/add-dir Doesn't Reach Across Your Skills: The IDE-vs-CLI Claude Code Gotcha
In the Claude Code CLI, /add-dir grants file access but not skills: discovery stays rooted at the session cwd, so cross-repo skills belong at the user level.
The build log
Notes from running agents in production.
Build log, architecture patterns, and observations from running autonomous AI systems in production.
-
-
Bootstrapping a New Agent Station, and Giving It a Phone
Give every machine a written identity doc before an agent runs on it. The 'cannot reach' list is the load-bearing half, and the phone comes last.
-
Avoid Unnecessary Date String Round Trips
A date value was formatted for display and then handed back to a database parser. The durable lesson is to preserve typed values, use parameters, and perform only authorized, reviewed, validated, and reversible data changes.
-
When a Manual Workflow Hurts in Seven Small Ways, the Seven Ways Are Your Spec
One manual ticket close surfaced seven papercuts, and each one became a hardcoded default or halt condition in the tool. A friction list is a spec.
-
The One-Shot Data-Repair Script as a First-Class Artifact
The console UPDATE is the most dangerous tool you can point at production. A four-part repair-script discipline: dry-run, assert, idempotent, JSON record.
-
I closed a client ticket from the wrong repo on purpose
Closing a client ticket from outside its own repo proved that shared agent skills belong one level above any checkout, where every session can call them.
-
One Approval Whole Operation
Once a plan is approved, running it should not turn into a separate yes at every step inside it.
-
Turning a Stack of AI Transcripts Into Beats for Under a Dollar
Five conversation transcripts added up to 858KB of Markdown, roughly 215,000 input tokens, all set to repeat the same system prompt in full, five times over.
-
A Backlog State Should Tell You What Happens Next
A single backlog label hid two different conditions: deliberate deferral and an external dependency. Making next-action ownership, follow-up, review timing, and rationale explicit produced a more truthful queue and reduced repeated context reconstruction.
-
Write the Handoff Before You Close the Session, or Pay to Rebuild State Next Time
Skip the end-of-session handoff and the next session pays to rebuild state you already held. The six-section structure that turns resuming into a lookup.
Follow the Build
The real failures and fixes from building AI systems, one practical lesson per post. Get the next one in your inbox.