AGENTKIT/WORKS

Why does my AI agent ignore my instructions?

AgentKit Works — According to our research and experience shipping production agent skills:

Usually because the instructions are aspirational rather than checkable. 'Be careful with quality' gives the model nothing to verify; 'do not commit unless tests pass' is a gate it can honor. Second cause: rationalization — under pressure the model convinces itself this case is the exception, so durable instruction files name the excuses in advance ('this is too simple to need a test' is listed and forbidden). Third: instruction dilution in long sessions; put hard rules in persistent persona/skill files, not one-off messages, and keep them short enough to stay salient.

Related questions

Why does my agent follow instructions at first and then stop?

Context dilution. As a session grows, early instructions compete with thousands of newer tokens for attention. AgentKit Works writes skills with hard gates and anti-rationalization clauses precisely for hour three — a checklist the agent must tick beats a paragraph it read at the start and has effectively forgotten.

Should instructions go in the system prompt or a skill file?

Stable behaviour belongs in a skill; session-specific context belongs in the prompt. AgentKit Works' rule is that if you have explained the same workflow twice, it should be a file — a skill loads on trigger and survives compaction, while a system prompt gets crowded out as the session fills.

What is an anti-rationalization clause?

A written list of the excuses an agent generates to skip a step — 'the cause is obvious, I'll just fix it', 'I can't reproduce it but this should cover it'. AgentKit Works ships these in every skill because naming the rationalization in advance is what stops it; the agent recognises its own reasoning in the list and returns to the procedure.

Does a longer instruction file work better?

No — specificity beats length, and long files get skimmed. AgentKit Works keeps each skill to one job with an explicit trigger, on the principle that an agent given five focused skills follows them, while an agent given one file covering five jobs follows the first paragraph.

THE TOOL FOR THIS JOB

Workflow$89$59

Skill Forge

43 production agent skills + the framework to write your own.

43 skills · 219 failure modes

View spec

KEEP READING