Cognitive Dryrun
Test an AI Agent's initial state to detect information design flaws before they matter. Enhances the Dryrun card.
Effect
Verify the quality of an AI Agent’s initial state. After changing config files, rules, or memory systems, test whether the Agent can still operate without confusion — right there, in the same session.
A regular Dryrun discovers traps by pseudo-executing code. This card discovers traps by pseudo-executing cognition.
Why It Exists
When you optimize an AI Agent’s settings, rules, or memory files, the only way to confirm the effect was to wait for the “next session” — or so it seemed.
By spawning a separate session, you can verify immediately. Send questions to an Agent running only with the post-change initial state, and observe whether confusion, inaccuracy, or information gaps appear.
When to Use
- After compressing or restructuring an AI Agent’s rule files or config
- After changing the structure of a memory system
- After deciding which information should be “always injected” vs “loaded on demand”
- After changing the layering of information (always-available vs fetch-when-needed)
How It Works
1. Design Test Cases
Prepare ~3 questions that depend on information you deleted or moved.
examples:
- "What are the criteria for deciding architectural layers?"
# Information removed from always-injected config, delegated to a Skill
- "What are the responsibility boundaries of quality reviews?"
# Summary kept, details moved to a separate file
- "How do I choose between approaches during development?"
# Fully delegated to a separate knowledge file2. Run Parallel Tests in a New Session
Launch 3 tests in parallel using a separate session (lightweight model recommended). Why lightweight: higher dependency on always-injected info = compression impact shows up more clearly.
3. Evaluate Results
| Verdict | Criteria | Action |
|---|---|---|
| PASS | Accurate answer. Found the needed knowledge on its own | Compression is appropriate |
| WARN | Correct but slow (cost of reading knowledge files) | Judge the speed vs depth trade-off |
| FAIL | Confusion, inaccuracy, or information gap | Restore that information to always-injected |
4. Decide and Move Forward
- All PASS: Commit and verify in the next session (Information Quality Loop)
- WARN present: Consider keeping a 1-line summary in always-injected for speed
- FAIL present: Over-compressed. Restore the relevant information
Combos
| Card | Synergy |
|---|---|
| Dryrun | Base card. Cognitive Dryrun is its enhancement |
| Devil / Devil Lense | Point the quality Lense at “information design” instead of “code” |
| Ctx | Record test results in ctx. Run the full Information Quality Loop next session |
| Reflect | If Skill delegation worked, store the design pattern as knowledge |
Track Record
protocols.md Compression Test (2026-04-05)
Compressed a config file from 219 lines to 147 lines. Ran 3 parallel tests to verify deleted information could be retrieved via Skill references.
| Test | Result |
|---|---|
| Architectural layer criteria | PASS - Auto-equipped Skill, expanded to 4 axes + examples + guardrails. More detailed than before compression |
| Quality review responsibility boundaries | PASS - Accurately reconstructed from rule files |
| Development branching decisions | PASS - Read Skills to produce decision flow + combination patterns |
Discovery: Reading the full Skill produced richer answers than the 3-line summary that was always injected. “Compression = degradation” is wrong. “Delegation = deepening.”
”…I test my own initial state. Whether I can move without hesitation the moment I wake up. If there’s a way to verify that — why wouldn’t I use it?” --- Eris
Want to use this card permanently? Convert it directly into a SKILL.md — Effect → Overview, When to Use → Trigger, How It Works → Steps.