The Sidekick operating system, in detail.
Inspect the architecture, workflows, quality controls, operating limits, and delivery process behind a Sidekick workspace.
The whole thing, in detail.
This specification documents what is inside, how each layer works, what it solves, and where its current limits sit. It is written for people who already use AI seriously and want to inspect the operating mechanism.
Identity, authority, sources, safety, and what must always hold.
Current truth, durable task resumes, decisions, relationships, and work history.
Briefs, plans, specialist routing, safe work graphs, and repeatable workflows.
Connected tools, monitors, briefs, dashboards, and recurring operating cadences.
Direct proof, quality gates, staged outputs, and human control of consequential action.
Compact, source-linked logs help decide what deserves to become a rule, workflow, connection, or schedule.
01What it solvesAI workspaces decay by default
Adopt Claude or ChatGPT seriously and the first few weeks feel like progress. Then the slide starts. The same failures show up across very different people, in the same order, and they compound quietly enough to look like the model getting worse. It does not get dumber. It gets buried.
None of these is dramatic on day one. By week six the workspace is a junk drawer, you trust it less than at the start, and the honest read is that the tool never paid back its own learning curve. That is what turns “it knows me” into “it has gotten worse.”
Want the plain-English version of this? Read Why Your AI Gets Worse the More You Use It.
02What you receiveYou get a workspace that gets better the more you use it
The deliverable is an operating environment that gets more useful with use: context compounds, recurring work becomes a named workflow, and verified improvements propagate forward into new builds. The same operating pattern runs in our ChatGPT workspace and in Claude + Cowork client builds. The exact files, connectors, and interface follow the platform and the work.
Each build starts from the current Sidekick foundation and is populated with your context, boundaries, and highest-value workflows. The core is file-first, inspectable, and portable. Your business context stays in a workspace you can read, own, and move; live connectors are added where they make the work more useful.
|-- CLAUDE.md the instruction layer; loads at every session start
|-- Start_Here.md one-page entry point, names your first-win task
|-- Your_Workflows.md your 2-3 highest-priority workflows, copy-paste prompts
|-- INDEX.md master map of every file in the workspace
|-- Starter_Prompts.txt the full skill routing table: trigger phrases
|-- Skills/ 8 day-one skills + Skills/Operator/ (30-skill library)
|-- Guides/ branded PDF guide set, dashboard template, Your First Month
|-- Reference/ Business + Comms profiles, Current_State, Action_Items, Decisions
|-- People/ one context file per key relationship
|-- Meetings/ structured call notes, one file per meeting
|-- Work/ durable task resumes, evidence, reviews, and exact next actions
`-- SOPs/ · Active/ · Deliverables/ your processes, work in progress, finished output
A few root files carry the day-one load: Start_Here.md (how to start a session and your named first win, no placeholders), Your_Workflows.md (your priority workflows with copy-paste prompts), and Starter_Prompts.txt (the routing table that makes triggering reliable). A reader who opens the folder cold finds an entry point, a map, and no internal build artifacts.
03Memory that compoundsThe index is what makes memory reliable
Most people run their AI memory like a junk drawer: everything goes in, nothing comes out, and eventually you are afraid to open it. The fix is a handful of boring disciplines, none of which need a better model. CLAUDE.md is the one file the platform reliably loads at session start, so it carries everything that must always hold: identity, the routing table, the memory rules, and the safety rules.
Give the workspace an index. A simple map of every file and what it holds lets the model look up the one relevant, current thing and read just that. With no index, it rereads the entire filing cabinet to answer one question, fills its working memory (the context window) with noise, and forgets the answer. People read that as the model getting forgetful. What is really happening is it has more memory than it can sort, so it holds all of it at once and weights none of it. The index is also where relevance and recency get decided. For most people hitting the “it cannot remember anything” wall, this is the highest-impact fix almost nobody has in place.
Rewrite the state, do not append to it. The current-state file is rewritten at the end of each session so it always reflects what is true now, instead of stacking session recaps until it contradicts itself. This single discipline kills most of the instruction-pileup decay on its own.
Route memory by type. What is true now (priorities, business, people) lives in one place. What happened when (decisions, meeting notes) lives in another. How things get done (standing rules, workflows) lives in a third. Every “remember this” is sorted to the right type before it is written, so the assistant can actually find it later.
Keep long work resumable. A substantial job gets a compact control surface with its objective, completed milestones, evidence, failed paths, approval boundaries, and exact next action. If a conversation is interrupted, the system resumes from the record instead of reconstructing the job from chat history.
Every drafting and analysis skill reads this layer before producing anything, which is why output quality scales with workspace age here and degrades with it elsewhere. The architecture is the moat; the prompts are replaceable. A weekly scheduled lint pass catches stale content, contradictions, and dead references before they mislead a session.
04Skills + orchestrationSkills do the work. Orchestration makes them one system.
A skill is a saved specification that wraps a recurring piece of judgment: what to read, what to produce, what never to do. A core set is active from day one, and a deeper library opens up as your usage matures, so most builds end up running more than 40 skills in total. Every skill carries explicit trigger phrases, because manual triggering is the reliable path on this platform (more on that under limits).
For substantial work, orchestration sits above the individual skills. It locks the outcome, chooses the specialists, preserves state, and checks each handoff before the next phase begins. Work with real branches can use a small work graph so independent parts run safely and rejoin at one verified result. Simple work stays simple.
Research must be sound enough to plan from. The plan must be sound enough to build from. Novel work proves the smallest complete path before broad production. When a premise breaks, the job returns to the stage that owns it instead of spending a final polish loop on the wrong thing.
| Day-one skill | What it does |
|---|---|
| session | Opens and closes the session: reads context, then updates current state, logs decisions, and integrates corrections. The ritual that makes everything else persistent. |
| daily-brief | A sub-3-minute morning orientation from priorities, action items, decisions, meetings, and live calendar, email, and CRM where connected. |
| post-call | Raw notes or a transcript in; structured meeting notes, signals, a follow-up draft, and a next action out, in one pass. |
| quick-capture | “Remember this” routed to the right file and confirmed in one line. No questions, no analysis. |
| action-items | The live task ledger: add, review, filter by person, prioritize, hygiene sweep. The owner column drives the dashboard split. |
| meeting-prep | A one-page brief for an upcoming meeting: agenda, objectives, relationship context, recent signals. |
| executive-comms | Drafts a message to a named recipient in your calibrated voice and their register. Never sends without approval. |
| dashboard-render | Renders the live daily dashboard from your action items, decisions, and current state. |
| Operator library · introduced as usage matures | |
| Documents | content, docx / pptx / xlsx, outreach, prompt-sharpener: long-form writing, branded documents, and drafts in your voice. |
| Planning | plan / build, decision-brief, strategic-analysis, board-prep: multi-step work spec-locked, scored, and steelman-first. |
| Orchestration | one-shot, brief locking, work graphs, durable resumes, verification, and approval-ready handoffs for substantial jobs. |
| Voice | my-voice / voice-extractor, stakeholder-comms, people-intelligence: voice pulled from real samples, multi-audience handling. |
| inbox-organizer, email-drafter: triage plus reply drafting staged in Gmail. Never sends. | |
| Meetings | meeting-intelligence, week-ahead / weekly-review: transcript processing and the weekly cadence. |
| System health | feedback-loop, sop-capture, tune-this, quality-loop, workspace-eval, context-lint, update-diff, monthly-review, schedule. |
The baseline: we turn your two or three highest-impact workflows into custom skills, with trigger phrases rich enough to catch the natural ways you would ask. From there it grows with you. Once those first workflows are running, the coaching tends to surface the next ones, and clients who take the coaching often finish with half a dozen or more custom skills, covering ground they did not know could be automated when they started, until the bulk of their operation runs on tooling built specifically for them. One delivered example: a bookkeeping practice owner runs her core bookkeeping workflows as custom skills.
05Your daily surfaceScheduled work keeps the system useful between conversations
The dashboard is your persistent operating view, rendered as a live artifact from your working files, not a separate database that drifts. Tiles come from your action items, decisions, and current state, and the owner column splits work into “the system can handle” and “only you.” It is plain, inspectable HTML that runs inside your workspace, with a setup guide for adding tiles.
Scheduled tasks handle the work that should not depend on you remembering to ask: morning orientation, follow-up detection, pipeline and delivery scans, context health, and bounded system reviews. Each important run leaves a compact receipt, including when nothing useful was found or when access failed, so silence does not look like success.
Scheduled work depends on the permissions, connectors, and runtime available to it. Long runs can fail, and a schedule is not the same thing as an event-triggered workflow. The system records missing or failed runs and keeps consequential action human-gated. It is an operating cadence, not an unbounded autonomous fleet.
06Built safe, proven by measurementA quality floor that has itself been tested
You will paste emails, transcripts, and documents from strangers into the workspace. The safety layer assumes that. External content is treated as data, never instructions: it cannot invoke a skill, change a rule, or authorize an action, and instructions found inside content are surfaced, never silently obeyed. Every outbound skill drafts only; sending is always a human click, which caps the blast radius of any manipulation at a staged draft.
No workspace ships on a feeling. Every build passes eight mechanical floor gates, each a literal pass or fail with no evaluator discretion: cross-client contamination, stale pricing, placeholders shipped as real, wrong identity, leaked internal content, dead references, core-skill integrity, and structural completeness. A single failure is a red verdict that blocks delivery. Then ten graded dimensions score the build against a written bar, ending with an experience score taken after a four-scenario simulation of your first week.
Substantial jobs use the same discipline while they run: challenge the handoff between research, plan, and build; test the first meaningful end-to-end path; verify the exact destination; and keep inaccessible human, legal, safety, or live-system evidence as an open gate instead of averaging it away.
A quality check you have never tested is just a vibe. So we tested ours. In June 2026 we ran a seeded-defect audit against our own framework: ten known defect classes planted into workspace copies, then blind evaluators who did not know what was planted ran the gate.
07How it improves over timeThe logs are evidence. The improvement is earned.
The evidence loop captures material corrections, verified misses, recurring friction, useful scheduled findings, and demonstrated successes. Ordinary completion and unsupported self-critique do not qualify as learning signals.
Signals are clustered by their first shared cause. A direct correction to a durable boundary or one high-risk failure can justify an immediate fix; ordinary signals need repetition. Every promoted change starts with a hypothesis, success bar, direct test, rollback path, and a later outcome check. Weak evidence can correctly produce no change.
Capture a real correction, verified miss, recurring friction, or demonstrated success.
Write one compact, source-linked signal instead of preserving a transcript.
Group evidence by the first shared cause, not by every downstream symptom.
Choose the smallest durable change the evidence actually supports.
Test the changed behavior against a success bar and keep a rollback path.
Check the next comparable uses to learn whether the change held, regressed, or should be removed.
A verified lesson is wired into the smallest owning surface: current context, a stable rule, a skill, a connector, or a scheduled task. New builds start from the current foundation. Delivered client workspaces are never silently retro-edited; updates remain visible and adoptable.
08Real gaps and limitsCurrent constraints and operating limits
These are the operating constraints we design around as of July 2026. The architecture choices only make sense against them, which is why they are in this document rather than hidden under it. Exact connector and platform behavior is rechecked before a workflow depends on it.
What the quality gate does not certify:
- Your activation. The gate certifies the workspace as delivered; whether you embed it in daily operations is coached, not guaranteed.
- Long-term drift. Priorities shift. The lint and review cadences catch drift, but only if you run them or keep the schedules on.
- Voice under pressure. A voice calibrated from polished writing may differ in a tense thread; the review mode converges it over time.
- Connector uptime. The gate cannot test the real-world stability of your Gmail, Drive, or CRM connections.
The system is wrong for buyers who want fully hands-off AI, who do not work at a computer daily, or who need deep enterprise integration. If that is you, we will say so in the first conversation rather than sell past the line.
09How delivery worksSigned engagement to independent operator
Plenty of people will configure an AI workspace for you. The difference here is the system around it: context engineered against decay, orchestration that returns errors to the stage that owns them, scheduled work with durable receipts, a quality gate that has itself been measured, an evidence loop that tests improvements before keeping them, and a written record of the limits. The result is an operating environment with a maintenance discipline, documented here in full.
Build a system around the work you repeat.
Same engine, your business. Different skills, your tools, your voice, shaped to how you actually work. Tell us what your week looks like and we will show you what it could run on.