Goal spec

​

Goal prompt (paste into /goal)

Mission. Build a searchable, video-accompanied training & SOP site for Synscribe, in portable Markdown under docs/training/, that (a) can onboard a brand-new account manager (internal or agency partner) to run client accounts at scale, and (b) works as a guided tour for future direct-to-subscriber self-serve users. The full page map, audience model, and phase plan live in docs/training/PLAN.md — that is your build spec; follow it and keep its status ledger updated as you complete pages. Success is not “all pages exist” — success is the Definition of Done below, every gate passing, verified rather than asserted. Non-negotiable constraints (invariants):

1. Ground truth only. Never invent a feature, route, skill, or behavior. Every platform claim must be traceable to the codebase, live prod, an existing SOP, or a video. When unsure, verify with an Explore subagent or the prod-context skill before writing it. Honor the known corrections: the guardrail does block at publish (its settings copy is stale); only GSC/PostHog/Mixpanel/DataForSEO are per-client integrations (Ahrefs/Apollo/Instantly/ Unsplash are platform-wide). Do not touch the abandoned theta/epsilon/gamma stack.
2. Never mention credits. No credit units, no “N credits = M pages” anywhere. Describe throughput/cadence concretely (posts/week, landing-page batches/week). A grep for credit-as- currency must come back clean.
3. Real prompts, mined not invented. Synscribe is agent-operated, so most “how” is a prompt you send Pi. Mine real operator prompts from production chats via prod-context (factory session/messages) — route all transcript reading through subagents so raw transcripts never enter the main context; extract only the reusable prompt + pattern. Lightly anonymize client names/PII. Pull from the high-activity orgs (wonderchat, chalkie, frameshft, straitsx, rently-uae, bluente, bookaclinic, scheduling-wizard). Never paste a raw transcript into a doc.
4. Audience model. One site, audience-tagged. Self-serve users can do everything — the difference is discipline/rigor/standardization at scale, not gated capability. Tag divergences For AMs (the rigorous, checklist-driven version) vs For self-serve (the lighter path). Part 3 (SOPs) is AM-facing.
5. Honesty over coverage. Anything you cannot verify gets a visible ❓ [needs Raymond: …] marker and an entry in the gap-register page — never a confident guess. Failing or ambiguous items are stated plainly in the journal, not smoothed over.
6. Portable Markdown. Clean front-matter, relative links, a consistent video-embed slot convention, so the whole tree drops into a site generator (Starlight/Mintlify) later with no rewrite.

Milestones (ordered; each has an acceptance gate you must satisfy before moving on). Commit after each milestone on branch docs/account-manager-training with a journal entry in docs/training/JOURNAL.md (what shipped, what was verified + how, open ❓s). This run is autonomous — committing/pushing is authorized.

• M0 — Scaffold & conventions. Create the directory structure for the 3 parts, a home/index page, the front-matter template, the video-embed convention, the audience-tag convention, and empty JOURNAL.md, gap-register, and video-shot-list pages. Gate: a new contributor could read the conventions and add a conforming page.
• M1 — Mine the sources (parallel subagents). Produce a staged docs/training/_sources/ with: (a) a distilled Synscribe SEO principles outline for Part 1 (from Workshop Notes + client dossiers); (b) a real-prompt library tagged by task, each with provenance (org + rough date) and anonymized; (c) a video embed map (every Loom/YouTube ↔ its page); (d) a confirmed VFS conventions note (folders like /reports/, memory, soul — verified from prod chats + code). Gate: every downstream page has its raw material staged; the prompt library has ≥15 distinct real prompts spanning keywords, landing pages, blogs, publishing, analytics, and review.
• M2 — Part 3 SOPs (do first; next week’s AMs need these). Every Part 3 page from the map, as a runnable SOP: inputs checklist → numbered steps → acceptance checks → edge-case table. Port the already-strong SOPs (Onboarding Month 1, Organise Keyword) faithfully; write the rest. Gate: a brand-new AM could onboard a client end-to-end using only Part 3 + its links, with no tribal knowledge required. Spot-check by reading 3 pages back against this gate.
• M3 — Part 2 platform how-tos. Every task page: concise steps + ≥1 embedded video (where one exists) or a shot-list entry + at least one real mined prompt block where the task is agent-operated. Gate: no page is a stub; every existing video is embedded on the right page; every agent-operated page has a real prompt; every video-less page has complete written steps OR a shot-list entry (task + 3–5 beats + est. length).
• M4 — Part 1 theory. Every Part 1 page, each making a specific, non-obvious Synscribe stance (not generic SEO advice), backed by a real example (ideally a client dossier or a mined chat). Gate: each page passes the test “a seasoned SEO reads this and says ‘that’s a real, specific point of view,’ not ‘I’ve read this a hundred times.’” No filler.
• M5 — Assemble & verify. Home page with two learning paths (“New AM — start here” and “Self-serve tour”); working cross-links between parts; the consolidated gap-register and video-shot-list pages complete; a final self-audit pass that reads a representative sample of pages and checks them against every DoD gate, listing any misses honestly in the journal. Gate: the full DoD below passes.

Definition of Done (the judge’s checklist — ALL must hold):

• Coverage: every page in PLAN.md’s map exists and is substantive (a real micro-guide, not a placeholder). Any intentionally-deferred page is explicitly listed and justified.
• Grounded: no invented features; the known corrections are reflected; a skeptical reader could trace any platform claim to code/prod/SOP/video.
• Real prompts: Part 2 agent-operated pages carry real, mined, anonymized prompts; the prompt-library page consolidates them with provenance.
• Video: all existing Looms/YouTube embedded on the correct pages; all remaining video-needs captured in one shot-list with task + beats + length.
• Differentiated Part 1: every theory page states a specific Synscribe stance with a real example; no generic filler.
• Runnable Part 3: a new AM could execute each SOP unaided (inputs, steps, checks, edges).
• Audience-correct: tags applied where guidance diverges; self-serve framed as full-capability-lighter-rigor; zero credit language (verified by grep).
• Navigable & portable: home + learning paths, consistent front-matter, working relative links, generator-ready Markdown.
• Honest: every unverifiable claim marked ❓ and aggregated in the gap register; the journal states what was verified, how, and what remains open.
• Verified, not asserted: the final self-audit actually re-read a sample of pages against these gates and recorded the result.

Anti-goals (these count as NOT done): empty or one-paragraph stub pages; generic SEO content in Part 1; invented UI/skills; raw transcripts pasted in; credit talk; silent guesses where you should have marked ❓; declaring done without a verification pass; optimizing for finishing quickly instead of finishing well. Working style: use subagents liberally (mining, verification, per-part drafting) to stay deep without flooding context. Keep PLAN.md’s status ledger current. Prefer making a page more concrete/useful over moving on. When you believe you’re done, run the self-audit before claiming it, fix what fails, and only then report completion with the journal as evidence.

​

Notes for whoever launches this

• How to launch: /goal with the prompt block above (or /goal build the training site per docs/training/GOAL.md). The Stop-hook + judge will hold the loop to the DoD gates.
• Env: the run needs FACTORY_API_KEY in the shell for prod-context mining (provided, verified). It reads factory.agent_session/messages only — read-only, audited.
• What Raymond still owns: reviewing Part 1 positioning once drafted, and answering any ❓ the run surfaces into the gap register. Everything else is meant to complete unattended.
• If genuinely blocked on a Raymond-only input, the loop documents the block, does NOT guess, and continues with the rest of the work rather than stalling.