Too many developers treat AI like a magic 8-ball. Throw a vague prompt at it, cross your fingers, hope the output does not break production. Prompt engineering is not where the leverage sits. It is in the Claude Code context you hand over.
Without the right data in front of it, you are asking a genius to work a puzzle blindfolded. Most of the standard advice about AI coding fails for that reason: it ignores the local knowledge bottleneck.
Why your Claude Code context strategy is failing
Most devs expect the agent to just know the architecture of their custom WooCommerce hooks or the shape of their Nginx config. It cannot, so it hallucinates or hands back a tutorial-grade answer. The fix is to treat your local files as a primary knowledge base.
Anthropic made a pragmatic call here. Rather than pushing everything through a slow vector search (RAG) by default, Claude Code runs bash commands against your directory, and they found raw bash search often beats embeddings on coding tasks. It is quicker at turning up the one helper function buried somewhere in a legacy plugin.
On the infrastructure side, I wrote up how to run Claude Code with local models.
Step 1: centralize your knowledge silos
I keep everything locally: marketing briefs, LinkedIn posts, API documentation, transcripts of client meetings. MacWhisper handles that last one, turning spoken conversation into searchable text files. Once those files sit inside the Claude Code context, the agent has the business logic behind the code and not only the syntax.
Make a /knowledge folder in your project root and dump the PDFs, text files and architecture notes in there. Claude picks them up with grep or find while it reasons.
Step 2: security and the .claudignore
This is the messy part. Hand Claude the whole directory and it can read your .env files or your SSH keys and ship them off to the LLM provider. Be strict about permissions.
Always keep a .claudignore file. It works like .gitignore, except that it tells the agent what to stay away from. That matters most when you are running in sandboxing mode or with the --dangerously-skip-permissions flag.
# .claudignore - Keep the agent out of the sensitive stuff
.env
*.pem
node_modules/
vendor/
.git/
wp-config.php # Unless you need it to debug DB issues specificallyThe gotcha: bash vs. vector search
Bash search is good at finding things and hungry with tokens. On a large directory the agent will happily read files that have nothing to do with the task. I have burned $10 of credits in an hour that way, purely because I never pruned the context. A script that gathers only what a given task needs pays for itself quickly.
For heavier workflows, there is more in my post on applying agentic coding to WordPress.
A custom context aggregator script
When you are working on one feature, do not make Claude guess which files matter. This bash script builds a single feature context file you can drop straight into the Claude Code context.
#!/bin/bash
# bbioon_context_bundle.sh
# Aggregates relevant files into one context file for the agent
OUTPUT="claude-feature-context.txt"
echo "--- FEATURE CONTEXT START ---" > $OUTPUT
# Add your specific local docs
cat knowledge/api-specs.md >> $OUTPUT
cat knowledge/business-logic.txt >> $OUTPUT
# Add relevant source files
tail -n +1 src/Plugin/Core.php src/Plugin/Helper.php >> $OUTPUT
echo "Context bundled in $OUTPUT. Feed this to Claude Code."If this kind of setup is eating your dev hours, I can take it off your plate. I have been wrestling with WordPress since the 4.x days.
What it comes down to
Personalizing the agent means moving knowledge out of your head, and out of your Slack messages, into local files it can read. Bash-based search plus strict ignore rules is most of the setup. Get that far and the agent starts behaving like a senior partner instead of a confused intern.
