Writing: Field Notes, AI Skills & Case Studies

01 · Writing

Everything we write in one place. Honest notes on building AI that earns its keep, plus open-source Claude Code skills and workshops, each ready to download, install, and go.

Field note
23 September 2026
CLAUDE.md Never Skips Permission Prompts
Writing always proceed in CLAUDE.md does not stop permission prompts. Claude Code enforces those rules outside the model. Use /permissions or an allow rule.
Read more
Field note
22 September 2026
The Formatter Ping-Pong
Anthropic's PostToolUse prettier hook rewrites the file Claude just edited, so the next Edit pays a re-read. Format once per turn with a Stop hook instead.
Read more
Field note
21 September 2026
CLAUDE.local.md Blocks AGENTS.md
Claude Code reads AGENTS.md only when no CLAUDE.md or CLAUDE.local.md is present. Keep both and AGENTS.md is ignored — set both-files mode, or import it.
Read more
Field note
18 September 2026
Bypass Permissions in Project Settings Does Nothing
Putting bypassPermissions in .claude/settings.json looks correct and fails silently. From v2.1.257 only user or managed settings, or --permission-mode, apply.
Read more
Field note
17 September 2026
Safe Mode Isolates Your Setup
Claude Code gets weird after a hook, skill, or MCP change. Official docs: `claude --safe-mode` turns customizations off so you can see if your setup is the break.
Read more
Field note
16 September 2026
CLAUDE.local.md Is Personal
People put sandbox URLs and personal prefs into committed CLAUDE.md. Official docs: local instructions live in CLAUDE.local.md — add it to .gitignore.
Read more
Field note
14 September 2026
Nested CLAUDE.md Never Loads at Start
A CLAUDE.md in a subdirectory looks finished and never loads at session start. Official docs: nested files load only when Claude Reads a file in that directory.
Read more
Field note
10 September 2026
Standalone Hooks File Never Loads
A standalone hooks.json looks finished and never loads. Official docs: hooks belong under the hooks key in settings.json; only plugins use hooks/hooks.json.
Read more
Field note
3 September 2026
How to Build Your First AI Agent: One Agent, One Job, One Owner
Learn how to build your first AI agent around one job, one accountable owner, clear guardrails and a measurable test before adding complexity.
Read more
Production AI
15 September 2026
Orchestration, Not Autonomy
A recurring theme among practitioners: the systems that survive production are the best-orchestrated ones, not the most autonomous. Task boundaries, explicit handoffs and deterministic checks at the joins.
Read more
Operating note
11 September 2026
Skills Beat Prompts
Reusable project-specific instructions are displacing ad-hoc prompting as the higher-value practice. What a skill actually looks like, when to promote work into one, and where the boundary sits.
Read more
Field note
9 September 2026
Lowercase Matcher Never Fires
Hook matcher set to lowercase bash looks finished and never fires. Official docs: matching is case-sensitive; tool names are capitalised: Bash, Edit, Write.
Read more
Delivery standard
8 September 2026
The Verification Bottleneck
Generating code stopped being the hard part. 85% of teams say the bottleneck moved to review. What actually shifts it: machine-produced evidence, deterministic gates and smaller units of change.
Read more
Field note
7 September 2026
`.mcp.json` Under `.claude/` Never Loads
Putting project MCP under .claude/, a VS Code servers key, or settings.json mcpServers never loads. The file belongs at root as .mcp.json under mcpServers.
Read more
Field note
4 September 2026
Migrating off budget_tokens
Setting budget_tokens returns a 400 error on Claude 4.7 and later. A practical migration to effort and max_tokens, plus the overthinking problem it was really solving.
Read more
Field note
3 September 2026
Permissions in ~/.claude.json Are Ignored
Putting permissions, hooks or env in ~/.claude.json looks correct and fails silently. Official docs: those keys belong in ~/.claude/settings.json, not there.
Read more
Field note
2 September 2026
Claude Code Ignores AGENTS.md
Claude Code reads CLAUDE.md, not AGENTS.md. Operators drop the cross-agent file in a repo and it is silently ignored. Import it, or symlink it — not /import.
Read more
Field note
1 September 2026
The CLAUDE.md Diet
Anthropic recommends keeping CLAUDE.md under 200 lines, because every line loads into every session. A decision tree for what stays, what becomes a skill and what becomes a hook.
Read more
Field note
31 August 2026
Auto in Project Settings Does Nothing
Putting defaultMode auto in .claude/settings.json looks correct and fails silently. Anthropic only honours auto from ~/.claude/settings.json or the flag.
Read more
Field note
29 August 2026
Cut Your Claude Code Bill
Cache hygiene, effort routing, model routing and keeping bulk reading out of the main window. Four documented levers that lower Claude Code spend without changing how you work.
Read more
Field note
29 August 2026
Grok Bot Templates Are a Recipe
A Grok Bot template is a recipe, not a clone. What ships, what stays private, and the one setup prompt that makes a published share actually usable today.
Read more
Field note
28 August 2026
Top-Level additionalContext Is Silent
A UserPromptSubmit hook can exit 0 with valid JSON and still inject nothing. Top-level additionalContext is silently ignored; nest it inside hookSpecificOutput.
Read more
Field note
27 August 2026
Hook Paths Don't Follow the Worktree
After Claude enters a worktree, $CLAUDE_PROJECT_DIR stays on the main checkout. Read cwd from hook JSON, or formatters and policy hooks hit the wrong tree.
Read more
Field note
26 August 2026
The Silent Opus Tax
A subagent with no model field inherits the model that called it, so mechanical helpers quietly run at Opus rates. One frontmatter line and one damping block fix it.
Read more
Field note
25 August 2026
Exit 1 Does Not Block
Claude Code hooks that exit 1 look like they blocked the tool. They did not. Official docs: only exit 2 stops it. Here is the pasteable PreToolUse gate.
Read more
Field note
24 August 2026
The Compact Drop
Path-scoped rules load into the conversation, so compact summarises them away. Move must-survive rules to root CLAUDE.md, or re-inject them with a hook.
Read more
Field note
22 August 2026
Stop Claude Over-Engineering
Anthropic documents that Opus 4.5 and 4.6 add extra files, unnecessary abstractions and flexibility nobody asked for. Here is the minimalism block that stops it, and the audit prompt that undoes it.
Read more
Field note
19 August 2026
The Conciseness Line
Opus 5 answers longer than earlier models by default, and the effort setting will not shorten it. Anthropic says to ask for conciseness explicitly. Here is the line that does it.
Read more
Field note
16 August 2026
The Output-Style Trap
A custom output style in Claude Code drops the built-in software-engineering instructions unless keep-coding-instructions is set to true. One line, easy to miss, and it quietly changes the quality of the work.
Read more
Field note
16 August 2026
Stop Claude Waffling: The ELI5 System
Claude got smarter and chattier. One instruction, report only in ASD-STE100 Simplified Technical English, removes the padding. Here is the output style you can paste once and keep.
Read more
Codex workflow
10 August 2026
Codex Automation: Turn Your Sidebar Into a Live Work Dashboard
Build a Codex automation that labels completed, ongoing and attention-needed tasks, turning your sidebar into a reliable live dashboard for parallel agent work.
Read more
Operating note
27 July 2026
The AI Advantage Is Initiative, Not Access
The same models and tools are available to almost everyone. The advantage comes from choosing one useful constraint, finishing the work and turning each result into a stronger operating system.
Read more
Production AI
29 July 2026
The Backend Is the AI Product
A polished AI interface is not a production system. The real product is the authentication, data, permissions, evaluation and recovery work behind the screen.
Read more
Founder note
31 July 2026
Founder-Led Distribution Is Now an Operating Skill
Being excellent in private is no longer enough. Founders need a repeatable way to turn real work into public evidence without becoming full-time content creators.
Read more
Customer work
4 August 2026
Customer Trust Compounds Before Revenue
The strongest customer relationships are built through clarity, useful challenge and consistent follow-through. That trust changes what a customer is willing to attempt with you.
Read more
Delivery standard
6 August 2026
Would You Trust Your AI Partner With a Customer?
AI capability is only one part of delivery readiness. A practical scorecard for deciding whether a partner can represent your standard when the work becomes real.
Read more
Build story
22 July 2026
Eight Builders, One Insight Engine: What Our First AI Build Challenge Taught Us
Eight builders had three weeks to turn a raw warehouse into a trusted AI product. Explore the captured results, the evidence behind every review, and an interactive improvement coach.
Read more
Open source
22 July 2026
I Open-Sourced My AI Advisory Team for Claude Code and Codex
One chief of staff, leads for the parts of your organisation, and a shared bench of specialists. Install the crew, update it safely and keep your own operating context.
Read more
Field note
8 June 2026
Your AI Agents Work While You're at the Gym, And That's the Point
Claude Code Remote Control and Codex mobile let founders monitor AI agents from their phone. How async AI workflows are reshaping work-life balance, who is already working this way, and how to set it up this week.
Read more
Field note
5 June 2026
How We Made navaigate.dev Agent-Native (And Scored 71)
We scored 21 on Cloudflare’s agent-readiness scanner. Here is exactly how we rebuilt three sites to Level 5, the three layers we built, why we refused to fake the endpoints, and the open-source playbook.
Read more
Field note
4 June 2026
Turn Any Photo Into a Full-Page Manga with One ChatGPT Prompt
We turned our founder into a manga hero using one ChatGPT image prompt. Steal it. Here is exactly how to do it yourself in minutes, in ChatGPT or straight from the terminal in Codex.
Read more
Field note
29 May 2026
The hard part of AI automation was never the tool
Everyone obsesses over which tool to use, n8n, Make, the latest agent framework. The tool was never the hard part. A field note on finding the automations that actually pay for themselves.
Read more
Founders' note
7 April 2026
Why I built NavBot
An honest founder’s note on why I stopped trusting hosted AI assistants and built my own. The architecture, the trade-offs, the security model, and the costs, for operators who would rather build than rent.
Read more
Skills + Workshop
5 April 2026
Claude Code + NotebookLM: infinite memory
Give Claude persistent memory, free content generation, and a growing AI Brain. Two open-source skills that connect Claude Code to Google NotebookLM, with a full training workshop included.
Read more
Claude Code skill
5 April 2026
LinkedIn Carousel skill
Generate professional branded LinkedIn carousel PDFs directly from Claude Code. Custom branding, AI-generated images, 7 slide types, and automatic caption writing. Five minute setup.
Read more
Case study
5 April 2026
Why I Rebuilt NavBot From Scratch, A Security Case Study
Most always-on AI assistant projects look great in a screen recording and fall over the moment a real adult tries to use one in production. The five things every operator should care about.
Read more

Put AI to work

Bring a real business problem.

NavAIgate helps you find the work worth automating, then builds the AI systems your team can use every day.

Book a callGet in touch

Field notes, skills, and the occasional case study.

CLAUDE.md Never Skips Permission Prompts

The Formatter Ping-Pong

CLAUDE.local.md Blocks AGENTS.md

Bypass Permissions in Project Settings Does Nothing

Safe Mode Isolates Your Setup

CLAUDE.local.md Is Personal

Nested CLAUDE.md Never Loads at Start

Standalone Hooks File Never Loads

How to Build Your First AI Agent: One Agent, One Job, One Owner

Orchestration, Not Autonomy

Skills Beat Prompts

Lowercase Matcher Never Fires

The Verification Bottleneck

`.mcp.json` Under `.claude/` Never Loads

Migrating off budget_tokens

Permissions in ~/.claude.json Are Ignored

Claude Code Ignores AGENTS.md

The CLAUDE.md Diet

Auto in Project Settings Does Nothing

Cut Your Claude Code Bill

Grok Bot Templates Are a Recipe

Top-Level additionalContext Is Silent

Hook Paths Don't Follow the Worktree

The Silent Opus Tax

Exit 1 Does Not Block

The Compact Drop

Stop Claude Over-Engineering

The Conciseness Line

The Output-Style Trap

Stop Claude Waffling: The ELI5 System

Codex Automation: Turn Your Sidebar Into a Live Work Dashboard

The AI Advantage Is Initiative, Not Access

The Backend Is the AI Product

Founder-Led Distribution Is Now an Operating Skill

Customer Trust Compounds Before Revenue

Would You Trust Your AI Partner With a Customer?

Eight Builders, One Insight Engine: What Our First AI Build Challenge Taught Us

I Open-Sourced My AI Advisory Team for Claude Code and Codex

Your AI Agents Work While You're at the Gym, And That's the Point

How We Made navaigate.dev Agent-Native (And Scored 71)

Turn Any Photo Into a Full-Page Manga with One ChatGPT Prompt

The hard part of AI automation was never the tool

Why I built NavBot

Claude Code + NotebookLM: infinite memory

LinkedIn Carousel skill

Why I Rebuilt NavBot From Scratch, A Security Case Study

Bring a real business problem.