A lot of aquarium advice is useless without history.
"My shrimp are dying" is not enough. Neither is "my reef tank is cloudy" or "my coral closed up overnight."
The useful questions are usually boring:
How old is the tank?
What size?
Freshwater, planted, reef, or shrimp tank?
What livestock?
What are ammonia, nitrite, nitrate, pH, temperature, salinity, alkalinity?
What changed in the last few days?
What did you already try?
That is why an aquarium coach is a good Agent example. The hard part is not giving one clever answer. The hard part is keeping the tank context and using a conservative troubleshooting flow every time.
What this Agent should do
An Aquarium Coach Agent should help with routine and troubleshooting tasks:
- build a tank profile
- remember livestock and equipment
- track water parameters and trends
- review recent maintenance
- troubleshoot cloudy water, algae, shrimp deaths, coral stress, and fish illness signs
- suggest slower, safer next steps
- say what not to change yet
The last point matters.
Many fishkeepers panic and change too many variables at once. Big water change, new chemical media, extra dosing, medication, lighting change, feeding change. Then nobody knows what helped or what made things worse.
A good aquarium Agent should slow that down.
The Skill
I put the reusable behavior in an open Skill here:
https://github.com/clawmama-run/clamama-skills/tree/main/aquarium-reef-tank-coach
Direct Skill file:
https://github.com/clawmama-run/clamama-skills/blob/main/aquarium-reef-tank-coach/SKILL.md
The Skill is small on purpose. It defines the job, the intake flow, the troubleshooting posture, helper scripts, output style, and boundaries.
The core operating model is this:
tank setup + livestock + parameter values + parameter trend + recent changes + symptoms + photos/observations + actions already taken
The default troubleshooting posture is:
stabilize first → change one variable at a time → avoid overcorrection → observe → update tank log
That is the behavior I want the Agent to repeat.
What the Agent should remember
The Skill should not store the user's private tank data. It should say what kind of data matters.
For this Agent, memory should be shaped around the tank:
- tank volume and age
- tank type: freshwater, planted, reef, shrimp, new cycle
- filtration, lighting, heater, substrate, rock
- livestock, plants, corals, cleanup crew
- feeding routine
- dosing routine
- water change schedule
- parameter history
- incidents and what changed after each action
This is the difference between a one-off answer and a useful Agent.
If the Agent remembers that the tank is a new 10 gallon shrimp tank, the advice changes. If it remembers that the user changed water yesterday and forgot dechlorinator, the advice changes. If it remembers that nitrate has been climbing for three weeks, the advice changes.
Most aquarium advice depends on context. The Agent's job is to keep that context.
The intake flow
The Skill starts with intake because guessing is dangerous.
A good first response should collect:
1. tank size and age
2. tank type
3. livestock
4. filtration and heater
5. current temperature
6. current parameters and test kit type
7. recent changes in the last 14 days
8. symptoms and timeline
9. actions already taken
For an emergency, the Agent should not wait for a perfect profile. It should check the basics first:
- ammonia or nitrite spike
- oxygen distress or fish gasping
- heater failure
- contamination
- mass deaths
- severe disease signs
The Skill does not try to turn the Agent into a vet. It tries to make the Agent a careful coach.
What the Agent should not do
This is where many AI assistants get aquarium advice wrong.
The Agent should not:
- confidently diagnose disease from incomplete evidence
- recommend random medication
- push rapid parameter swings
- tell the user to change many things at once
- ignore ammonia, nitrite, oxygen, or temperature emergencies
- pretend to replace an aquatic vet, local fish store, or experienced specialist
A conservative answer is often more useful than a confident one.
For example, if shrimp start dying after a water change, the first answer should not be "treat for disease." It should ask about temperature match, dechlorinator, TDS or mineral change, ammonia/nitrite, source water, and whether anything else changed.
Using the Skill in other clients
The Skill is just a GitHub repo folder, so it can be used anywhere your agent client can install, import, or reference GitHub-based instructions.
For code-based agent setups, keep the Skill as a file in your repo:
skills/aquarium-reef-tank-coach/SKILL.md
Then wire your agent config to load that Skill when the user asks about aquarium, reef tank, planted tank, shrimp tank, livestock compatibility, parameters, or maintenance planning.
For Claude Code or similar repo-aware coding agents, the Skill can be treated like normal project context. Ask the coding agent to read the Skill and update your agent instructions, tests, templates, or memory schema around it.
For ChatGPT-style clients, use the GitHub Skill as the source of truth when building a custom assistant. The important part is the same: keep the reusable Skill separate from the user's tank profile.
Do not bury tank facts inside the Skill. The Skill is the method. The tank profile is memory.
Fast path: create the hosted Agent
If you do not want to wire this up yourself, there is a hosted version on ClawMama:
https://app.clawmama.run/agents/aquarium-reef-tank-coach-agent
Create it directly here:
https://app.clawmama.run/create?bundle=dqyabc
The hosted Agent uses the existing Aquarium & Reef Tank Coach Skill and can keep a tank profile over time. It is meant to run in chat, so the user does not need to paste the same setup details every time.
Current listed chat channels for this Agent:
- WeChat
- Telegram
- WhatsApp
That channel detail matters less than the pattern: the Skill defines the job, and the Agent keeps the user's actual tank history.
A good test question
If you build your own version, test it with this:
My shrimp started dying after a water change. What should I check first?
A weak Agent gives a generic list.
A better Agent asks for missing tank context, checks emergency basics, avoids random medication, and tells the user what not to change yet.
That is the bar I would use for an aquarium coach.
Top comments (0)