Customer Intelligence Report

Customer Intelligence Report

Turn a period of Fireflies meetings and Circle community posts into a recurring intelligence report: what themes keep coming up, where people are in pain, which technical patterns recur, and who is stuck. The deliverable is a BenAI-branded instant-ui HTML dashboard rendered from a data file, so it is customizable and never static (see references/report-config.md).

Secrets: the Circle and Fireflies connectors authenticate through the MCP connection. Never put an API key, token, space id, or email into the skill as a hardcoded secret, all targeting values come from the run config. If you add publishing, use an env-var token and a gitignored deploy dir (see report-config.md, Secrets and safety).

Steps

Track progress:

Task Progress:
- [ ] 1. Scope: confirm scope, cadence, sections (checkpoint)
- [ ] 2. Collect: pull the period from Fireflies and Circle
- [ ] 3. Analyze: cluster transcripts into themes
- [ ] 4. Assemble: build the WEEKS + LEXICON data
- [ ] 5. Build: render the instant-ui dashboard

1. Scope (checkpoint, before any heavy pull)

Confirm the run with the user. Present the options and wait for a pick, do not assume. Read references/report-config.md for the choices. Confirm: sources (Fireflies, Circle, or both, plus their targeting: organizer emails and Circle space ids), scope (all / person / team), cadence (weekly / daily / custom), the trailing recurring window, and which sections to include. Offer the default section catalog plus 3 to 5 optional sections so they can shape it. Targeting values are config, never hardcoded secrets.

2. Collect

Pull the period from the configured sources. Both are optional and degrade gracefully: if one errors, continue with the other and note the outage.

• Fireflies meetings: read references/fireflies-guide.md.
• Circle posts: read references/circle-guide.md. List meetings and posts first and report the counts to the user, then fan the transcript and post reading out to subagents that return small structured notes.

3. Analyze

Cluster the notes into themes. Read references/analysis-framework.md. Assign each theme a section, canonicalize its id against the LEXICON, compute counts, streak, and resolved, and detect stuck members. Every theme needs a real verbatim quote or it is cut.

4. Assemble

Shape the analysis into the data contract. Read references/data-schema.md. Produce valid JSON for WEEKS (append this period to prior periods if they exist) and update LEXICON with any new phrasings. Present a short chat summary of the top themes and stuck members.

5. Build the dashboard

Render the report in the BenAI instant-ui design language. Load the embedded guides first: read references/instant-ui/design-tokens.md (the :root token block and typography rules), references/instant-ui/page-shell.md (gradient stripe, dark BenAI header, smiley SVG, breakpoints), references/instant-ui/components.md (copy component CSS/HTML verbatim if a section needs a block beyond the template), and references/instant-ui/build-rules.md (hard constraints and the after-build checklist). Then read references/dashboard-template.html, which is already built in this language: set the config constants and SECTIONS, paste in the WEEKS and LEXICON JSON, and write the filled file to customer-intelligence-report-YYYY-MM-DD.html in the working directory. The dashboard must be self-contained and offline, and must keep the instant-ui look: cream #fffef8 background, 3px solid #111 borders, hard zero-blur shadows, square corners, weight-900 headings, SF Mono for every number and label, the 4px gradient stripe, and the BenAI smiley in header and footer. Run the after-build checklist before confirming. No em dashes in the rendered copy. Never invent a number or quote, every figure traces to a real Fireflies or Circle result.

Self-improvement

This skill is never finished. Improve it as you use it.

• When the user corrects a section, a theme call, or the scope handling, update the relevant reference file (or this SKILL.md) so the correction sticks.
• When a correction is a hard rule ("always", "never"), add it as a permanent rule here.
• Grow the LEXICON every run: fold new phrasings into existing canonicals or add new emerging entries, so recurrence tracking improves over time.
• When the user says a report was genuinely good, save its data file to references/examples/ as a model for future runs.
• Keep the skill small: when you add something, run the deletion test and cut anything that no longer changes behavior.

Routing