AmalgamRx · AI Enablement Session
The question is whether you’re driving it. How agents actually work — context, skills, loops, goals — mapped to what you each do here.
Chatting with AI: you ask, it answers, once. You still gather the context, take the next step, and check the work — every time, from zero.
Directing an agent: you point it at an outcome. It gathers what it needs, does the work, checks itself, and only comes back when it’s actually done — or stuck.
Same model. Different relationship to the work.
The anatomy of an agent: the model (the part that thinks), the goal (the finish line you set), context (its working memory right now — finite, like a desk), memory (what carries over between conversations), tools and connectors (its hands: calendar, Gmail, SharePoint, search), channels (where it meets you: chat, Cowork, browser, a scheduled run), and skills (recipes on the shelf, written once and reached for when a task calls for them).
Within a thread: every message stays in view, so corrections stick and the work compounds. Across threads: Projects and memory carry your role, tone, and history into every new thread. Over time: the prompt you save, the skill you write, the doc you refresh Friday. It learns the way a new hire does: from what you write down, correct, and hand back.
Token: a chunk of text, roughly three quarters of a word — your quota is a pile of them. Context window: how much the agent can hold in view at once; big, and still finite. Connector: a one-time link to a tool you already use — it turns reading into doing. Skill: a saved recipe the agent reaches for. Loop: how it runs — gather, act, verify, repeat until the work holds up.
Gather context. Looks at what’s actually there first — the notes, the last version, the real data.
Take action. Doesn’t describe the fix. Makes it — drafts the email, pulls the report, edits the file.
Verify results. Checks its own work before it lands in front of you.
A simple question might only need step one. Messy work cycles through all three until it holds up.
Dump everything in and it drowns. The quality gap isn’t 10% — it’s 5 to 10 times better with the right context loaded.
The deliberate practice of building, curating, and reusing your context — instead of rebuilding it from zero in every conversation. AI has no memory; every new thread starts blank. The fix is a standing asset you build once.
Five types to farm: Role Context, Company Context, Task Templates, Output Preferences, Relationship Context. Build plan: personal context doc (30 min), company block (1 hr), task templates (2 hrs), relationship profiles (1 hr), 15 min/week maintenance. Total setup 4–5 hours; payback starts day one.
Each one sits on the shelf with just a name and a label; the full recipe only comes down when a task calls for it. Examples: Contract Redline Checklist, QBR Builder, Config Spec Template, Prospect Outreach Brief, Status Roll-up Format, Module Dependency Check.
Plan first: for anything broad or risky it reads, analyzes, proposes a plan — no actions until you say go. Define “done”: one measurable finish line it checks itself against. You’re managing an outcome, not typing instructions for every step.
One Claude, four doors. Claude Chat + Projects for quick drafts and single documents. Cowork for multi-step knowledge work with a finished result to review. Claude Code for the same loop pointed at a repo. Claude in Chrome + Excel for the tools you already use. Rule of thumb: one answer, chat. One deliverable, Cowork. One codebase, Code. One website or spreadsheet, Chrome and Excel.
Before: agenda pulled from the last thread in two minutes. Right after: notes become follow-ups before the call closes. Standing: one pull each morning. It can already see your calendar, inbox, and docs — if you connect them.
The AmalgamRx toolbelt, connected once: Mail + Calendar (Gmail, Google Calendar), Meetings + Transcripts (Google Meet, Microsoft Teams, Fireflies), Docs + Knowledge (SharePoint, Confluence), Research + Notebooks (NotebookLM and the open web through Claude in Chrome). Setup is one time per tool, in Settings.
A scheduled task is a standing goal on a timer. Every morning: inbox, calendar, and overnight tickets distilled into one brief. Every Friday: status pulled across workstreams, reds flagged, roll-up drafted. Always watching: a watchlist that pings you only when something moved. Start with one — the recurring task you dread most is usually the best candidate.
Stay in the same thread. Give it the real thing, not your summary. Say the format up front. Don’t stop at the first draft. Save what you’ll reuse. Push back directly when it’s wrong. Let it ask you questions first on anything ambiguous or high-stakes.
One good brief beats ten retries. Load the doc, skip the retyping. Start fresh when a thread gets heavy. Match the effort to the stakes. Batch the small stuff. Reuse what worked. Ask for a plan before the big run.
Starter prompts for Program Managers, Product Managers, Client Success, and Legal/Regulatory — copy, paste, adjust in the interactive version.
Role-by-role before/after and a first move for Program Managers, Product Managers, Client Success Managers, Project Managers, Legal & Regulatory, Engineering Leads, and Sales & BizDev — explore them interactively with JavaScript enabled.
Signs a draft needs a harder look: overly confident tone on uncertain topics; specific-sounding numbers with no traceable source; instant unanimous agreement with your framing. At AmalgamRx, clinical judgment, regulatory interpretation, and client commitments stay human — always.
This week: build one context doc; give one recurring task a real finish line; ask it to plan before it acts, once; point it at your calendar or inbox. Context you build once keeps paying off every week after.
Framing adapted from Anthropic’s engineering docs and the Claude Code documentation. Role scenarios build on AmalgamRx’s internal 201 Playbook.
Two quick things — the page tailors its examples to your world. Both optional.