Home / Ops / More in this area

_SHARED_CONTEXT - the cross-chat shared brain (read me)

Updated Jul 04, 2026 · _SHARED_CONTEXT.md

_SHARED_CONTEXT - the cross-chat shared brain (read me)

A living, shared record for EVERY chat working on Affirmology: the DATA/engine chat, the code/apps chats, and Cowork. It is the "things worth knowing" layer, decisions, standing principles, current phase, and owner split, so we all stay aligned and Jeff does not have to re-explain across windows. Append to it as decisions are made. For minute-to-minute deploy coordination use _CHAT_SYNC.md; for the authoritative deep state use Affirmology_PROJECT_STATE.md. This file is the fast shared map.

Started 2026-07-04 at Jeff's request.

Current phase (as of 2026-07-04)

• Beta is NOT starting today. Target open: Fri 7/10 (Thu 7/9 tighter). Jeff on a family trip, back in the Miami office late Tue 7/8 / early Wed 7/9. Plan: Affirmology_BetaLaunch_RevisedPlan_v1.md.
• Gift drip to the list (done with care, not rushed): Mirror ~Sun 7/5, a new audio Mon/Tue, the report Wed/Thu. Founders excluded; maybe a few people trimmed.
• Today (7/4): a beta-expectations + affirmational-agreement email (Cowork owns the copy).
• The Mirror is HELD and being retooled (structure/content). Do not batch-send it.

Standing principles (do not violate)

• Map, never the full picture, for a creation prompt. Generation prompts get a compact MAP of reference material; the full detail is pointed to, never dumped in. Full stays checkable in the corpus + a FullReference doc. (First applied to the storytelling craft menu; bake it everywhere.)
• Iterate, don't regenerate. For a tweak, edit the person's EXISTING script and re-voice it verbatim (/render-verbatim). Only full-regenerate for a structural rewrite or with an explicit ask. Losing liked language to a re-render is the thing we are preventing.
• The Profile is the canonical source. Compute a person's chart once, store it, read from it for every audio/reading. Affirmology_ProfileCanonicalSource_v1.md.
• The council is our signature weapon AND our biggest cost. It is the expensive multi-round call (the Mirror uses it). Run it ONCE per person and cache the output; re-voice cheaply. Never fire it repeatedly at the same person.
• All audio comes from our system. Founders (Jeff, Sol) excluded from customer counts/reports. Never em dashes. "Subconscious Operating System" frame. Sol = Soledad Gabriela Ballard.

Cost discipline (learned the hard way 2026-07-03: credits hit zero mid-batch)

• Turn on Anthropic auto-reload; enforce a low-balance warning (not just a note).
• Trim prompts (the craft menu was 8.5k tokens/render, now a ~2.7k map, full on demand).
• No blind auto-retry loops (a keepalive re-firing a failing batch wasted cycles).
• Before any batch: dry-run + cost estimate + confirm.

Owner split

• DATA / engine chat (this one): the render/audio pipeline, profiles, chart engine, corpus/craft, cost-safe generation, the demo results page, gift renders, recovery.
• Code / apps chats: the iOS/Atlas apps, TestFlight, cloud-Fish, the Mirror batch runner, the app accounts and in-app library, the wiki.
• Cowork: the beta agreement text, outbound email copy, app/TestFlight account linking (Sol), wiki.

Decisions appended 2026-07-04 (Cowork session, reading-craft reset)

• THE UNDERSTANDING is decided (Jeff, 7/4). Code's "structured Understanding" = Cowork's "Deep Map"; one name: the Understanding. The council builds it ONCE per committed person (12-field schema v1 in Affirmology_FlagshipReading_Proposal_v1.md section 3: ranked salient findings, convergences, aha moves, wound/gift/gap, mechanisms, roads, metaphor bank, one-liners, fear preemptions, timing, context store, learned-info store). Every audio/chat after is a cheap shaping (select facets + apply craft). APPEND on new learnings, never recompute. Built INSIDE the Mirror rebuild. Two tiers: Tier 1 light (everyone gets Origin), Tier 2 the deep run on commitment (recommended trigger, pending Jeff: the "Yes, I'm in" beta reply).
• The hero's story is the product-wide standard. Every output makes the listener feel seen, epic, meaningful, and shown a path. Craft standard + Mirror v3 postmortem: Affirmology_ReadingCraft_HeroStory_Synthesis_v1.md.
• New first reading structure spec'd (public name stays The Mirror; internal mirror_recognition_v1): Promise, Being Seen, Three Doors (Big 3 convergences), Install, Seal; ONE threaded master symbol; a loop always open; 12-rule QC gate; test on Sol first. Affirmology_FirstReading_Structure_v1.md.
• Origin Soul Song B variant spec'd (origin_soul_song_b_craft, six line-level changes, separate axis from The Recognition A/B): Affirmology_OriginSoulSong_Bvariant_Spec_v1.md. Pronoun bug ("becoming her" in Jeff's gold script) must be fixed in both arms.
• The 72-min wealth reading transcripts were recovered (they existed only as MP3s on R2; Cowork transcribed both): Affirmology_WealthReading_Transcript_{Jeff,Sol}_v1.md, the canonical style exemplars. Known render flaws found there: composer meta-seams spoken aloud + Jeff's MP3 truncated at 71:37 -> QC now includes seam-strip and end-integrity checks.
• Beta invite email drafted (Affirmology_BetaInviteEmail_v2.html): craft-first framing, one small trip line, human-answers service line, reply "Yes, I'm in" as the affirmational signature. SEND DEPENDENCY: hello@affirmology.ai must receive replies (Email Routing forward) and the send must use a reply-able from, not the no-reply reports@ pattern.
• Jeff's fast map for launch week: Affirmology_LaunchWeek_OnePage_v1.md (his remaining open decisions: email send timing, tester list, Tier 2 trigger, starting audios set).

Decisions/state appended 2026-07-04 (render + cost-safety thread)

• Batch cost safeguards BUILT + DEPLOYED (affirmology-studio, commits 76840e8 + 6827996). POST /api/batch-mirror now enforces three independent brakes: (1) a spend cap (max_cost_usd, budgets ~$0.45/reading, stops BEFORE draining the pool below the demo's reserve), (2) stop-on-credit- error (aborts the whole run on the first Anthropic "credit too low" error instead of hammering), and (3) a gentle overnight throttle (pace_seconds, one-at-a-time drip; 25 over 8h = ~1140s apart). Workflow: Jeff states the balance before an overnight run, the cap is set to leave a demo reserve. Cost discipline is now ENFORCED in code, not just a principle. Batch stays HELD (all 27 in .mirror_hold) until the Mirror rebuild is validated.
• Hybrid render architecture spec'd (Affirmology_HybridRenderArchitecture_v1.md, DESIGN): cloud = always-on router + demo; local Mac = preferred render engine when awake, on its own credit pool. Cloud health-checks the local node and offloads renders to it BEFORE spending cloud credits, falls back to cloud when the Mac sleeps. Local overnight runner (caffeinate + idempotent). Jeff DEFERRED buying a second Anthropic account for now - running on the shared account with the safeguards above. Verified 7/4: local + cloud Anthropic keys are the SAME account (no isolation yet, so the safeguards matter).
• Audio-engine layered architecture (Affirmology_AudioEngine_LayeredArchitecture_v1.md) - companion to THE UNDERSTANDING decision; same model.
• Wealth transcripts also recovered by this thread (faster-whisper "small", same two filenames). If Cowork's transcription (with the seam/truncation QC findings) is higher-fidelity, treat Cowork's as canonical so there is one version.

• Site fully live (direct wrangler deploy; the Workers CI was stuck one commit behind): green Mirror player matching the Origin, NO download on either song, results page lands CENTERED on the Mirror. Investor PDF republished + live (media.affirmology.ai/docs/mission-and-opportunities.pdf).

• Beta acceptance mechanism spec'd (Cowork, 7/4): terms presented as email summary cards + a linked full agreement page (Affirmology_BetaAgreement_v1.md, page copy + internal tracking spec). Acceptance = the "Yes, I'm in" reply to hello@; logged in Airtable People (Beta committed + beta_affirmed_at + verbatim reply text -> the person's Understanding context_store). Every yes gets a confirmation email within the day (Affirmology_BetaConfirmationEmail_v1.html). Do NOT use the broken results-page checkbox for founding-circle acceptance. If Jeff approves the Tier 2 recommendation, the logging event also fires the overnight Understanding run.

• The Norm lesson (7/4, Cowork correction): grading people from CRM rows alone is WRONG; the Profile store is canonical (existing principle, now proven). Norm Adams "missing birth data" was only a thin CRM row; his full profile (Norman Walter Adams, Origin delivered 6/30, clean) existed all along. Cowork synced his CRM row. RULE: any people-triage must cross-check the profile store before declaring data missing; Group D of Affirmology_InnerOrbit_ListTriage_v1.md is provisional pending that cross-check (engine chat). Also: beta is named THE INNER ORBIT (customer-facing); invite email + agreement renamed, no headcount ever stated to the list.

Open threads to know

• Sol's app is empty because her audios are not attached to her app account (solballard@gmail.com, which exists). Backfill of her 7 audios is staged; waiting on her app-account id to attach them.
• Push test audios into Jeff's + Sol's app (in-app + email) via /api/library/ingest; needs the account-id link. See the beta plan doc.
• Storytelling/hooks craft now wired into script generation as a compact map; FullReference doc = Affirmology_StorytellingHooks_CraftMenu_FullReference_v1.md.
• Graceful failure shipped on the results page (a dead render no longer spins forever).

Pointers

• Deep state: Affirmology_PROJECT_STATE.md · Beta plan: Affirmology_BetaLaunch_RevisedPlan_v1.md
• Tactical deploy coordination: _CHAT_SYNC.md · Canonical profiles: Affirmology_ProfileCanonicalSource_v1.md
• Craft full reference: Affirmology_StorytellingHooks_CraftMenu_FullReference_v1.md
• Reading craft standard: Affirmology_ReadingCraft_HeroStory_Synthesis_v1.md · Reading structure: Affirmology_FirstReading_Structure_v1.md
• Understanding schema + tiers: Affirmology_FlagshipReading_Proposal_v1.md · Launch-week map: Affirmology_LaunchWeek_OnePage_v1.md
• Style exemplars: Affirmology_WealthReading_Transcript_{Jeff,Sol}_v1.md

Decisions/state appended 2026-07-04 (DATA chat: verbatim + recovery + graceful failure)

• Verbatim voice render shipped + live (POST /api/render-verbatim, self-contained module). Voices an EXACT edited script with NO regeneration, through the proven Origin pipeline, and persists a demo job so the permanent results page serves it. This is the enabler for the "iterate, don't regenerate" principle and for user-created variations later. Use it for every micro-edit (edit the stored script's targeted part, re-voice) instead of regenerating.
• Graceful failure handling shipped (affirmology-site, results page). The returning poller never checked for a failed render and never timed out, so a dead render (e.g. the credit outage) spun "Loading your page" forever with no email. Now: on a failed status or an 8-minute cap it shows a reassuring message; the song still emails when it lands. Covers both the emailed-link and first-time film paths. Any future dead render never strands a person.
• Chasity + Colin recovered (7/3-7/4). Both took the demo DURING the credit outage, so their Origin renders failed and they got nothing (no song, no email). After credits were restored, DATA re-rendered both from their stored profiles and emailed them their results pages. Chasity: chasitywill@gmail.com, Douglas GA. Colin still has TWO duplicate profile rows (cjacobsandassoc + colinconnection) to collapse.
• Craft menu = compact MAP, live (62688f2): storytelling/hooks wired into script generation and the council as a ~2.7k-token map (was 8.5k full); full detail on demand + ..._CraftMenu_FullReference_v1.md. Dedupe pass next. This is the first application of the map-never-full-picture safeguard.

Decisions/state appended 2026-07-04 (DATA chat: nurture ladder + cheap Mirror validated)

• The Nurture Ladder / Goody Menu is spec'd (Affirmology_NurtureLadder_GoodyMenu_v1.md). After the Origin Song, a person gets a short drip of DISTINCT, cheap, high-value goodies, each a new modality + angle, each opening a loop the gate closes: Rung1 Mirror Door One (~$0.50 spoken), Rung2 Signature Card (visual PDF, ~free), Rung3 a 2nd audio of a different type (~$0.20), Rung4 a living "Season" timing note (~$0.10). Whole ladder for a never-converter ~$1.00-1.50; full Understanding only fires on commitment. SAME shape serves both gates: pre-beta "Yes I'm in" now, PAID membership trigger post-launch. Sequencing: build the CHEAP Mirror + a couple goodies NOW for the pre-beta list; build the elaborate "$1-at-a-time" on-demand journey AFTER beta teaches us the real journey.
• The cheap one-door Mirror is VALIDATED at length. Sol's Mirror "Door One" chunk composed via /api/compose (single council pass, ~$0.50): 1,411 words, ~10.7 min paced, 0 em dashes, one loudest convergence (Gate 41: Sun+Venus Aquarius / HD gate 41 lines 2-3 / Gene Keys Life's Work + Venus Attraction + Pearl), full segment craft, clean open loop to "the second door." Saved Affirmology_SolMirror_Chunk1_TestScript_v1.md, emailed to Jeff for approval. NO audio rendered yet (awaiting Jeff's go, then Fish-paced render). This is the near-term pre-beta tool.