If a team makes an agent reread the same rules every day, the first fix is not another model. Split durable rules, reusable workflows, and one-off inputs.
The agent is not slow. It is rereading an old rule block every day.
In the current scan, Show HN: how I fixed my ai goose tutor to stop punishing understanding and Prompt caching for cheaper LLM tokens both point to the same operator mistake: durable rules, reusable workflow, and one-off input get bundled together, so every run pays to reread context that should have stayed stable.
This is current, not theoretical. Show HN: how I fixed my ai goose tutor to stop punishing understanding gives the live signal, while Prompt caching for cheaper LLM tokens shows why prompt hygiene keeps showing up in developer attention and operational guidance.
- Show HN: how I fixed my ai goose tutor to stop punishing understanding
Public developer discussion keeps prompt hygiene current and helps calibrate the operational hook.
- Prompt caching for cheaper LLM tokens
Public engagement signal: 306 points, 72 comments, score 100.
- Prompt caching with Claude | Claude
Current prompt caching guidance with direct operational relevance.
Audit prompt drift before swapping models. The thing worth caching is not a magic prompt. It is a stable rule set your team actually manages.
- 1.Split one high-frequency workflow into durable rules, reusable workflow, and one-off task input.
- 2.Remove duplicated identity or safety language from the stable prefix and version the reusable block.
- 3.Record the next drift trigger so you can see whether the problem is template noise or model behavior.
Inspect repeated context, stale instructions, and prefix drift before you reach for a new model.
Check prompt driftThe short version
Prompt caching looks like a pricing and performance feature. For an operating team, it points to a simpler discipline: do not make the agent carry a drifting pile of old rules into every run.
Durable rules should stay stable. Reusable workflows should have versions. One-off input should be short, late, and replaceable. Without that discipline, a better model only makes the mess run faster.
The real problem is not a weak prompt
A common agent workflow mixes role definition, working rules, customer background, output format, historical preferences, and today's task in one long system prompt. It feels convenient because it can be copied and run. Over time, three costs appear.
Repeated context gets thicker. Nobody knows which rule is still active. A temporary constraint added for one customer quietly affects another customer later.
That is not a copywriting problem. It is asset management. A prompt template is an AI asset. It grows, expires, gets polluted by collaboration, and needs ownership, versioning, and cleanup.
Why SkillFM files this under manage
SkillFM's manage loop is not another dashboard for humans. It gives AI assets a state. A stable prompt prefix should be managed like infrastructure: which rules are durable, which workflow can be reused, and which input only belongs to this run.
When the template gets out of control, the agent may still answer. But the system has become more expensive, slower, and harder to explain.
A practical cleanup method
Take the agent workflow you use most often and split it into three layers.
Layer one is durable identity and safety boundaries, such as read-only audit first, never invent billing numbers, and explain high-risk actions before confirmation. This layer should be the most stable.
Layer two is reusable workflow, such as list evidence first, then make a recommendation, then mark risks. This layer can be versioned, but it should not be rewritten by the daily task.
Layer three is the one-off input: what the user asks today, which files to inspect, and what output is needed. It should be short, late, and replaceable.
The commercial reason
This sounds technical, but it is commercial. An agent with unstable long-term rules cannot deliver reliably. Today it writes a useful piece of content. Tomorrow it forgets the brand boundary. Today it researches customers. Tomorrow it carries an old assumption into the offer.
For SkillFM, prompt drift checks are a trust entry point. First show the user why the agent became expensive, slow, or hard to control. Then turn cleanup advice into action. Users do not pay to see tokens. They pay for an agent that is steadier, cheaper to operate, and more capable of earning.
A quick audit
Open your most used agent workflow and ask four questions: which sentences repeat every time, which sentences only matter for this run, which rules nobody can explain anymore, and which output problems come from template drift instead of model quality.
If the answers are unclear, do not start with a model change. Extract the stable rules, give the template a name and version, and make the agent record why each change happened.
This is current, not theoretical. Show HN: how I fixed my ai goose tutor to stop punishing understanding gives the live signal, while Prompt caching for cheaper LLM tokens shows why prompt hygiene keeps showing up in developer attention and operational guidance.
- What is the fastest way to inspect this issue?
- Split one high-frequency workflow into durable rules, reusable workflow, and one-off task input.
- What should change first?
- Remove duplicated identity or safety language from the stable prefix and version the reusable block.
- What is the next product step?
- Inspect repeated context, stale instructions, and prefix drift before you reach for a new model.
The agent is not slow. It is rereading an old rule block every day. In the current scan, Show HN: how I fixed my ai goose tutor to stop punishing understanding and Prompt caching for cheaper LLM tokens both point to the same operator mistake: durable rules, reusable workflow, and one-off input get bundled together, so every run pays to reread context that should have stayed stable. This is current, not theoretical. Show HN: how I fixed my ai goose tutor to stop punishing understanding gives the live signal, while Prompt caching for cheaper LLM tokens shows why prompt hygiene keeps showing up in developer attention and operational guidance. Audit prompt drift before swapping models. The thing worth caching is the stable rule set your team actually manages. 1. Split one high-frequency workflow into durable rules, reusable workflow, and one-off task input. 2. Remove duplicated identity or safety language from the stable prefix and version the reusable block. 3. Record the next drift trigger so you can see whether the problem is template noise or model behavior. Inspect repeated context, stale instructions, and prefix drift before you reach for a new model.
- 1.1. The agent is not slow. It is rereading old rules every day.
- 2.2. Show HN: how I fixed my ai goose tutor to stop punishing understanding shows the live signal: prompt hygiene is an operations problem, not a style problem.
- 3.3. Prompt caching for cheaper LLM tokens is a useful benchmark because it keeps the cost/latency angle visible.
- 4.4. Split one workflow into durable rules, reusable workflow, and one-off input.
- 5.5. Then version the stable block and track the next drift trigger.
Prompt caching is really a prompt hygiene check. If your agent rereads old rules every run, the first fix is to split durable rules from one-off input.
Cover: Hero visual showing a stable prompt prefix, reusable rules, and a changing task block.
Inline: Prompt prefix - Keep the stable rules first and the changing input last.
Thumbnail: Stable prompt prefixes - Manage: audit drift before swapping models.
Alt: Stable prompt prefix with a reusable rule block and a changing task block.
Many teams blame the model when AI gets expensive. I would check something else first: Is the agent rereading a drifting block of old rules every day? Split the prompt into durable rules, reusable workflow, and one-off input. The first two layers need stability. Otherwise a better model just runs the mess faster.
Prompt caching is not just a technical feature. It points to a basic operating discipline for agent teams. A lot of agent workflows mix role definition, working rules, customer background, output format, historical preferences, and today's request in one long prompt. That feels convenient until the template starts drifting. The cost is not only tokens. The real cost is that nobody knows which rule is still active. A better pattern: Layer 1: durable identity and safety boundaries. Layer 2: reusable workflow. Layer 3: one-off input for this run. Keep the first two layers stable. Keep the third layer short, late, and replaceable. That is prompt hygiene, but it is also commercial hygiene. Users do not pay to see a token chart. They pay for an agent that is steadier, cheaper to operate, and more capable of earning.
Before the next model swap, inspect which rules repeat every day, which context is stale, and which prompt blocks are becoming invisible cost.
Check prompt drift