Built in the open

The roadmap

Rendered from the same file the work updates, one source of truth. Statuses move in the commit that ships them; course-corrections are marked honestly. Below: what's live, then the version trains coming next (about five to six features each).

Shipped

72 live · tap to open the ledger
  • live Day-one artifacts (contract, product definition, playbook, roadmap, plan)
  • live Hero-moment mechanics prototype (see-it / do-it / keep-it) Mechanics only, visual design intentionally undecided.
  • live App scaffold, SvelteKit + Cloudflare, Capacitor-ready v0.1.0.
  • live The cost gateway, one choke point for every model call v0.6.0: DEPLOYED to Supabase (project FULL AUTHORITY). Real Anthropic call + token→cost logging, per-kind quotas, $5/day ceiling. Live spend readable in the god console.
  • live Low-friction auth + the user profile + cloud sync v0.7.0: magic-link sign-in (soft-gate); on sign-in the local playbook + profile union-merge with the cloud and follow you across devices. Sign in from the Playbook.
  • live Content schema, chapters, drills, rubrics, templates as data v0.1.0, validated on every push; seeded with the Chapter 5 hero drill.
  • live THE COACH, the full guided curriculum (See it → Do it → Keep it) v6.0.0: the flagship is complete, all 9 chapters of the operator’s craft (mindset → modes → epochs → feedback → verification → memory → money → idioms → anti-patterns), 18 drills across four varied mechanics (choose / write the prompt / order the steps / spot the flaw), deterministic on-device grading, INK + Blueprint player, at /learn. Chapter map is Trust-Ladder grouped; progress derives from the Playbook; rules sync to Supabase.
  • live THE LIBRARY, your prompts + a large managed starter catalog v7: /library saves, tags, searches your prompts and snippets with [[fill-ins]], local plus cloud sync. Explore ships a big starter catalog (23 categories, 250+ prompts) you copy or save. The operator manages the whole catalog from the God Console (add, edit, hide) with no deploy, over an admin-gated catalog_items table. v9 (THE LIBRARY, REBUILT) is expanding it now.
  • live CROSS-MODEL, run any prompt in any AI, and a which-model guide v7.2.0 (Phase 3): the neutral moat. From the Library, run a saved prompt in Claude / ChatGPT / Perplexity / Gemini / Grok / Copilot; the prompt is copied first so a paste works even where the tool ignores the prefill URL. A plain "which model for the job" guide at /models. All client-side, no model cost. feature.crossModel flag.
  • live THE GRADER, paste a prompt, get an operator’s critique Live at /grader through the gateway (metering + quota + magic-link sign-in for more). Real model output with ANTHROPIC_API_KEY set. The one live-model surface. THE FORGE (v11) turns this into a compose→grade→keep workbench.
  • live THE PLAYBOOK, accumulation + cloud sync + one-file export v0.7.0: earned directions persist locally AND sync to your account (union-merge, duplicate-protected); export to MY_OPERATOR_PLAYBOOK.md. Paid export gating attaches at P4.
  • live THE GENERATORS, agent contract + kickoff + handover v5.1.0: /generators, template-based, personalized, instant, exportable. Three artifacts (CLAUDE.md agent contract, kickoff prompt, session handover) that bake in the craft (scope / plan-first / prove-it / read-back). Operator flag feature.generators. No model cost.
  • live THE GOD CONSOLE, total operator visibility + control, feature-complete by construction v7.9.0: real live control, server-driven app_config (maintenance kill-switch, feature flags, announcement banner, one-tap Broadcast to in-app + web push) written through admin-gated set_config; live spend/reach stats; operator grant/revoke. Registry self-updates and the push-gate fails if a live feature is unsteerable.
  • live THE BLUEPRINT, the v5 art language v5.0.0: technical line-art schematics (Aamir picked Blueprint). BlueprintDiagram (the blast-radius drawn as a wiring schematic, strokes draw in), bp-grid engineering-grid surfaces. Colour still only means something. Broadens into a full illustration system in v12 (CHARACTER & DELIGHT).
  • live THE PERSONAL ENGINE, meet every user in their world v4.0.0: persona intake (identity + purpose → domain) drives a translation layer (lib/world.ts) that re-skins the hero, path teaser, and drill framing to the user’s world. Signed-out gets a curiosity hook; the persona points the path at their goal. Operator flag feature.personalize.
  • live THE COMMONS, share a prompt with its provenance, curate, moderate v7.1.0 (Phase 2): un-parked the Marquee as THE COMMONS at /commons on the same directions backend. Calm feed (recent + featured, no ranking games); a shared prompt carries what it prevents; the social act is keeping one into your Library. Share from the Library goes to a moderation queue; the operator approves/hides/features at /god/commons (admin RLS). /marquee redirects here. v7.10.0: moved the feed + moderation off the wedge-prone supabase-js client onto the lock-free REST path (they were reading empty despite live rows). The big feature expansion is v14 (THE COMMONS, ALIVE).
  • live Installable PWA + share cards v3.1.0: manifest + service worker (offline shell), intelligent install prompt (value-moment, cooldown, iOS-aware), Open Graph share image + Web Share. Install prompt + share are operator flags.
  • live DAILY CHALLENGE + streaks, the growth loop v5.0.0 + v5.1.0: one deterministic drill a day (/daily) with a spoiler-free Wordle-style result grid + Web Share; a streak (freeze every 5 days, loss-aversion) that SYNCS across devices (user_streak table) and logs daily_completed events. The shareable daily card gets sharpened into the viral spine in v13.
  • live Prompting 101, the beginner track v7.11.0: live at /start. 8 short chapters (6–10 min each, one idea per chapter) on the Task · Context · Role · Format spine, everyday examples (the hard email, the medical report, the trip), all four drill mechanics, deterministically graded through the same engine as THE COACH (lib/content/beginner.ts). Every chapter ends with a keeper in the shared Playbook (b101- ids). Operator kill-switch feature.prompting101.
  • live The 90-second Aha v7.11.0: live at the top of /start (AhaWidget). The vague ask gets the generic answer; you add the three facts only you know (days, equipment, a limit) and watch the same question come back transformed. Fully client-side (template substitution, no model call) so the highest-traffic page costs nothing to serve. Verified interactively via the Playwright harness.
  • live The soul statement, vague works, but directing it does wonders v7.11.0: the thesis leads /start ("Any AI will act on a vague ask. Learn to direct it, and it does far better work for you."), is FELT via the Aha demo directly beneath it, and About routes newcomers there. Shown, not told.
  • live Choose-your-track front door v7.12.0: three doors on the unified /learn home for a fresh visitor, I’m new to AI (→ Prompting 101), I use AI already (→ the craft, chapter 1), just need something useful now (→ quick wins). One tap, right rung, no quiz.
  • live The seamless bridge to THE COACH v7.12.0: the bridge is structural, /learn is now THE one learning home, with Prompting 101 as Part 1 and the operator’s craft as Part 2 of a single ladder (one progress bar, a visible seam, chapter completion hands off to the next rung). Lessons are never scattered across pages again.
  • live Everyday quick-wins, useful on day one v7.12.0: eight ready prompts for real life on /start (understand a medical report, the hard email, decode a contract, dispute a bill, homework that teaches, the trip that fits, resume bullets, summarize-with-the-action), copy into any AI or keep straight into the Library, no signup.
  • live Command-palette + faceted catalog v7.14.0: live. Ctrl/Cmd-K (or the ⌘K chip) opens the quick-find palette on /library: one search across YOUR items and the whole catalog, one-hit copy (snippet references resolve on the way out). Explore gains kind facets (system / prompt / snippet) beside categories and search.
  • live Everyday-life categories v7.13.0: five everyday shelves in the starter catalog (lib/starter/everyday.ts, 25 prompts): Health & documents, Money & consumer, Family & home, Job search & career, Life admin. Medical reports, bill disputes, homework that teaches, interview grilling, contract decoding. The Library serves daily life, not only job work (most real AI use is non-work).
  • live Manage-your-own at scale v7.14.0: live. Collections (assign in compose/edit, filter chips, bulk Move), a Recent view (sorted by last use), Duplicate in the detail sheet, and select-mode bulk actions (star / archive / move) with a floating bulk bar. Archive stays the non-destructive default.
  • live Variables + snippet composition 2.0 v7.14.0: live. Composition: [[snippet:Title]] inside any prompt expands to your snippet of that title (3 levels deep, loop-safe) before fill-ins run, every use / fill / run path resolves it, and the compose sheet has an Insert-a-snippet picker. Prompts now build from reusable parts.
  • live Operator rewrite-on-add (quota-limited) v10.5.0: THE QUIET HAND. One small Rewrite button in the corner of every text field (Library body, Tricks body and why, the Forge bench, the share sheet fields): it rewrites your text IN PLACE on the fast model (~2s) and turns into Undo, which restores your original exactly. Edit the rewritten text and the button returns for the new human text (the no-mirror rule falling out naturally). THE STRICT-TAG RULE holds: only text inside rewrite tags ever lands, so refusals and malformed replies can never touch a draft. The on-device lens chips stay. The heavy read panel is gone; the full critique lives at /grader.
  • live Versioning + diffs v7.14.0: live. Every body edit snapshots the previous version (capped history per prompt); the detail sheet lists history with one-tap Restore (itself versioned, nothing lost). Local-first by design; the cloud syncs the current text.
  • live THE TRICKS, a living tips & tricks shelf v7.13.0: live at /tricks. 31 house tricks on TWO live filter axes, topic (memory, handovers, context hygiene, output shaping, recovery, cost & speed) × tool (Claude web / desktop / mobile, Claude Code CLI, ChatGPT, Gemini, Copilot, Perplexity, any AI). Each trick: the move, why it works, and a copy-ready line where there is one; Keep-to-Library on every card. Add-your-own (local-first) with share-to-Commons through the normal moderation queue. feature.tricks kill-switch.
  • live The Playbook format v7.16.0: live. lib/playbooks: role, what/when, a [[fill-in]] document sharing the proven skeleton (role, context, hard rules, the loop, first message). Fill, copy, keep-to-Library, or download as.md at /playbooks; the push gate validates every playbook structurally.
  • live Seed from the house playbooks v7.16.0: live. The four house originals distilled into fill-in form under Technical, provenance on every card: THE WEB FACTORY, THE GAME FACTORY, THE GAME-FEEL BIBLE, THE DESIGN BIBLE. The raw sources stay in docs/reference/playbooks/.
  • live The role catalog, “we cover everyone” v7.21.0: 44 playbooks across 43 roles and all five categories (batch 2 added student, caregiver, freelancer, small business, customer success, researcher, ops, EA, nonprofit, coach, e-commerce, community, podcaster, journalist, security). New roles keep landing as content commits.
  • live Everyday-life playbooks v7.16.0: live. Household CFO, Homework Coach, Patient Advocate, Trip Architect, Job Campaign, Tenant File, Refund Machine, Event Runbook: eight standing workflows for real life.
  • live White-collar role playbooks v7.16.0: live. Lawyer, teacher, accountant, HR, recruiter, product manager, support, consultant, plus marketer, salesperson, and real-estate agent under Sales & marketing. Each interviews before it writes and knows where the professional line is.
  • live Technical starter playbooks v7.16.0: live. The four house seeds plus the Analysis Charter (read-only data discipline), the Test Strategy, and the Infra Runbook (plan-first change discipline, tested backups).
  • live THE FORGE, surgical prompt-making v7.17.0: live at /forge. Pick a base (a playbook, one of yours, or a catalog prompt), stitch and edit on the bench, and the whole Library is raw material. feature.forge flag.
  • live Grade-in-place v7.17.0: live. Grade-this sends exactly what is on the bench through the gateway (same grader kind, quotas, ceiling, ledger); the operator’s read comes back inline.
  • live Mix-and-match templates v7.17.0: live. [[snippet:Title]] stitches your snippets in, fill-ins resolve live on the bench, and the result saves as a new prompt.
  • live Run from the Forge v7.17.0: live. Run-in-a-model from the bench reuses the CROSS-MODEL targets, copy-first.
  • live Keep-it into your Playbook v7.17.0: live. Keep lands the forged prompt in your Library (tagged forged) with sync, history, and collections like anything else.
  • live THE VAULT, the knowledge corpus as copy-ready building blocks v7.21.0: /vault is open. The corpus atomized into 25 copy-ready blocks across six domains (scoping, verification, sessions, design systems, feel, shipping), each with the moment to use it and where it was proven. Keep-as-snippet feeds the Library and the Forge. First pack: builders.
  • live Art-direction pass (Aamir picks) v7.18.0: Aamir picked Direction B (the Raycast model) from three rendered candidates on: chrome stays void + one cyan; saturated color lives ONLY inside illustration panels. The law is canonical in DESIGN_BIBLE.
  • live The illustration system v7.18.0: SpotArt.svelte, ten named code-drawn spots in Direction B, applied across /start, /learn, the Library empty state, and the tricks/playbooks/forge headers. Zero assets, zero licensing, crisp at any scale.
  • live Concept spot-art per chapter & category v7.21.0: every chapter now carries its own code-drawn concept mark (17 chapter glyphs on the /learn ladder, drawn from the idea each chapter teaches) and every playbook category its own (5 more on /playbooks). Same SPOT LAW language: color lives in the mark, chrome stays void and cyan.
  • live The Operator’s visual identity v7.18.0: the operator mark: a lit prompt-chevron workbench panel, not a mascot (cuteness costs trust, per the voice bible research). It leads the front door.
  • live A beautiful first run v7.18.0: /start opens with the operator mark over the soul statement, then the Aha. The first screen invites.
  • live THE FEEL LAYER, motion & micro-delight as acceptance criteria v7.18.0: the standard is codified and applied: JuiceManager haptic grammar, instant primary actions, the.rise staggered list entrance (reduced-motion safe) on the new shelves. DESIGN_BIBLE documents it.
  • live THE GRADER GRID, daily challenge + spoiler-free share card Sharpened: the daily share is now a rendered PNG card (grid + streak, on-device canvas, zero server cost) where file-share exists, the emoji grid everywhere else, plus a save-as-image path and an honest "one of N clean runs today" pulse.
  • live Before → after share One tap on the Aha reveal shares the vague ask next to the operator ask, pointing at /start, the most honest recruiting asset the app has, because the receiver can run the same experiment in ninety seconds.
  • live The Operator report card Your level, rules kept, longest run, and medals rendered on-device as a postable card (canvas PNG, native file share with a download fallback). Numbers only, no prompt text ever leaves the device.
  • live Referral loop with visible progress Your invite link lives on the profile; arrivals through it are counted once each (aggregate-only RPC, raw events stay admin-side) and feed the Recruiter and Founding circle medals. No signup wall, no dark patterns.
  • live Calm cohorts (status without leaderboards) Presence, not ranking: "climbing with N operators this week" on the profile and "one of N clean runs today" on the daily, shown only when the numbers are real (3+ / 2+), never a leaderboard, never a name.
  • live THE OPERATOR’S LADDER, collection meta on the Playbook Ten medals, every one a pure function of real work (rules kept, days shown up, forges, tricks, clean dailies, invites landed) with endowed-progress bars under the unearned ones. Boss scenarios stay charted for a later train.
  • live Themed shelves Live: the operator files any live post onto a free-text shelf from the cockpit; the feed groups by shelf (biggest first), the rest flows under "Fresh from operators".
  • live Search-by-situation Live: the search box on /commons asks "What are you trying to do?" and ranks matches on the PREVENTS field above title and body, so the situation wins over the keyword.
  • live Provenance-rich cards Live: cards now carry what the prompt prevents, a collapsible worked example, the model it fits, and its author credit. Authors fill example + handle in the share sheet.
  • live Keep-with-credit + calm author pages Live: every keep credits the author by name ("Kept. Credit to X."); tapping a name opens their calm view (shares + total keeps, deep-linkable via ?by=). No follows, no ranks.
  • live Weekly Operator’s Pick Live: the operator sets one post + a WHY note from the cockpit (stored in app_config); it leads the feed with a cyan seam and the note quoted underneath. Editorial, not algorithmic.
  • live Moderation cockpit 2.0 Live: reports triage on top, bulk select + approve/hide over pending, quality auto-flags (very short / link-heavy / all caps / no provenance), shelf filing, pick setting, and one-tap seed-from-my-Library.
  • live THE FIVE ROOMS: the whole app on five intent-named tabs, every page two taps away (v12.0.0). Navigation rebuilt around what people came to DO: Home (launch), Learn (the ladder + daily + drills), Create (library, grader, playbooks, forge, tricks, vault, generators, builders shelf), People (Commons, team shelf, model guide), You (playbook, progress, spend, guide, why, roadmap). One shared map (lib/rooms.ts) renders the TabBar, the desktop rail, and the room hubs, so they can never drift. The More drawer retired by folding; the home page became a launcher and its story moved to /about.
  • live THE WARDROBE: six switchable looks + the morph ritual The whole app re-voices through one token contract: the refined Instrument plus five owner-picked dark looks (Night Library, Arcade, Neon Run, Ink & Brass, Aurora). A handcrafted picker reachable from every screen; switching washes the new look over the app from your fingertip. Choice follows the device. v9.0.0 polish: surgical color passes over every theme (primary-button tokens, contrast audit), desktop responsive passes, and the morph slowed into a 2.6-second evolution with a light ring riding its edge.
  • live THE SMART LAYER, the app organizes for you Two tiers by design. Free on-device heuristics everywhere: Organize-for-me in the Library compose sheet (title, tags, prompt-or-snippet from the text itself), auto-filing in Tricks (topic + tools from your words), and quick-find now searches the whole workbench (playbooks, tricks, Vault) beside your items and the catalog. A tiny fast-model polish rides on top where quota allows (new gateway kind assist, fast tier, ~250 tokens, 3/day signed-out and 25/day signed-in, inside the $5 ceiling). AI is never a dependency: the heuristic answer always stands. feature.smart kill-switch.
  • live THE VELVET ROPE, everything behind one beautiful sign-in The owner flipped the model: members only. A signed-out visitor gets exactly ONE thing, the 90-second Aha on the home (feel the craft work, no signup), then a shelf of locked doors and one honest line: everything opens with a magic link. Walk to any route and a warm gate names what is waiting there instead of a wall. No password, no card. The operator can drop the rope with one flag (feature.velvetRope). Sign-in stays the only method until OAuth providers are enabled.
  • live THE BRIDGE, your library inside your tools (MCP + CLI) v11.0.0: live at /connect. Mint a personal token (shown once, only its hash stored, revocable) and FULL AUTHORITY mounts inside Claude Code, Cursor, or any MCP client: search_prompts, get_prompt, save_prompt, latest_handover, save_handover, get_contract. The terminal gets the same through the fa script (curl install, no npm chain). All calls are scoped to the token owner; verified live end to end.
  • live THE RELAY, session handovers that travel between tools v11.0.0: live at /relay. End a session by saving one structured handover (the house template: state, decisions, next steps, gotchas, how to verify); load it tomorrow here, via fa handover load, or by asking any MCP-connected agent for your latest handover. The research-verified wish ("session handoff should be built in") made real.
  • live THE ONE CONTRACT, one rules file compiled to every tool v11.0.0: live at /contract. Author one agent contract (in-place rewrite + undo included); export it compiled to CLAUDE.md, AGENTS.md,.cursorrules,.windsurfrules, and.github/copilot-instructions.md, copy or file download each. Agents fetch it live via the Bridge (get_contract). One-click GitHub pull requests are charted, honestly labeled coming.
  • live THE PROOF PACK, verification demands in one tap v11.0.0: live at /proof. Describe the task (type, stakes, files, data, tests) and get the verification block to paste at the end of any ask: reproduce-first for bugfixes, unchanged-behavior for refactors, row counts before data changes, plan-first for infra, evidence demands throughout. Deterministic and on-device, zero model cost. The direct answer to the deepest verified pain: the review burden.
  • live THE METER, your AI spend, honestly v11.0.0: live at /meter. Every model call the app makes for you, at real token prices: totals, by day, by feature, straight from the same ledger the operator console reads (owner-only RLS). No surprises by design; cross-tool capture (a CLI wrapper) is charted and labeled coming.
  • live THE TEAM SHELF, prompts stop dying in chat v11.0.0: live at /team behind feature.teamShelf. Start a team or join by invite code; share prompts with provenance (what it prevents, who proved it); anyone keeps anything into their own Library with credit. Small by design: one shelf, no ranks, no feeds.
  • live CAPTURE, paste anything and it files itself v11.0.0: live at /capture. Paste a prompt from a Slack thread, a Discord message, or a gist; the smart layer names, tags, and shelves it into your Library in one tap. On Android the app registers as a share target, so Share-to-FULL-AUTHORITY works from inside any app. Direct Slack/Discord connectors are charted and labeled coming.

Coming

every planned train has shipped

All fourteen version trains are live. The backlog below holds what gets trained next.

Backlog

not yet scheduled
  • charted Subscription gating + full share layer (P4) The paywall (gate breadth/depth, never the daily habit) + the fuller share layer. Attaches once the trains above give the free tier enough to be generous and the paid tier enough to be worth it.
  • envisioned Team spaces · scenario packs · the arena · certification · grader duels · operator benchmarks Candidates, not committed. Team spaces and certification are the enterprise up-sell once the Library + Playbooks land.

Changed course

honest notes
  • diverged THE STAGE, the film-set four-beat loop Retired in the v3 INK rebuild. Replaced by the interactive blast-radius drill (see coach-ch1-2). Kept in git history.
  • diverged THE MARQUEE (share-out), turn any direction into a shareable one-sheet Parked in the v3 rebuild, folded to an honest note + a way back at /marquee; code/data intact. Re-enable via the feature.marquee flag when there is an audience. General share cards ship instead (see pwa-share).

← Home