State lives on disk, never in a conversation. A fresh session with no memory of this project can pick up exactly where the last one stopped by following the four steps below. Compaction, a session limit, or a machine restart costs nothing beyond the entries that were mid-write.
cd ~/repositories/patterns
python3 tools/next-batch.py --statusPrints per family, done over total, plus the overall count. The definition of
done is a PASS from tools/check-structure.py. Nothing else counts, so a
partially written entry is correctly reported as unfinished.
python3 tools/next-batch.py --size 8Prints a JSON array of the next eight unfinished entries, already shaped as the
args payload the authoring workflow expects. Add --family 17-ai-agentic to
restrict it to one family.
Invoke the Workflow tool with the lean authoring script and pass the JSON from
step 2 as args.
scriptPath: ~/.claude/projects/-Users-mirzaiqbal-repositories-patterns/
<session>/workflows/scripts/patterns-lean.js
args: <the JSON array from step 2>
The script runs in chunks of eight, skips any entry that already passes the gate, and writes nothing for an entry that is already complete. Re-running a finished batch costs one cheap gate run per entry.
python3 tools/gen-indexes.py
python3 tools/check-structure.py
python3 tools/check-prose.py
git add -A
git commit -m "feat(<family>): add <entries>"
git pushNever let more than one chunk of work sit uncommitted. A limit hit between commits is the only way to lose finished work.
The authoring plan used to live in the assistant's working memory, which made compaction expensive. It now lives in three files.
| File | Holds |
|---|---|
docs/AUTHORING-QUEUE.json |
Every remaining entry, its slug, family, and path |
docs/ENTRY-TEMPLATE.md |
The 18-dimension contract every entry must satisfy |
tools/next-batch.py |
The only source of truth for what is still outstanding |
An assistant needs no prior knowledge of this project. It needs this file.
Entries in the queue carry an empty focus field. A focus brief tells the
authoring agent which distinctions, failure modes, and citations matter for that
specific pattern, and it is the difference between a competent entry and a sharp
one. Fill it in before running a batch where the pattern has a well known trap,
for example Singleton and double-checked locking, or Observer and the lapsed
listener leak.
An empty focus is acceptable. The template alone produces a valid entry.
- Flipping the repository from private to public.
- Refreshing the Kimi API key if the second council engine is wanted.
- Raising
CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTSbeyond the default. - Approving any change to the 18-dimension contract.
Everything else runs without input.