lucid

Lucid (Lucid Software / Lucidchart)

Current truth

• Active engagement: ChatGPT-app eval testing under a statement of work (SOW contact was Mallory, product marketing, now on maternity leave ~until Aug; her manager Harley is the interim commercial contact). Current SOW = one testing/optimization cycle (~5 rounds) + one submission. Baseline delivered 2026-06-18 — 100-prompt eval on the live ChatGPT host on GPT-5.5 (their best model): 66 pass / 32 partial / 2 fail. Now in baseline-review → prioritized-fix-list phase ahead of their second ChatGPT submission.
• Second ChatGPT submission imminent: triggered by switching auth from OAuth-only to mixed auth (enables unauthenticated users, nick-ridge’s workstream). Final pieces in QA as of 2026-07-02; target ~week of Jul 6 but hackathon + annual kickoff week may slip it. Tool count has grown from ~12–15 at submission #1 (April) to ~26; kelsey-nield holds the diff of added tools.
• Claude connector submission ~2 weeks behind ChatGPT. Nick confirmed interest in a separate SOW for managed Claude discovery optimization (Elliot to send proposal to Nick + Harley). We need their prompt list (~50, prioritized) 1–2 weeks before their Claude submission to build a week of baseline run data.
• RunScript architecture shift incoming (disclosed by apeksha-awasthi 2026-07-01): Lucid AI team building “RunScript”, a coding/scripting layer for multi-step canvas tasks (~100 tools) that will replace most existing MCP edit tools. Timeline per Apeksha: Lucid AI side done ~July, MCP integration ~August; existing tools stay live 3–4 months post-release while it’s validated. Creation stays separate for now (RunScript needs an existing document; one-shot “create from specification” outperforms multi-step creation in their testing). Fix-list prioritization principle agreed: focus on issues RunScript won’t solve and that risk churn in the transition window.
• Eval method: deterministic pass/fail (did the expected widget invoke?) plus a vision/model judge on five criteria — completeness, shape semantics, label legibility, connector quality, layout/spacing. Baseline run on the best model; weaker-model degradation testing still planned. Methodology aligned 2026-07-01: evals test what the product can do today (happy path + edge + clear negative cases), not aspirational scenarios; several partials were mislabeled.
• Scope is ChatGPT by agreement, but Lucid gets far more connector adoption on Claude than ChatGPT (Nick said 3–4x on 2026-07-02; the 2026-06-18 call had it at ~5x). Nick: “definitely more interested in Claude — doubling down on the success we’ve seen there.”
• [idea] Apeksha’s team pivots next quarter from tool-building to enterprise admin controls/governance (IP allowlisting, client restrictions, audit logging for MCP access) — per her 2026-07-01 read of her roadmap once RunScript absorbs tool work. Hard constraint she stated: Lucid enterprise policy prohibits storing or even viewing user prompts (even anonymized), so prompt-level usage telemetry is off the table; their workaround is 2–4 user interviews/week.
• Our benchmark data (shown 2026-07-02): Lucid ranked 3rd in Claude discoverability for the tracked category prompts, score trending up (~45 → 89, then ~84) with no Lucid-side changes in that window.
• Their benchmark standing (our data): 26 tools = 90th percentile across ~1,500 ChatGPT apps; second-longest tool descriptions of any app. Their “codified”/deterministic tool-description style is validated by Excalidraw’s similar approach. Their custom discovery tool is smart and frequently called — uncommon among comparables.
• What each side owes (as of 2026-07-02): Ghost — prioritized fix list from partials/fails (Vincent), Claude SOW proposal to Nick + Harley (Elliot), review of their updated tools/descriptions/submission copy once sent. Lucid — baseline review + prioritized critical-issues list with repro prompts (Apeksha, async before ~Tue Jul 7 sync), upcoming tool-change list (Apeksha), updated tools + all submission/marketing copy ~1 week before submitting (Nick/Apeksha), Claude prompt list (Nick).
• Evidence (Nick’s Snowflake chart, shared on the 2026-06-18 call): “adoption” = daily distinct users hitting beacon.accessedApiEndpoint for the embed/export actions (post/get.documents.documentid.embeds|export), Mar 24–Jun 12 2026. The dominant line (peaks ~1,800, strong weekday sawtooth, stepping up in early June) is Claude; the other (~100–200, flat) is ChatGPT — the chart’s legend doesn’t color-label them, but the high line matches both of Nick’s claims (the ~5x gap and the early-June spike). Their ChatGPT number is held down by the locked submission #1 tool set.
• Standing optimization-hypothesis buckets (proposed 2026-06-18, feed the fix list): (1) sharpen tool-description wording/syntax (small changes like “must” measurably shift model behavior), (2) improve render output (text-label overlap, right-side cutoff; optional mobile-viewport testing), (3) a pre-tool reference/readme or chain-tool pattern to guide the model across 26 tools (trade-off: ~1s added latency).
• Contacts mapped: nick-ridge (growth PM, owns the registration funnel / unauth-user mode and the second submission), apeksha-awasthi (PM for the MCP team — main point of contact for MCP server changes), ryan-ferguson (developer who makes the MCP changes), kelsey-nield (unauthenticated-user-mode launch; holds the tool diff). Commercial contacts: Mallory (product marketing, SOW contact, on maternity leave) and her manager Harley (interim).
• Live upsell thread: ADO discovery. We track Lucid daily on agentdiscoverability.com in the charts/diagram/visual-collaboration category. Pitched our method: track category + brand-specific prompts daily, reconcile against MCP server logs to estimate discovery-channel users, optimize for the connector picker. As of 2026-07-02 this has concretized into the pending Claude SOW.

Open questions

• Is the engagement paid? An SOW exists (Mallory was the contact), but paid/continuation framing still not explicitly closed (Q2 goal: get Lucid “onto platform, priced higher”). The Claude SOW is the live expansion vehicle.
• Budget owner: Harley (Mallory’s manager) is the emerging commercial contact — Nick said he’d talk to him about the Claude SOW. Role/seniority unconfirmed.
• Which new MCP capabilities were auto-inherited by the ChatGPT app vs which require the resubmission to activate? (Apeksha to confirm; Nick flagged it as the key open item on 2026-07-02.)
• Persona split for the prompt set: existing authenticated users vs top-of-funnel unauthenticated users may need different intent sets — casual/low-monetization (brainstorming, mind mapping) vs high-value professional (network diagrams, BPMN). Long-standing open question for Lucid’s marketing team.
• Which baseline failures survive RunScript? (The basis for Vincent’s prioritized fix list — needs Apeksha’s upcoming tool-change list.)
• Product-marketing attendee (listings owner) and “Alicia/Alyssa” on the organic-discovery side — names/roles not yet pinned. Unmapped invitees: jbrown@lucid.co, abadger@lucid.co, mnielsen@lucid.co.
• Early-June Claude adoption spike (+30% WoW, step-function) has no confirmed cause — no known Claude-side change after the Apr 23 organic-discovery launch. Candidates: ranking shift, “popular” registry tag, organic word-of-mouth. Nick wants an attributable event for leadership.
• Does ChatGPT offer a tools-only re-submission, or does updating tools require the full submission flow? Partially answered 2026-07-02: changing the auth setting (OAuth → mixed) cannot be done without a resubmission (per Nick, “could be wrong”).

Timeline

• [2026-07-02] (call) Bi-weekly sync (recorded by both founders; Vincent, Elliot, Apeksha, Ryan Ferguson, Nick Ridge — Nick joining from PTO at a wedding). Resubmission scoped: initial app submitted ~April; resubmission triggered by OAuth → mixed-auth switch for unauthenticated users; final pieces in QA, target ~next week but hackathon + Jul 4 may delay; tools grown ~12–15 → ~26 (Kelsey has the diff, working with Augustina); open item = which MCP capabilities auto-inherit vs need resubmission (Apeksha to confirm). Ghost’s submission role confirmed as part of the current SOW: review all updated tools/descriptions/marketing copy, run Lucid’s intents through our discovery platform, suggest keyword/text optimizations, and flag common submission errors to cut OpenAI review cycles; sequencing agreed — no need to do Claude first, optimize the ChatGPT submission now with existing Claude-derived data. Elliot presented the ADO framing (SEO → ASO → ADO; store search + organic surfacing; ~30k prompt runs backing keyword→pick attribution) and showed the benchmark: Lucid 3rd in Claude discoverability, ~45 → 89 → ~84 with no Lucid-side changes. Nick: Claude adoption 3–4x ChatGPT; motivation for unauth mode = zero-click search eroding marketing-site traffic (“if people are going to AI instead, how can we get into AI”); confirmed interest in a separate Claude SOW — Mallory (prior SOW contact) on maternity leave ~a month, loop in her manager Harley. Vincent on submission mechanics: ~5–6 page process, golden prompt set = 3 example + 3 negative prompts (for OpenAI’s test team, not optimization), now auto-reviewed/approved in ~1–2 days (was 1–3 months). [idea] Elliot’s stated prediction (his word) on ChatGPT organic discovery: currently only among already-connected apps; organic surfacing to unconnected users likely follows the Codex-under-ChatGPT merge + ads rollout. ~2,000 ChatGPT apps approved (±40), ~30–40 approvals/day. After Nick dropped: Vincent asked Apeksha for two async items before a ~Tue Jul 7 sync — (1) baseline review (edit the sheet directly or send bullets; “50 microservices” agreed pass, the DB-pull case agreed fail), (2) prioritized critical issues with repro prompts; her availability limited by Jul 4 + hackathon week.

• [2026-07-01] (call) 1:1 baseline sync, Vincent + apeksha-awasthi. Eval methodology aligned: evals should test what the product can do today (happy path, edge cases, clear negative cases), not dream scenarios; several partials mislabeled — the “50+ microservices” prompt (model grouped logically instead of literally) agreed borderline-pass; the “pull live DB data via MCP” case agreed a true fail because the model attempted a workaround instead of stating the limitation upfront — Apeksha named that limitation-honesty failure mode as the pattern she most wants caught. Jake recently added data linking + custom fields to the MCP (custom shape data now readable), so that specific case may now pass. RunScript disclosed (all detail from ~the week before this call): Lucid AI team building a coding/scripting tool for chained multi-step canvas operations, ~100 tools, replacing most edit tools; Lucid AI side ~1 month out, MCP integration the month after; existing tools live 3–4 months post-release; creation tools stay separate (RunScript needs an existing document; one-shot create-from-specification beat whole-board multi-step creation in her own testing). Agreed the baseline stays valuable as a before/after benchmark across the RunScript transition, and that fix prioritization should target issues RunScript won’t solve + churn risk in the window. Apeksha’s team pivot: next quarter focus = enterprise admin controls/governance (IP allowlisting, client connect restrictions, MCP audit logging). Vincent floated capturing user-question keywords via tool schema (OpenAI allows it in practice); Apeksha ruled it out — enterprise policy: no storing/viewing prompts, even anonymized; their insight workaround is 2–4 user interviews/week. Codex plugin skills were written in ~a week from known customer flows + marketing submissions; she expects they’ll need a more robust process; Vincent offered Ghost could spend time testing the Codex plugin. Next steps: Vincent to finish reviewing partials/fails → prioritized fix list; Apeksha to share the upcoming tool-change list; continue next day with Elliot + Nick (resubmission, unauthenticated demo-mode testing, whether Ghost supports the new-user-acquisition effort).

• [2026-06-29] (internal discussion — Ghost Team stand-up, Vincent + Elliot, dual recording) Deliberate slow-roll. Engagement is stalled on apeksha-awasthi — she messaged Friday she’d review the baseline report “today” (06-29); Vincent still needs her input on the partial approvals (can’t tell pass vs fail without it). Read: her pace reflects how slow the org is, not disinterest (Elliot has seen it pre-engagement; landing the project took ~a month). Plan: don’t push — use their slowness to buy time; propose moving the Thursday call to next week (draft a light message together, citing the short turnaround after her feedback) so Vincent isn’t scrambling. Framed as the normal reality of client work that the managed-service side will always carry.

• [2026-06-18] (call) Bi-weekly sync (recorded by both founders). Baseline eval delivered: 66/32/2 on GPT-5.5 in the live ChatGPT host; most partials judged to be miscategorized (ambiguous eval prompts, e.g. “make this chart accessible” with no chart → model asks a clarifying question → counts partial). Benchmarked at 90th percentile / second-longest tool descriptions across ~1,500 apps; codified style validated vs Excalidraw. We proposed three optimization buckets (descriptions, render, readme/chain pattern). Lucid disclosed Claude adoption is ~5x ChatGPT (new tools auto-update in Claude but are locked to submission #1 on ChatGPT marketplace) and a ~30% WoW Claude spike in early June. Lucid committed to a second ChatGPT submission ~2 weeks out — unauthenticated user mode (Figma-style, no-account diagram generation, growth-driven), engineering done ~mid next sprint, bundling all new tools since submission #1. We shared our Claude organic-discovery tracking + connector-picker explanation. Roles mapped (Apeksha=MCP PM/main contact, Ryan=dev, Kelsey=unauth mode, Nick=growth). The 100-slide eval deck/Excel was shared in the Lucid Slack channel (artifact lives there + Granola, not in the brain).

• [2026-06-09] (call) Bi-weekly sync. Prompt eval set ~100 prompts shared in Slack; 3 tools missing (lucid_fetch_item_image, lucid_submit_feedback, get_mcp_resource) — Nick to add 1 prompt each by Thursday. Weekly cadence agreed: baseline this week, results reviewed next week. They’re submitting a Codex plugin this week. Nick shared user-interview findings (layout #1 complaint, personalization gap, persona split). We demoed our discovery tracking on their category; Nick asked for separate ADO call. We shared intel: Codex–ChatGPT “super app” integration expected, organic discovery live in Claude and expected in ChatGPT within months. Agreed next week: define critical paths vs acceptable free-tier degradation.

• [2026-06-10] (internal discussion) Strategy: pace deliverables (don’t reveal all fixes week 1), aim for continuation ideally paid, tease discovery optimizations as upsell. Need to reach a more senior contact.

• [2026-06-09] (internal discussion) Stand-up before the Lucid sync: Vincent to merge Lucid’s comments into the prompts Excel and “validate that Excel today, and that’s it” — then figure out how delivery work starts.

• [2026-06-08] (internal discussion) “Lucid will hit us tomorrow” — engagement work starting 06-09. Elliot: delivery should ultimately need minimal founder involvement (we manage/lead/close; someone else delivers) — recruitment trigger if more such engagements land (“three Lucids would make hiring top priority; we have one”).