Concepts - What Goes Where

System prompt vs skills vs memory vs wiki - the four places an agent’s brain lives, and how to decide where to put what.

A SlackHive agent’s brain has four layers. Each layer has a different purpose, a different lifespan, and a different place in the UI. Getting this right is the difference between an agent that sharpens over time and one that forgets everything.

The four layers

Layer	What it is	Where it lives	Lifespan
System Prompt	The agent’s identity - role, tone, rules	Instructions → System Prompt	Forever, until you edit it
Skills	Reusable procedures - how to do specific tasks	Instructions → Skills	Forever, until you edit them
Memory	Facts the agent learned from conversations	Instructions → Memory	Forever, until you delete them
Wiki	Shared reference material the agent looks up on demand	Knowledge Library → Wiki tab	Lives in shared folders, agent is assigned what it needs

1. System Prompt - who the agent is

The agent’s always-loaded identity. Every conversation starts with this in context.

Put here:

• Role: “You are a senior data analyst for Pelago”
• Tone: “Concise, direct, no filler words”
• Hard rules: “Never run DELETE or UPDATE queries”
• Domain scope: “Focus on bookings, revenue, and churn - not marketing”

Don’t put here:

• Long procedures (use a skill)
• Facts about specific people (use memory)
• Large reference material (use the wiki)

Kept short. A bloated system prompt wastes context on every turn.

2. Skills - how the agent does things

Slash commands the agent can invoke on demand. Each skill is one Markdown file with step-by-step instructions for a specific task. Skills stay out of the agent’s context until someone - a user, a scheduled job, or the agent itself during a chain - invokes them by name.

Put here:

• /weekly-report - steps to pull, format, and post the weekly digest
• /onboard-new-user - the full onboarding checklist
• /debug-checkout - the triage flow for checkout errors

Don’t put here:

• One-off logic (just tell the agent in Slack)
• Identity or tone (use the system prompt)
• Data the agent needs to read (use the wiki)
• An agent’s single always-on job - if the agent has one job it always does (e.g. birthday bot, daily digest), put the instruction in the system prompt and trigger via scheduled job. A skill adds ceremony without payoff.

A skill runs only when called. It doesn’t cost context until then. That’s the whole point - skills let an agent carry dozens of procedures at near-zero idle cost.

3. Memory - what the agent learned

Structured facts the agent writes down during conversations. Every memory is inlined into the system prompt every turn, so the agent always sees what it has learned - no skill call, no retrieval step.

Put here:

• user - “Kai leads the data team, prefers concise answers”
• feedback - “Don’t mock the database in integration tests - last quarter’s incident”
• project - “Merge freeze starts 2026-03-05 for mobile release”
• reference - “Pipeline bugs tracked in Linear project INGEST”

Don’t put here:

• Identity (use the system prompt)
• Procedures (use a skill)
• Large reference docs (use the wiki)

Memory grows organically. Every conversation is a chance for the agent to learn.

4. Wiki - what the agent can look up

Wiki folders live in a shared, platform-level Knowledge Library. Anyone on the team can create a folder, ingest a repo / file / URL into it, and SlackHive builds a Karpathy-style wiki (overview.md, concepts/, flows/, modules/). The agent is assigned the folders it needs and reads them on demand via the /wiki skill.

Put here:

• The codebase the agent works with
• Internal runbooks, design docs, API specs
• Product documentation

Don’t put here:

• Quick facts (use memory)
• Procedures (use a skill)
• Identity (use the system prompt)

Wikis can be large. The agent only reads the pages it needs.

Decision flow

Ask yourself two questions:

1. How often is this relevant?
   • Every turn → System Prompt
   • Sometimes, on demand → Skill, Memory, or Wiki
2. Where does the content come from?
   • You write it → System Prompt or Skill
   • Agent learns it from a conversation → Memory
   • External source (repo, docs, URL) → Wiki

                    Relevant every turn?
                   /                    \
                YES                      NO
                 |                        |
          System Prompt              Written by me?
                                    /              \
                                 YES                NO
                                  |                  |
                          Procedure or fact?     From external source?
                         /         \              /            \
                      Procedure    Fact         YES              NO
                         |          |            |               |
                       Skill      Skill        Wiki           Memory
                            (short, reusable)   (agent       (agent
                                                 reads        learns,
                                                 on demand)   always sees)

Worked examples

You want the agent to…	Where it goes
Always write in a polite, formal tone	System Prompt
Run the same 8-step SQL triage flow when revenue drops	Skill
Remember that the CEO prefers charts over tables	Memory
Reference the codebase at github.com/org/repo on demand	Wiki
Never edit production config files	System Prompt
Draft a standard PR summary in the team’s format	Skill
Note that Tuesdays are always low due to maintenance	Memory
Look up the API reference when asked about an endpoint	Wiki

How they interact

The four layers are not isolated - they compose:

1. System Prompt loads every turn. Sets the lens.
2. Memories load every turn, inlined into the system prompt. The agent always knows what it has learned.
3. Skills load when invoked. /wiki and your custom skills.
4. Wiki loads page-by-page when the agent reads it via /wiki.

Memories and the system prompt are always in context; skills and wiki pages load on demand. That keeps the turn cost bounded while everything the agent has learned stays visible.

Quick reference: when to use each

You want to…	Use
Give the agent a permanent role, tone, or hard rule	System Prompt
Define a repeatable workflow (triage, report, audit)	Skill
Trigger that workflow on a schedule, hands-free	Skill + Scheduled Job
Let the agent remember something it learned in a conversation	Memory
Give the agent a codebase, runbook, or doc to look up	Wiki
Delegate to the right specialist automatically	Boss Agent

Next steps

Creating Agents

Set up the system prompt and pick a persona in the 5-step wizard.

Coach

Tune skills, system prompt, and wiki via chat - no hand-editing.

Memory

How agents save and manage memories.

Knowledge Library

Shared wiki folders. Owners ingest repos / files / URLs; agents are assigned the folders they need.