Claude Code for Knowledge Work

And other learnings about building agents
And practical advice for building a personal agent

The Type of Work I Do

Understand
  • Customer calls
  • Competitive research
  • Data deep dives
  • Closed-lost deal reviews
  • Domain learning
  • Reading industry trends
→
Choose the Right Bets
  • Spec & memo writing
  • Prototypes to stress-test ideas
  • Prioritization & trade-offs
  • Data analysis
  • Stakeholder alignment
→
Bring to Market
  • Specs for engineering
  • Unblocking the team
  • GTM & launch planning
  • Writing code
  • Validating outcomes

Goals

No Repeated Context
When I tell the agent something, I don't have to repeat it session to session.
Optimal Context per Task
Each session loads exactly what it needs through progressive disclosure. Research shows LLM performance degrades as context grows.
Intern → Co‑worker
Intern: extends my work — does my tasks faster.

Co-worker: does work I didn't even imagine that pushes me and my team closer to our goals.

Basic Principle

My Space — only I write
~/Obsidian/
├── CLAUDE.md ← the harness
├── projects/
│   ├── construction/
│   ├── dev-api/
│   └── personal/
├── goals/
│   └── current/
├── resources/
│   └── my-ideas/
├── inbox.md
└── journal/
~/.claude/skills/
├── daily-brief/
├── deal-review/
├── process-meeting/
└── … 40+ skills
Agent Space — only agent writes
agent/
├── registry.md ← map of the vault
├── memories/
│   ├── self-assessment.md
│   ├── knowledge-gaps.md
│   └── skills/
├── events/
│   └── 2026-03-04.md
├── projects/ ← self-directed
└── scratch/ ← ephemeral
Tool Layer — MCPs · CLIs · Slack · Obsidian · Granola · Git

Guardrails live in the harness. The agent manages its own space freely within them.

The Harness: Two Layers

Goal: keep total harness under ~4k tokens

Layer 1 — Claude Code itself
Claude Code system prompt
Built-in instructions that ship with the tool
Ramp enterprise CLAUDE.md
Org-wide guardrails pushed to all managed Macs
Layer 2 — mine
~/.claude/CLAUDE.md
Personal global rules — applies everywhere
~/Obsidian/CLAUDE.md
Vault harness — orientation, memory protocol, guardrails
Skills loaded per task
Prompt templates injected on demand

All of these stack into one system prompt. Keep it lean — every token costs attention.

The Memory Layer

Just flat markdown files. No vector store, no database. Portable and inspectable.

Tiered retrieval
L1 — Hot Registry
~20 most-used files. Regenerated nightly. Agent reads this first for fast routing.
L2 — Full Registry
Complete routing table of the vault. Fallback when L1 doesn't cover it.
L3 — Memory Content
The actual files — self-assessment, knowledge gaps, skill calibrations. Loaded only when needed.
How it stays healthy
Provenance tags
observed Agent noticed something
corrected I confirmed or fixed it
Corrected entries never get pruned.
/sleep — nightly consolidation
Runs at midnight. Sets line budgets per memory file, then lets the agent freely organize within those constraints. Goal: optimize for progressive disclosure. Prunes stale, consolidates duplicates, enforces provenance.
Progressive disclosure
The agent decides its own memory structure — what to keep, how to organize — within the constraints. The harness just says: load the least context needed to do the task well.

The Agent Workspace

Event logging
agent/events/2026-03-04.md
sleep — 00:00
Pruned 20 entries, consolidated duplicates
daily-brief — 06:00
3 priorities, 7 Slack drafts, 3 meeting preps
independent-work — 06:00
Shipped construction PM brief
autopilot — 07:00, 10:00
Slack scans, inbox updates, no new urgents

Every session logs topic, actions, outcomes, and files accessed. Builds a queryable history.

/independent-work — self-directed projects
How it works
  1. Scan goals, inbox, Slack, journals for opportunities
  2. Triage candidates — max 3 active at a time
  3. Execute — research, write, build, analyze
  4. Ship artifact, surface to Teddy via DM
Constraints
  • · One hour budget per session
  • · Must produce a concrete artifact
  • · Must trace back to a goal
  • · Never send messages — draft only

The agent decides what to work on. I review what it shipped async.

Skills: Repeatable Workflows as Code

How to build a skill
01
Tell it what, not how
Define the goal and constraints. Let the model figure out the steps.
02
Build in progress tracking
Let the agent measure its own quality over time so you can see improvement.
03
Monitor closely first
Run the skill 4-5 times manually in real workflows. Give it feedback, let it iterate. Once you trust it, let it run autonomously.
Example: /daily-brief
What it does
  1. Load goals + mental model for prioritization
  2. Run through all inputs — Slack, calendar, inbox, news, weather, sports, lunch order, restaurant recs
  3. Build recommended brief with top priorities and focus areas
  4. Re-filter through original goals lens, then write the brief
  5. Ask for feedback, track accuracy over time
Accuracy over time
Day 1 36%
Day 10 100%

Ran it manually ~10 times, gave feedback every run. Now it auto-runs at 6am.

The Slack Bot

Headless Claude Code processes with an identity layer on top

Architecture
Claude Code harness
  ├── CLAUDE.md ← vault rules
  ├── Skills ← 40+ workflows
  ├── Memories ← learned context
  └── MCPs ← Slack, Clockwise, Linear…

+ Identity (a markdown file — voice, personality)
+ Bolt app (routing, scheduling, process mgmt)
What it does
  • Responds to DMs and @mentions
  • DMs me results of recurring tasks
  • Every hour, scans Slack for new signals
  • Surfaces completed follow-ups
  • Quick summary: what needs attention
Guardrails
DM: full vault, all skills + memories
Channels: project context only, restricted skills + memories, no personal data

~800 lines of JS glue. All the intelligence lives in the Claude Code harness.

Automated Processes

Independent Work

Runs daily at 9am.

Agent proposes its own tasks based on goals and open work. I review what it did async.

Proactive Scan

Monitors signals without being asked — Slack channels, deal signals, customer asks. Processes Granola transcripts. DMs me the results.

Sleep

Midnight nightly.

Consolidate memories. Prune stale entries. Validate org health.

Coming soon — still manually refining
  • Closed-lost deal reviews
  • Monitoring Gong for signals that match roadmap features
  • Updating Linear projects and Notion
  • Keeping internal documentation up to date

Limitations

01
Runs Locally
Everything runs on my laptop. If it's closed, the agent is dead. No cloud deployment yet.
02
Slack Permissions
The bot needs explicit approval for every scope — DMs, reading channels, posting messages. Each permission is a separate ask.
03
Cost
All those automated processes use tokens. It adds up.

Best Practices

Handcraft Your Harness
Claude is not good at writing its own harness — it doesn't understand its own limitations well enough. You need to handcraft this yourself. It's the highest-leverage thing you can do.
Tell It What, Not How
Set the goal, set the framing, set constraints, give it the tools it needs. Don't give step-by-step instructions. Always have a final calibration step where it reruns output through the original goals and constraints.
Refine Hands-On First
Your first version won't be good. Run it, watch it, refine it — with real workflows, not test cases. This is the hardest part of building for users too: you need hands-on sessions watching them run skills and refining together.
Measure Progress
During calibration, you need a way to evaluate whether you're hitting the outputs you want. I haven't fully nailed an evals framework, but the agent needs benchmarks to assess itself — like the daily brief accuracy score.

Starter Kit

I'm sharing a repo with everything you need to start.

CLAUDE.md Template
Annotated harness — your role, projects, rules, file layout, memory protocol. ~30 min to customize.
Agent Space Scaffold
Pre-built directory structure — registry, memories, events, scratch. Ready to go.
Four starter skills
Daily Brief
Morning priorities with calibration scoring
Proactive Scan
Hourly signal monitoring across channels
Sleep
Nightly memory consolidation and pruning
Independent Work
Self-directed goal-aligned projects
Start with Claude Code. Write your harness. Refine a handful of skills.
Protect your own space — the agent can draft for you, but your harness is your source of truth.

github.com/riker-t/claude-for-knowledge-work
1 / 13 ← → to navigate