Second Brain — an AI executive secretary vault

A file-based knowledge vault operated by an AI agent (Cline) acting as your executive secretary. You dump raw material and talk; the agent files, indexes, and cross-references everything; you ask questions and get cited answers; you request reports and get polished, offline-ready HTML/PDF documents whose formatting improves with your feedback.

Built for a senior engineer / architect / group lead: team dossiers and performance evaluations, meeting minutes and prep packs, decision records, project RAG rollups, commitments tracking, interview debriefs, and a brag log that turns into CV bullets.

Inspired by Karpathy's LLM wiki (agent-maintained markdown + indexes) and DOX (small root docs, detail in leaves — so navigation files never outgrow a weaker model's reliability).

Quick start

1. Get a private copy. This repo is a data-free template. Create your private instance from it (GitHub: Use this template), keep it private — it will hold sensitive people data. Optionally add the template as an upstream remote to pull future improvements.
2. Requirements: Cline (VS Code extension); python3 (stdlib only, for the health check and hooks); Chrome/Chromium (optional, for agent-side PDF export).
3. Open the folder in Cline. Rules in .clinerules/ load automatically; hooks and skills are picked up from .clinerules/hooks/ and .cline/skills/.
4. Start using it: drop files into inbox/ and say /inbox.md — or just tell the agent about your team, meetings, and projects. Ask it anything later.

How it works

Three layers, strictly separated:

inbox/    →   library/               →   memory/
your dumps    originals, untouched,      the agent's synthesis:
              renamed + catalogued       dossiers, minutes, projects,
              by year                    decisions, notes — indexed + logged

• Originals are sacred. Files are moved (never edited) into library/YYYY/ with date-prefixed names. A PreToolUse hook hard-blocks agent edits to them.
• Everything is indexed. Navigation is a two-level tree: root files (memory/index.md, library/catalog.md) list only children with counts; entries live in leaf indexes/catalogs. A mechanical 300-line rule triggers deeper sharding, so no navigation file ever grows past what a small model handles reliably.
• Everything is logged. memory/log.md is an append-only journal of every ingest, filing, and report.
• Every fact is dated, answers cite their source pages, and recorded fact is kept separate from inference.

Daily use

You say	The secretary does
"Sam shipped the migration two weeks early"	Files it (dated) into Sam's dossier's accomplishments log
"/prep.md — steering meeting with KDDI folks tomorrow"	Prep pack: attendee one-liners, open threads, landmines, talking points
"/one-on-one.md Sam"	1:1 arc: recognise / follow up / probe / goal pulse / carry-over
"Draft Sam's performance eval"	Evidence-based draft from goals + dated entries; flags gaps loudly
"/project-status.md"	Portfolio RAG rollup; stale statuses report as "unknown", not old green
"/brief.md"	Overdue reminders/commitments, upcoming dates, loose ends
"Make that report a deck instead" + feedback	Report re-done; your formatting persisted as a template for next time

Workflows (type / in Cline)

Command	Does
/inbox.md	Process everything in inbox/ into library/ + memory/
/remember.md	File facts you just said in chat
/recall.md	Answer a question with sources cited
/meeting.md	File meeting minutes; fan out decisions/actions/people facts
/prep.md	Meeting prep pack
/one-on-one.md	1:1 prep and post-1:1 capture
/decision.md	Record a decision (ADR-style: options, rejections, revisit date)
/project-status.md	Single-project deep dive or portfolio RAG rollup
/cv-update.md	Turn the principal's brag log into CV-ready bullets
/brief.md	Daily brief
/weekly.md	Weekly review: sweep, nudge list, people pulse, look ahead
/handoff.md	Out-of-office handoff and re-entry brief
/report.md	Generate a report (markdown or HTML), template-aware
/reindex.md	Rebuild the navigation tree from disk
/checkup.md	Health check + judgment-level hygiene

Skills (auto-trigger or type /name)

Skill	Does
performance-evaluation	Evidence-based eval drafts from dossier + goals + meetings
interview-debrief	Structured candidate feedback vs your rubric; HR-sensitive handling
report-builder	The template-evolution loop for recurring reports
lazyway-io-design	Rules for lazyway.* HTML reports (kit's own skill)
claude-report-design	Rules for claude.* HTML reports (light/dark kit)

Hooks (.clinerules/hooks/, Python, fail-open)

• TaskStart — injects a briefing into every session: inbox backlog, overdue reminders/commitments, unresolved ⚠ markers.
• PreToolUse — blocks any file-edit tool targeting library/ originals (catalogs excepted). Moving/renaming via shell stays allowed.

Reports, design kits, PDF

HTML reports are fully offline — kits are vendored under reports/design/<kit>/ with self-hosted fonts; no CDN, no CORS/referrer issues opening from disk. Kits: lazyway (IBM Plex, brand blue, one amber accent) and claude (warm ivory/charcoal light+dark, terracotta, persisted theme toggle, print always light). Index and add-a-kit recipe: reports/design/README.md.

Three template shapes per kit in reports/templates/ — document (Word-style A4), deck (PowerPoint-style 16:9 slides), interactive (responsive one-pager). Export to PDF via the browser's Print → Save as PDF, or scripts/export-pdf.sh <report.html> (headless Chrome). Formatting feedback is persisted to reports/templates/ so it's never given twice.

Maintenance

• python3 scripts/health.py — mechanical invariants: inbox backlog, shard placement, catalog/index tree consistency and counts, broken links, navigation size rule, overdue items, stale dossiers. Run by /checkup.md and /weekly.md.
• /reindex.md — rebuilds the tree when drift is found.

Repo layout

Path	What	Written by
inbox/	Raw dumps, unprocessed	You
library/YYYY/	Originals, byte-for-byte untouched, per-year catalogs	Agent (move/rename only)
memory/	People, meetings, projects, decisions, notes, ideas, topics + root index + append-only log	Agent
reports/	Generated reports (YYYY/), templates, design kits	Agent
scripts/	health.py, export-pdf.sh	—
.clinerules/	Always-on rules, workflows/, hooks/	You + agent
.cline/skills/	On-demand skills	You + agent

For AI agents (non-Cline)

If you are an agent other than Cline working in this repo: read .clinerules/00-role.md and .clinerules/10-structure.md before changing anything. The invariants bind you too — never modify files under library/ (except catalogs), never rewrite memory/log.md history, update indexes + log with every filing, and run python3 scripts/health.py before finishing.

Privacy

The private instance will hold performance evaluations, attrition risks, and interview feedback. Keep it private; don't sync it anywhere you wouldn't email it. Initialise git only after confirming no personal data is present (this template ships clean).

License

Template: 0BSD — use it for anything, no attribution needed. Vendored material (IBM Plex fonts under SIL OFL 1.1; the lazyway design kit): see THIRD_PARTY_NOTICES.md.