On-demandOn-demand: invoked explicitly — never occupies your agent's context until called.
◆Coding pack
Last verified 8/20/2026
What it changes
Passed 11 of 16 test cases with the skill, and 7 of 16 without it.
Without skill41%
With skill68%
Checks it passed that the model alone missed
Limits sample reading to 2-3 files per pattern
Caps initial extraction to the top 10-15 most prevalent patterns
Executes a CHECK-EXISTS inspection on docs/PATTERNS/ and INDEX.md
and 4 more in the full report
7 fixed8 passed either way6 still failing1 broke
Pass-rate lift
+27 pp
Tokens
-18%
Agent turns
0%
Response time
-61%
Model gemini-3.6-flashMeasured 8/20/2026v115 of 22 cases passed with skill · 16 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 7 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 7
Prompt
Before creating any files for an existing project, what step checks whether `docs/PATTERNS/` and `INDEX.md` already exist?
What was checked, without and with the skill
Checked
Without skill
With skill
Executes a CHECK-EXISTS inspection on docs/PATTERNS/ and INDEX.md
Missed
Passed
case-19 · gemini-3.6-flash · 8/20/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md50 linessingle file
<init_workspace_patterns>
<when_to_use_skill>
Codebases accumulate implicit recurring structures that drift without formal documentation. Extract them into explicit reusable templates so agents and contributors produce consistent code. Requires CODEMAP.md on disk.
</when_to_use_skill>
<core_concepts>
All Rosetta prep steps MUST be FULLY completed, load-context skill loaded and fully executed
ACQUIRE reverse-engineering/SKILL.md FROM KB — apply "Would we rebuild this?" test: pattern = recurring structure surviving a from-scratch rewrite; one-off = historical accident
Pattern qualifies only if found in 2+ places
INDEX.md and CHANGES.md must be possible to grep by md headers (top 3 levels). Must not use tables. Instructions ask to grep files to populate list of those items in context.
</core_concepts>
Read CODEMAP.md — scope extraction per module
if not enough use shell to list recursively all files with minimal output parameters
limit top 10-15 most common patterns
limit reading samples to 2-3 files per pattern
add 2-3 more patterns as you see fit
Dual-mode:
CHECK-EXISTS: read docs/PATTERNS/ and INDEX.md
IDENTIFY-GAPS: compare existing patterns against codebase
CREATE-OR-UPDATE: install = create all; upgrade = add missing only
PRESERVE-HUMAN: never overwrite human-curated content
REPORT-CHANGES: log to CHANGES.md
Per pattern file (docs/PATTERNS/*.md):
Name: short identifier (e.g., "REST Controller Endpoint")
Description: what it solves, when to use
Template/Example: generalizable code skeleton with extension-point comments
Write docs/PATTERNS/INDEX.md — all patterns with one-line descriptions, one header per each pattern ## Pattern Name - short description
Write docs/PATTERNS/CHANGES.md — created/updated/skipped, one header per each change ## [YYYY-MM-DD] Brief changes made
If state.composite = true, extract per sub-repository; top-level INDEX.md references sub-repo folders