on the model you pick, and nothing is fired at a provider in the background. Benchmarks on a prompt page are outputs people chose to post into the thread, with the model and date attached.
LabsBecause the failures are different shapes. A prompt can be devastatingly effective and quietly unsafe. It can be beautifully clear and burn four times the tokens it needs. Averaging those into a single 8.4 destroys exactly the information you needed to choose. All six stay visible, and you weight them yourself.
ScoringNo, and we would rather say so than dress up a launch. Promptdec is pre-launch. The screens on this page are the build, the numbers are targets, and the first two hundred prompts set the baseline every future score is measured against.
Join the waitlist Status
Pre-launch · founding contributors
The first 200 prompts set the baseline every future score is measured against. Founding contributors are the ones who write them.
One email when the beta opens. Nothing else, ever.
What happens next
One confirmation, then silence until the beta is actually open.
Explorer, Creator or Engineer. Five questions, changeable any time in settings.
Submit one that works, rate a few others, and the first scores start to mean something.
Free tier stays free. No card, no trial timer.