v2.1 · Opus Architecture Review · June 2026

A personal AI
operating system
for creative work.

Every Claude session used to start cold — no memory of what happened yesterday, what decisions were live, or what patterns were building over weeks. This system fixes that. Persistent memory, daily rituals, a synthesis engine that compounds over time, and a quarantine layer that ensures observations are always reviewed before they shape strategy.

4
Memory layers
5
Installed skills
5
Synthesis moments
30+
Context files
11
Design principles
Overview

Three components. One compound system.

Memory makes Claude informed. Rituals make it proactive. Intelligence makes it generative. All three feed each other continuously — and each component gets smarter as the others do.

📂
Memory
Persistent knowledge in files
Your permanent identity Who you are, how you work, what applies to every session — always the first thing loaded. profile.md
Your Floward context Role, chain of command, session checklist, where everything lives. Stable. Rarely changes. CLAUDE.md
What's happening now Tasks, risks, today's brief, synthesis inbox, done log. Rebuilt every day. Where sessions begin. this-week.md · state-of-play.md · today.md · synthesis-inbox.md
Deep expertise on demand Photography project, AI strategy, team profiles, decisions. 20+ files with archive pairs — none loaded until needed. ai-photography-handover.md · team-org.md · 18 others + 3 archives
feeds ↓
⚡
Rituals
Installed skills that run daily
Morning orientation Adjudicates synthesis inbox first, then reads all layers + live signals. Writes today's brief. morning-brief skill — Sonnet session
End-of-day close Closes open loops, routes updates to their domain files, stages synthesis to inbox silently. eod-brief skill — Opus session
On-demand intelligence Skips task tracking — reads everything, surfaces connections, stages to inbox. /synthesis mode inside eod-brief
New workstream kickoff Pre-reads all context, pre-fills the interview, creates the project file in one step. new-project skill
One-command publishing Ships any HTML output to a live URL via GitHub and Vercel. publish-vault skill
produces ↓
🧠
Intelligence
Synthesis that compounds weekly
Synthesis inbox EOD stages observations here. Morning adjudicates — confirm, edit, or discard. Nothing reaches live files without review. synthesis-inbox.md — hard cap 8 entries
Strategic nucleus Observations accumulate into patterns. Patterns graduate to principles. Explains why things work — never expires. ai-strategy-nucleus.md · nucleus-archive.md
Ideas pipeline Each idea has a trigger condition. The pipeline pays out at the right time — not a backlog. ideas-pipeline.md · ideas-archive.md
Pre-meeting synthesis When a high-stakes meeting is detected (Asil, Al-Arifi, probation), morning runs a scoped 3-op synthesis into today.md automatically. built into morning-brief — automatic trigger
writes back to files ↑

The daily closed loop

☀️ /morning
→
Adjudicates synthesis-inbox first (confirm / edit / discard). Reads all layers. Pulls live signals. Writes today.md. Pre-meeting synthesis if Asil/Al-Arifi/probation detected.
Day of work
→
Mid-session reflex fires when significant new intel arrives. /synthesis available on demand. /checkpoint writes state if context fills.
🌙 /eod
→
Updates all context files. Stages synthesis observations to synthesis-inbox.md silently (no confirmation required at EOD). ✅ completions logged to done-log.md.
📥 synthesis-inbox.md new
→
Quarantine layer. Hard cap 8 entries. EOD writes here, morning clears here. Nothing reaches live files without morning review.
ideas-pipeline.md · ai-strategy-nucleus.md
←
Only confirmed observations land here. Entries accumulate, patterns graduate to principles, shipped ideas move to archive pairs.
Why the quarantine layer?

EOD is the highest-context moment for Claude — but not always the best moment for evaluation. Morning is better: fresh, no synthesis bias, just "does this hold up?" Separating generation (EOD) from confirmation (morning) prevents both overclaiming and idea loss. If the inbox hits 8 entries, EOD stops writing until morning clears it — deliberate backpressure so nothing gets buried.

Memory Architecture

Four layers. Loaded in order.

Context loads progressively — only what the task needs, when it needs it. Loading everything upfront degrades output quality by 15-25%. The system is deliberately stingy about what it reads at session start.

Layer 0 — Global identity
Your permanent operating instructions
Who Nour is. How he works. What applies to every session: writing rules, clarification behavior, email documentation discipline, filter contract, synthesis confirmation rule (EOD writes to inbox; morning confirms), and the mid-session reflex. Written once. Changes only when behavior changes.
profile.md
Always — first
Layer 1 — Project identity
Your Floward context
The stable frame around this specific role: title, chain of command (Nour → Omar → Asil → Al-Arifi → CEO), session-start checklist, full file map, top priorities. Session-start step 5 now reads synthesis-inbox.md first, then uses today.md surface list if available — spares direct pipeline/nucleus scans on most mornings.
CLAUDE.md · _index.md · system-guide.md
Always
Layer 2 — Session state
What's actually happening right now
Task priorities (⚠️ today / ⏳ this week / ✅ done), active risks, flags, pending briefings, today's brief, and two new files: synthesis-inbox.md (EOD observations staged for morning review, capped at 8) and done-log.md (append-only completion record for probation). These files change daily — reading them is what makes each session start informed rather than cold.
this-week.md · state-of-play.md · today.md · synthesis-inbox.md · eod-log.md · done-log.md
Every session
Layer 3 — Domain knowledge
Deep expertise, loaded when needed
30+ specialized files covering the photography project, AI strategy, team intelligence, decisions, probation milestones, campaign calendar, writing standards, and the playbook. Each live file has an archive pair — when entries are superseded or completed, they move to the archive instead of bloating the active file. No session loads all of these — each loads only when the task explicitly requires it.
ai-photography-handover.md · ai-strategy-nucleus.md · nucleus-archive.md · ideas-pipeline.md · ideas-archive.md · decisions.md · decisions-archive.md · playbook.md · team-org.md · 20+ others
On demand only
Active / archive split

Every intelligence file that grows unboundedly gets an archive pair. The active file has a soft cap maintained by monthly consolidate-memory. Superseded entries move to the archive pair — never deleted, never re-read unless explicitly needed. This keeps active-session context tight while preserving the full institutional record. The done-log.md is the exception: it only appends, never compresses — it's the probation case file.

Ritual System

Daily habits. Automated.

Most AI use is reactive — you describe a problem, Claude helps with that problem. Rituals change the model. The morning brief adjudicates last night's observations before you even see the brief. The EOD brief closes loops you might not have noticed were open. Synthesis fires on a schedule without being requested.

Daily closed loop · Sonnet → Opus → Sonnet
Morning · every session start · Sonnet
/morning /resume /checkpoint

Morning Brief

Adjudicates synthesis-inbox first (confirm, edit, or discard each EOD observation). Then reads all four layers, pulls live signals from Slack, email, and calendar. Detects high-stakes meetings and runs pre-meeting synthesis automatically. Surfaces unanswered questions from last night first.

Synthesis-inbox adjudicated — confirmed entries routed to live files
Today's non-negotiables in priority order
Pre-meeting synthesis if Asil/Al-Arifi/probation detected new
Unanswered EOD questions surfaced first
The one thing that most needs to go well today
Evening · every session close · Opus
/eod /synthesis

EOD Brief

Checks sent mail (not received — sent items reveal actual commitments made). Asks targeted questions. Routes every update to its correct domain file. Stages synthesis observations to synthesis-inbox.md silently — no confirmation at EOD time. Logs completions to done-log.md.

All context files updated and accurate
Synthesis observations staged to synthesis-inbox.md new
Completions logged to done-log.md new
On Fridays: extended weekly synthesis (all 3 rotating ops)
On demand · any time · Opus
/synthesis

Synthesis Pass

No task tracking. No questions. Just intelligence. Three-phase framework: (A) Diverge — structural sweep, personified bisociation, inversion. (B) Attack — mandatory disconfirmation pass, stakeholder divergence, reaction pass, outside view. (C) Filter — two-axis worthy filter (novelty AND consequence). Observations staged to inbox.

Phase A: structural gaps, multi-hop chains, bisociation prediction
Phase B: beliefs stress-tested against their own evidence
Phase C: 2-5 observations + 1-3 independent ideas, staged to inbox
Session continuity pair
Mid-session · when context fills · Sonnet
/checkpoint /save

Checkpoint

One question, one write, under two minutes. Writes a timestamped block to state-of-play.md under Mid-session Checkpoints. Keeps the last three checkpoints. No synthesis, no email pull, no routing. This is the bridge that makes /resume precise rather than generic.

Exact thread captured: in-progress, done, decisions, next step
Next session can /resume instead of re-reading everything
When picking up a paused thread · Sonnet
/resume

Session Resume

Reads all files (Steps 1-3 of morning brief), skips the live signal pull entirely, focuses synthesis on the most recent checkpoint in state-of-play.md. Delivers a compact "where you left off" summary. Ends with a prompt to paste the last few exchanges from the previous session — files give strategic context, chat history gives the tactical thread.

Synthesis-inbox still adjudicated at session start
Compact "where you left off" rather than full brief
Project tools
When starting any new workstream
/project

New Project Kickoff

Reads all existing context before asking a single question — so it already knows Floward's priorities, active risks, team relationships, and what similar projects looked like. Pre-fills the kickoff interview with confidence ratings. You confirm or correct. Writes the project file, updates the task table, surfaces relevant intelligence from nucleus and pipeline.

Pre-filled interview based on what's already known
Project overview file created and indexed
Task section added to the weekly tracker
When sharing any HTML output
/publish floward /publish personal

Publish to Vault

Any HTML file goes live as a permanent shareable URL in one command. Commits to a private GitHub repository, triggers a Vercel deployment, verifies the link is live, returns the URL. Two vaults: Floward work outputs and personal projects.

Live URL at floward-kappa.vercel.app or personal vault
Committed to private GitHub repo for version history
Permanent link — shareable with anyone, no login needed
Intelligence System

Creative synthesis, turned into infrastructure.

Five specific moments. Three synthesis phases. One quarantine layer. A pre-meeting trigger. The system converts intuition into a formal, repeatable process — and a review gate that ensures only tested observations shape strategy.

Moment 1
Daily · lightweight

Session start — synthesis-inbox adjudication

The first active step of every morning brief: review synthesis-inbox.md. Each pending EOD observation gets a verdict — confirm (routes to recommended file), edit (update then route), or discard (removed). Observations from the prior session reach live files only after this gate. If nothing is in the inbox, the brief continues normally.

Moment 2
Triggered · targeted

Mid-session reflex

Encoded as a standing instruction: when significant new intel arrives mid-conversation, Claude immediately reads the pipeline and nucleus and asks what this fires, validates, or contradicts. Significant means: does this change what Nour would do? If yes, surface it. If not, say nothing. A full session with zero reflex output is unusual and worth noting.

Moment 3 new
Triggered · targeted

Pre-meeting synthesis

When morning brief detects a high-stakes calendar event (Asil, Al-Arifi, or a probation review), it automatically runs a scoped three-operation pass — B1 disconfirmation (what belief should be stress-tested before this meeting?), B2 stakeholder divergence (what does each participant define as success?), and an asymmetry pass (what is the riskiest move in this meeting?). Output: 3-5 bullet points written directly to today.md under a pre-meeting section. No command needed.

Moment 4
Daily · comprehensive

EOD synthesis — three phases, inbox-staged

After all context files are updated: Phase A (Diverge) — structural sweep for missing edges and non-obvious chains, personified bisociation with rotating practitioners (prediction-forcing, not analogy), inversion. Phase B (Attack) — mandatory disconfirmation of every load-bearing belief, stakeholder divergence, reaction pass, outside view. Phase C (Filter) — two-axis worthy filter (novelty AND consequence, both required). Output staged to synthesis-inbox.md silently. Disconfirmation observations get automatic novelty-clearance — the additive bias of the framework cannot suppress them.

Moment 5
Weekly · generative

Friday extended synthesis

Runs automatically at the end of every Friday EOD brief. Steps back from the whole week — not just today. Runs all three rotating operations (negative-space, asymmetry, decay) rather than one. Asks what's in the pipeline with no corresponding nucleus pattern, and vice versa. Explicitly looks for whitespace. Generates 2-4 new pipeline ideas from scratch and stages them to the inbox.

The synthesis inbox in practice.

EOD generates silently. Morning adjudicates. Nothing reaches live files without a review gate.

Synthesis inbox flow — quarantine pattern
🌙
EOD generates

Observations written to synthesis-inbox.md silently. No confirmation needed. Up to 8 entries maximum.

📥
Inbox holds

Hard cap of 8. If full, EOD stops writing and flags it. Morning must drain before fill resumes.

☀️
Morning adjudicates

First active step. Each entry: confirm → route, edit → update then route, or discard → remove.

✓
Live files receive

Only confirmed observations land in ideas-pipeline.md or ai-strategy-nucleus.md with [AI-surfaced — review] tag.

Two files. Separate by design.

The nucleus explains why things work. The pipeline holds what to do about it. Each has an archive pair — active files stay tight, the full record is always preserved.

AI Strategy Nucleus

Strategic intelligence layer

Patterns, observations, and principles that explain how Floward operates. Grows from lived experience through the Emerging Patterns staging area. Low barrier to add, high barrier to graduate. Never expires. Superseded entries move to nucleus-archive.md.

EOD observation staged to synthesis-inbox.md
Morning confirms → lands in Emerging Patterns [AI-surfaced — review]
Evidence accumulates across weeks and sessions
Monthly review promotes to Strategic Principle → superseded entries → nucleus-archive.md
Ideas Pipeline

Ideas bank with trigger conditions

Specific things to build, run, or do — from Nour's thinking and from Claude's synthesis. Each entry has a trigger condition: the right moment for it to be surfaced and actioned. Status: Raw → Parked → Triggered → Active → Shipped. Shipped ideas move to ideas-archive.md.

AI-surfaced idea staged to synthesis-inbox.md
Morning confirms → lands in pipeline as Raw with [AI-surfaced — review]
Trigger condition fires → status moves to Triggered
Shipped → ideas-archive.md (full record preserved)
Design Principles

Eleven decisions that shaped everything.

These constraints apply to every part of the system. When something needs to change, check it against these first — most proposed changes violate at least one. The four marked "new" were added in the June 7 Opus architecture review.

01

Progressive disclosure

Context loads in tiers. Only what the task needs, when it needs it. Loading everything upfront degrades quality by 15-25% — so the system is deliberately stingy about what it reads at session start.

02

Projects vs Areas

Projects end. Areas don't. Treating an ongoing responsibility as a project creates a permanent sense of being behind on something that has no finish line. The classification changes how work is tracked and reported.

03

The routing rule

Every update goes to its natural home — never to the central identity file. The identity file holds a one-line pointer. The domain file holds the detail. This is what prevents context clutter from compounding over time.

04

Synthesis is expected

A session with zero synthesis output is unusual and worth noting. The default posture is: always looking for what to add, not waiting to be asked. This converts Claude from a question-answering tool into an intelligence layer.

05

The 64-occasion multiplier

Every workflow runs 64 times a year across Floward's occasions calendar. Always state the per-occasion saving AND the annualized number. The per-occasion number feels small. The annualized number makes the ROI case.

06

AI as enabler, always

Every output frames AI as an enabler of human craft — never a replacement. The human does the storytelling. AI handles production mechanics. One wrong framing in a leadership presentation undoes weeks of relationship-building.

07

Files over conversation history

State lives in files, not in chat. Files are readable, editable, auditable, and portable across any Claude interface or model. Conversation history is ephemeral and unpredictable — it cannot be the foundation of a persistent system.

08 — new

Active / archive split

Every intelligence file that grows unboundedly gets an archive pair. Active files have a soft cap. Superseded entries move to the archive — never deleted, never re-read unless explicitly needed. Active sessions stay fast; institutional memory stays complete.

09 — new

Synthesis-inbox quarantine

EOD generates observations silently to a quarantine file. Morning adjudicates them. Nothing reaches live strategy files without a review gate. Hard cap 8 entries creates deliberate backpressure — the inbox cannot quietly fill without being noticed.

10 — new

Domain file header standard

Each Layer 3 file leads with a four-part header: Status, Recent, Open items, Detail. A session can read just the header (8-15 lines) to get the domain picture. This enables true progressive loading — detail is earned, not assumed.

11 — new

Pipeline status taxonomy

Every pipeline idea has a status: Raw → Parked → Triggered → Active → Shipped. Off-ramps: Dormant, Archived. Without clear status labels, the pipeline is a flat list that never gets acted on. Status makes every idea explicitly actionable or explicitly not-yet-actionable.

Command Reference

Every slash command. Where it lives.

All commands are entry points into an installed skill. The skill handles everything — you just need the right trigger phrase and the right session.

morning-brief skill
Open in your Sonnet session. Run at the start of every work day and for mid-session continuity.
Sonnet
/morning Full morning brief. Adjudicates synthesis-inbox first. Reads all layers. Pulls Slack, email, calendar. Runs pre-meeting synthesis if high-stakes meeting detected. Writes today.md. Delivers brief.
Also triggers on: "good morning", "brief me", "what's on today", "start my day"
/resume Session resume. Reads files, skips live signal pull, surfaces latest checkpoint. Compact "where you left off" output. Use when picking up a paused thread — not for true day-starts.
Also triggers on: "resume session", "picking up where we left off", "continue from last session"
/checkpoint Mid-session state write. One question, one write to state-of-play.md. Under 2 minutes. Use when context is running low or pausing a long working thread. No synthesis, no email pull, no brief.
Also triggers on: /save, "save where we are", "write progress", "checkpoint"
eod-brief skill
Open in your Opus session. Run at day-end and for standalone synthesis passes.
Opus
/eod Full end-of-day brief. Reads context, pulls sent mail + Slack + calendar, forms questions (MUST cap 5), routes answers to domain files, stages synthesis to synthesis-inbox.md silently, logs ✅ to done-log.md. On Fridays: extended weekly synthesis runs automatically.
Also triggers on: "end of day", "debrief me", "close out the day", "wrap up the day"
/synthesis Synthesis-only mode. Skips questions and task tracking. Reads context files (Step 1 only), then goes straight to the three-phase synthesis pass (Diverge / Attack / Filter) plus independent idea generation. Stages output to synthesis-inbox.md. Use any time, any day, after a significant meeting or decision.
Also triggers on: "strategy synthesis", "give me a synthesis", "synthesis pass"
new-project skill
Use when starting any new Floward workstream. Works in any session.
Any session
/project New workstream kickoff. Reads all existing context first (no cold start), pre-fills the setup interview with confidence ratings, creates the project overview file, updates this-week.md and _index.md, suggests next actions.
Also triggers on: "new project", "start a project", "spin up a project", "I'm starting X", "let's kick off X"
email-triage-floward skill
Dedicated inbox triage for the Floward context. Works in any session.
Any session
/inbox Full email triage. Converts inbox into a prioritized work queue organized around Floward's context — Creative AI deliverables, Omar/Asil stakeholder chain, vendor relationships, peak-season planning.
/inbox 48h Triage scoped to last 48 hours. Same output, narrower time window.
Variants: /inbox this week, /inbox quick (stripped action list only)
publish-vault skill
Publish or update any HTML file to a live Vercel URL. Works in any session.
Any session
/publish floward Publish to the Floward vault (floward-kappa.vercel.app). For work outputs shared with teammates — reports, dashboards, briefings.
/publish personal Publish to the personal vault (personal-two-azure.vercel.app). For personal projects.
Both return a permanent shareable URL. Committed to private GitHub for version history. To update an existing artifact: use the same command — the skill handles replacement.
Manual operations (not skills)
/reconcile Check synthesis framework drift. Compare version headers on eod-brief SKILL.md (v2.1) and synthesis-prompt.md (v2.1). If they diverge, identify differences and update the out-of-date one. Run whenever the synthesis framework is edited.
Not a skill — just describe it to Claude and it will check the two files manually.
Skill files — install by double-clicking
eod-brief.skill outputs/skills/eod-brief.skill
morning-brief.skill outputs/skills/morning-brief.skill
new-project.skill outputs/skills/new-project.skill

The -final.skill files in the same folder are from a prior version — install the ones listed above. Both eod-brief.skill and morning-brief.skill were repackaged on June 7 (v2.1). New versions include the synthesis-inbox pattern, pre-meeting synthesis trigger, done-log routing, and version headers.

Maintenance

What needs attention, and when.

The system is mostly self-maintaining — daily rituals keep files accurate. Two maintenance windows keep intelligence files from drifting.

Weekly · every Friday

Extended synthesis

Automatic — no action required. When you run /eod on a Friday, the extended weekly synthesis runs automatically at Step 8.

  • All three rotating ops run (negative-space, asymmetry, decay)
  • Full week of context reviewed, not just today's
  • Generates 2-4 new pipeline ideas, staged to inbox
  • Friday morning: inbox adjudication clears the week's observations
Monthly · first Monday

consolidate-memory

Run before anything else on the first Monday of each month. Three files, ~10 minutes total.

  • state-of-play.md — remove resolved flags, compress intel older than 4 weeks
  • ideas-pipeline.md — review trigger conditions, move Shipped/Archived to ideas-archive.md
  • ai-strategy-nucleus.md — promote Emerging Patterns with 3+ touchpoints to Strategic Principles, move stale entries (no new evidence in 3 months) to nucleus-archive.md
Why three files, not all of them?

These three drift fastest. The session state files (this-week.md, state-of-play.md) are rewritten by the daily rituals — they stay accurate automatically. Domain files like the photography handover and decisions log only need attention when a project closes. The three intelligence files (state-of-play, pipeline, nucleus) are the ones that accumulate without a natural stopping point — they need a deliberate review cycle to stay useful.