Every AI surface in the product runs exactly one skill per context (FS_9 rule). Each skill is a SKILL.md file with a mode key, an injected-context contract, and an entry surface. This registry is the single source of truth for the skill dropdown, the skill bar, and routing. New skills are added here first, then to the dropdown.
.sk-grid — auto-fill minmax(255px,1fr).sk-card — bg1, bd2, hover border accent.sk-key — mono 2xs accent, is the routing mode key.sk-meta — Context in / Entry / Output rows.sk-group-h — mono label + 1px rule fill
shipped / build / spec — drives dropdown visibilityspec skills hidden from dropdown behind flag skills.previewskills.registry.json renders both this grid and the dropdown
The existing searchable combobox, upgraded from a flat 7-item list to grouped sections matching the registry. Generation skills appear but are gated: selecting one without a confirmed section schema routes to Opportunity Response first. Type to filter across all groups; group headers hide when empty.
.dd-group — mono 8px label row, bg2.dd-item — name left, mode key rightgrant_generation / contract_generation from dropdown → opens opportunity picker → Opportunity Response → write gatedocument_rating with no rubric → rubric picker firstspec-status skills hidden unless skills.preview
Shown on the opportunity detail page after the user clicks Respond. The opportunity_response skill reads the opportunity fields + attached solicitation and renders this card. Four derived stats, a factor list reusing the GO/NO-GO chip language, and a footer that advances to section confirmation. While analyzing, the card shows skeleton rows with status ANALYZING…
analyzing (skeleton rows + pulsing dot) → complete → error (retry). Factor chips reuse the Eligibility Analysis vocabulary (Qualifies / Risk / Gap). Gap items write to the same Unmet Requirements inventory the contract writer reads (Q48) — one inventory, two consumers. The footer CTA is the only path forward; analysis is never skippable when a solicitation is attached..an-card — 680px max, bd2.an-stats — grid-4, mono values, hi/med color.an-row + .an-chip — chip-ok / chip-warn / chip-gap.an-foot — note + Re-run + primary CTA
analyzing — skeletons, CTA disabledcomplete — shown aboveerror — danger band + Retryanalysis_result, section_schema[], unmet_requirements[]
The user confirms the section schema before anything is written. Sections extracted from the solicitation carry a “From your solicitation” tag; the user can rename inline, reorder by drag, toggle include/exclude, and add custom sections (tagged “Added by you”). Confirming locks the schema and opens the write gate (Q12). The FS_4 missing-instructions rule renders the amber gate below — never a silent preset fallback.
From your solicitation / Added by you) · include toggle. Excluded rows dim to 45% and are skipped (count as resolved per FS_4). Confirm writes the locked section_schema the writer consumes — section labels and ordering come from this list only, which also answers the Q48 sub-question. The eight-section default list exists only as the analysis fallback suggestion and is never used silently.The single checkpoint between analysis and generation. It shows exactly what the writer will use, surfaces the one thing worth fixing first (readiness), and sets expectations for the hybrid auto-approve first pass. This copy is final — ship it verbatim.
The Grant Writer will draft 8 confirmed sections for NSF STEM K-12, in order, using everything below. You review and approve each section — nothing is submitted anywhere.
Eligibility complete → Grant Writing (the FS_9 named transition). Title = “Ready to write your response?” Body = “The Grant Writer will draft N confirmed sections for [opportunity], in order, using everything below. You review and approve each section — nothing is submitted anywhere.” Checklist rows: Sections confirmed · Solicitation attached · Eligibility · Knowledge Center · Profile readiness (warn row links to the fixing skill, e.g. Story Collection). Note block explains hybrid auto-approve + [VERIFY] in two sentences. CTAs: secondary = Not yet — add impact stories (contextual to the warn row); primary = Start writing →. Contract variant swaps eyebrow to Eligibility complete → Contract Response and the body noun to “response volumes.”Two handoff patterns, both visible. Intake → anywhere: a full “Profile ready” card replaces the chat when intake confirms, itemizing exactly what was carried forward. Analysis → generation: a persistent context strip pinned at the top of the writer showing what it inherited — dismissible, recoverable from the toolbar.
.ho-strip in the writer: five key-values (sections, match, eligibility, source file, limits). Dismiss collapses it to a toolbar chip — the context never disappears, only the strip.From anywhere, the skill dropdown is always reachable (chat panel header, writer toolbar). Switching is named and announced — a switch banner states what's being left, what happens to it, and what's shared. Drafts always autosave; switching is never destructive. The only guard: a generation in flight.
Project Narrative is generating. Switch now and it finishes in the background — you'll find it in ready to review when you come back. Your draft and chat are saved.
(skill, opportunity). Mid-draft switch: autosave, banner with resume path; in-flight generation continues in the background and lands in ready_review without auto-approving. Returning to a skill resumes its thread exactly — no restart, no replay.The writer is one shell, two skills. Layout, state machine (not_started → generating → ready_review → approved + skipped/error), progress math, [VERIFY] footer, autosave, and export are identical. The active skill changes the vocabulary and switches three contract-only surfaces on. Section labels and ordering come from the confirmed schema (Q11) only — never hardcoded per skill.
| Labor category | Rate | Hrs | Ext |
|---|---|---|---|
| Sr. Bridge Inspector | $142/hr | 1,200 | $170,400 |
| NBIS Team Lead | $168/hr | 640 | $107,520 |
| CAD / Reporting | $96/hr | 480 | $46,080 |
.wr-secref mono prefix (L.4.2 / M.1) when the analysis extracted solicitation numbering. ③ Contract-only surfaces, flag-driven by skill: the GAP Unmet Requirements panel (separate from [VERIFY], per FS_4), section-number references, and the rate table component inside the Price Volume section. Grants show the [VERIFY] panel in the same slot. ④ Mixed grant + contract on one opportunity: not supported in one document. One response draft = one skill. An opportunity may hold two drafts (one per type), listed separately under All Drafts..wr-status — st-approved / st-review / st-gen / st-notstarted (+ skipped, error)
skill.flags.gap_inventory → Unmet Requirements panelskill.flags.section_refs → .wr-secref prefixskill.flags.rate_table → .rate-tbl in price sectionsgrant_generation
Document Rating requires a rubric — funder evaluation criteria extracted from the solicitation, a saved internal rubric, or one the user defines on the spot. No rubric → the skill asks before scoring, never invents criteria. The scorecard rates each criterion with evidence, and the overall score feeds the writer's per-section rating recompute.
.ru-score — 52px boxed total, ok/warn/danger border by band.ru-row — criterion + evidence note + bar + n/NThe full FS_9 open-question list, resolved by the designs above. Hand this table with the page to Cursor; each row maps to a directive in SKILLS_DS_DIRECTIVES.md.
| Q# | Question | Answer (design on this page) | Status |
|---|---|---|---|
| Q10 | Analysis card wireframe | Single card on opportunity detail: header + 4 stats + Qualifies/Risk/Gap factor rows + footer CTA. States analyzing/complete/error. Gap rows feed the shared Unmet Requirements inventory. #q10 | Answered |
| Q11 | Section confirmation UI | Dedicated confirm card: drag-reorder, inline rename, limits, source tags, include toggles, add-row. Missing-instructions amber gate replaces silent preset fallback. Confirm locks section_schema. #q11 | Answered |
| Q12 | Write-gate dialog full copy | Final copy shipped verbatim in the modal: eyebrow transition name, 5-row checklist, hybrid/[VERIFY] expectation note, contextual secondary CTA. #q12 | Answered |
| Q37 | Visible handoff moment | Intake → “Profile ready” summary card (6 carry rows + readiness gap + path-routed CTAs). Analysis → pinned carried-context strip in writer, collapsible to toolbar chip. #q37 | Answered |
| Q38 | Mid-flow skill switching | One grouped dropdown everywhere + writer toolbar chip. Shared: profile + KC. Independent: per-skill thread keyed (skill, opportunity). Autosave + named switch banner; in-flight generation finishes to ready_review. #q38 | Answered |
| Q48 | Writer UI: grant vs contract | One shell. Schema-driven labels/order only. Contract flags: GAP inventory, L/M section refs, rate table. Mixed grant+contract = two drafts, never one document. #q48 | Answered |