⇠ chan.dev / posts

2026

24

Jun 8-14

Games

Simple Apps

23

Jun 1-7

Personal System

  • Added Captioner, QR, and YouTube Audit as new personal-system tools.
  • Mapped middle-click to Wispr Flow with Karabiner.
  • Idea: build an MCP inspector that connects to a server and shows available tools and descriptions.

AI / Agents

  • Grok STT was not as strong as advertised in Captioner testing. Best results came from Whisper plus an Anthropic auditor; Sonnet models seemed to be the sweet spot, though xAI API credits are available via the xAI Console.
  • Poke — new AI tool to explore.
  • Is It Agent Ready? — combine with /goals; strong fit.
  • Clearly AI-built sites that seem to be connecting in AEO research for my tools: ScribeForge and Framewise.

WorkOS

Video Ideas

  • What is a basic AI concept — agent, harness, etc.
  • Flue Agents exploration.
  • Building on Cloudflare.

22

System

  • Added captioner.chan.dev to the cloud system inventory: a personal captioning workflow app built with Flue workflows.
  • Updated shutdown ritual practice to keep tracking notable changes to the personal system: tools, services, domains, workflows, automations, and conventions.

Agents / Workflows

Auth / APIs

Aesthetic / Culture

20

Shutdown

  • Ran update-deps; upgraded SQLite from 3.53.0 to 3.53.1 and refreshed global Node tools.
  • Ran sync-repos; all synced except .agents, which had local shutdown ritual changes.
  • Updated shutdown ritual to run maintenance, review local repos, and ask before committing or pushing.
  • Committed and pushed .agents: b28d067 Update shutdown ritual maintenance flow.

19

18

17

16

  • Open Tabs

15

Video ideas:

  • Build a skill skill
  • Unversal Skill directory
  • Distribution-shifting vs direction (on building skills)
  • 5 Coding Agents, one prompt
  • Building a Text Agent (Blue Bubbles)

NCC Message notes

Mess -> Message -> Messanger — Chris Brown

  1. Embrace interruption
  2. Obey instructions
  3. Don’t stop intercessions

Make the swap (Indiana Jones): Jesus for whatever my idol is.

14

Video ideas:

  • How to Say Goodbye. On dealing with death of loved ones.
  • How to write an AI Skill
  • DevTools MCP
  • Better than AI. (Gmail scripts)
  • Firecrawl.dev

13

Pi First Look

Chan + Nisi Livestream on Pi

  • Livestream with Nick Nisi exploring Pi — model switching, extensions, plugins, Case tool, vendor lock-in discussion
  • Full pipeline processed (cut → polish → zoom)

WorkOS 2026 Spring Launch Week

  • Wrote and recorded launch week recap script (five features in five days)
  • Recorded/edited Pipes MCP video (Day 4)
  • Features: Multiple Apps, AuthKit Analytics, Agent Experience, Pipes MCP, Widget Skills
  • Downloaded WorkOS Employee LinkedIn Banner
  • WorkOS Launch Week recap script

VPN Setup

  • Downloaded OpenVPN configs

Agentic Fever

Universal Skills / Agent Interop

AI Agents

Dev Life

Watch Stuff

  • Pseudo world time: add a worldtime bezel to a normal watch, match timezone to hour hand. Manual but fun.
  • Seiko SKX007 (eBay)

Lie Detector Science Project (with Ruby)

Carryover from Week 12

  • #thumbpulling
  • Moar chanisi
  • Migrate old projects off of Frame, to Lawn
  • Watch Akira (1988)

12

Browser Use

Watch Deals

11

Harness Engineering

One-sentence summaries

  • OpenAI frames harness engineering as building the environment around agents, with docs, tools, evals, and review loops that let teams ship reliably.
  • Anthropic argues long-running agents work best when harnesses favor clear state, guardrails, and measurable checkpoints over orchestration complexity.
  • This stream shows a practical harness pattern where an orchestrator delegates implement, verify, review, and close stages, then learns from retros.

Highlights

  • Fixing the harness is often more valuable than fixing one failed run.
  • Agents need machine-checkable evidence like tests, logs, and recordings, not just good prompts.
  • Persisted run state like task.json makes long workflows resumable and less fragile.

Hooks

  • If the agent cannot prove success, it has not earned merge.
  • Your next leverage point may be better rails, not better prompts.
  • Every failure can become training data for a stronger harness.

Stream with Nick Nisi

“It’s going on the fridge. But it’s not going on the wall. (Somebody else’s father won’t accept it)” — Cooridor Crew

Stream assumptions vs. the articles

CORRECT
- "The harness is what you build on top of generic tools for your
  specific project." Both articles confirm this. OpenAI: "the full
  environment of scaffolding, constraints, and feedback loops."
  Anthropic calls the Agent SDK a "general-purpose agent harness,"
  so both layers exist — Chan and Nick were each right about a
  different one.
- "I will steer, I will never execute." Dead on. OpenAI says
  engineers "shift from writing code to designing environments,
  specifying intent, and providing structured feedback."
- "Fix the harness, not the individual problem." Almost verbatim
  from OpenAI: "When the agent struggles, we treat it as a signal:
  identify what is missing — tools, guardrails, documentation —
  and feed it back into the repository."
- The retrospective/learning loop maps to OpenAI's third pillar:
  entropy management — periodic agents scanning for drift,
  inconsistencies, and pattern deviations.
- State persistence via task.json matches Anthropic's
  claude-progress.txt and feature_list.json — external memory so
  agents can resume across sessions.
- Cryptographic hash hooks to prevent agent cheating maps to
  OpenAI's architectural constraints enforced mechanically, not
  through suggestion.
- "The harness is the actual product." OpenAI says exactly this.


INCOMPLETE
- Architectural constraints as a full pillar. OpenAI enforces
  dependency flow rules (Types → Config → Repo → Service →
  Runtime → UI) with custom linters, structural tests, and CI
  blocks. CASE focuses on orchestration and retros but doesn't
  describe this kind of structural enforcement.
- Documentation as infrastructure. Both articles treat repo docs
  (AGENTS.md, structured docs/) as foundational infrastructure,
  not afterthought. The stream touches philosophy.md and skill.md
  but doesn't frame documentation investment as a core pillar.
- Entropy management at scale. The retro agent is a solo-dev
  version; OpenAI describes background agents continuously
  scanning for doc inconsistencies, architectural drift, and
  dependency issues — proactive, not just reactive per-run.


MISGUIDED
- "Is harness engineering basically building your own agents?"
  Both articles frame it as environment design — repo structure,
  CI, docs, constraints, feedback loops. The multi-agent pipeline
  (orchestrator → implementer → verifier → reviewer → closer →
  retrospective) is more "agent engineering" in OpenAI's taxonomy.
  The harness is the reins, saddle, fence, and road — not the
  rider giving commands.
- The level of orchestration complexity. Anthropic uses just two
  agents (initializer + coding agent) with different prompts.
  OpenAI doesn't emphasize multi-agent orchestration at all — they
  focus on making the environment so well-constrained that a
  single agent can succeed. If you need elaborate orchestration,
  it may signal the harness needs more investment.
- The terminology confusion. OpenAI drew a clean taxonomy: prompt
  engineering (what to ask), context engineering (what to show),
  harness engineering (how the environment is designed), agent
  engineering (agent internals). The terms have converged more
  than the stream suggested.


TL;DR
Core instincts were right: the harness is the product, steer don't
execute, fix the system not the symptom, build learning loops.
Incomplete on architectural constraints and docs-as-infrastructure.
Slightly misguided in conflating agent orchestration with the
harness itself — the articles define the harness as the environment,
not the agent pipeline running inside it.

For Next Week

  • #thumbpulling
  • Moar chanisi
  • Migrate old projects off of Frame, to Lawn
  • Watch Akira (1988)

10

Databrix roadshow intro

  1. Let’s talk about passwords
  • What was your first? Blurt it out nobody cares
  • (play up options)
  1. Now acct protection requires a meaningful ammount of infrastructure. MFA, Emails, Passwords, JWT-checking.

Reading

For Next Week

  • What is Harness Engineering?

9

8

Writing

Alternative window managers

  • OmniWM — soft-fork of Aerospace with Hyprland dwindle layout and Niri tabs
  • dwmac — hard fork of Aerospace

Molly Graham — frameworks for rapid career growth

https://youtu.be/twzLDx9iers

  • J-Curve vs. Stairs — jump into roles you’re unqualified for. You fall for 6-9 months, then climb far beyond where steady promotions would take you.
  • Fear as signal — financial fear is concrete (calculate your burn rate). Capability fear (“am I good enough?”) is a green light for growth. Run toward it.
  • Give away your Legos — as a company scales, hand off what you’ve mastered to new hires. Holding on stalls your career.
  • Two-week rule — most emotional reactions to change settle in two weeks. Only act on feelings that persist beyond that.
  • Bob the Monster — externalize irrational reactions by personifying them. Acknowledge Bob, don’t follow his advice.
  • Waterline model — diagnose team problems in layers: structural, dynamics, interpersonal, intrapersonal. 80% of problems are structural or dynamics. Fix those first.
  • Three rules for goals — no more than three goals; one must win when they conflict; a new intern should understand them on day one. Each goal gets one owner (not the CEO).
  • Strategy means painful choices — if setting goals doesn’t hurt, you’re not prioritizing hard enough.
  • Founder effect — ~80% of culture reflects the founder’s personality. Reinforce it rather than inventing new values. Culture lives in hiring/firing decisions, not posters.
  • Invest in high performers — highest ROI comes from developing your best people, not rescuing struggling ones.
  • Hiring is 50/50; firing is a skill — even great managers are right about half the time. Fire quickly and directly.
  • Never promise what you can’t control — don’t guarantee titles or promotions during hiring. They become letter bombs when reality changes.
  • ~50% annual growth is optimal — beyond 100% headcount growth causes severe org pain. Under-resourcing forces focus.
  • Escalation isn’t failure — when equals disagree, escalating to a decision-maker is productive.
  • LinkedIn Crush test — if you’re more excited about announcing the title than doing the work, that’s misalignment.
  • Be a professional learner — learning speed > existing expertise. Asking obvious questions brings clarity everyone else is afraid to seek.

7

6

WorkOS NYC office opening and time with Zack.

5

Kids made dinner!

4

  • Restarted Wellbutrin/Bupropion
  • Improved personal accounting by starting *Recurring expenses acct”
  • Streamed on CC + Remotion
  • Medication helping me lock into full task
  • Enjoyed new ghostbusters after long week of screens
  • I didn’t buy or obsees over watches
  • Served at church with Rock and discussed favorite verses.
    • Shared sadness over not being invited.
  • #video on “asking the right questions of AI”/interview

Ideas

  • No retained sweets in the house
  • #video on materialization function

Learnings from Cloned Consultant (Claude Code)

  • “Interview me about this…”
  • Podcast transcripts are great for research
  • Summarize to markdown/html
  • Be specific about ouput (“5 star chef”, “Senior dev”)

Things I need from Raycast

  • Clipboard history, bound
  • Snippets, bound
    • email
    • expense reports

3

“Great chokes out God” (in our lives) — Chris Brown “The worries of this world ad decietfulness of wealth” - Mark 4:19

“I pray to god that they (congegrants) understand that they teach 90% of the message.” - William Barkley

WorkOS vision for year

  • “WorkOS TV”, constant conference
  • “Build fast” mantra/theme
  • Merch drops
  • Don’t do evergreen. Focus on what’s fresh.

Hourly materialization

  • Materialize a descriptor
  • collect descriptors based on date

Improvements

  • Add Recurring expenses

2

Video ideas

Build an AI Saas on WorkOS

  • WorkOS CLI
  • An Org primer and auto assignment
  • Vault for Keys (Anthropic), shared across org
  • Stripe and Entitlements
  • Widgets for user management (and others)
  • RBAC (FGA?)
  • Feature flags
  • Pipes (Slack integration?)
  • Connect

1

2026 Pig Zodiac

Instant Pot Recipes