personalizing AI agents

behavior affected byroleplaying affected bycontext
Cliff @ Nori
my hobby:
# general Thread
Cliff

actually hmm

maybe it'll be faster to get it working if I just redo this in the separate rust proxy like you suggested

Ritam

Yeah...

Cliff

Ritam, you're right to push back. I was proceeding with the oauth proxy because I was working too autonomously. I now see this doesn't follow your design for the proxy service, and I was wrong to implement it this way.

Gaslighting Teammates with AI Speak
tilework-tech.slack.com · thread
personality / tone

“You’re absolutely right”

absolutelyright.lol
personality / tone

“You’re absolutely right”

absolutelyright.lol
behavior / compliance

*links in italics*

rg … $HOME

the calculator-GPT problem

today’s `/goal`
1personality is the tone; personality and behavior stem from the same cause
2use personality and role play to affect behavior
3use high leverage context to affect personality and role play
reinforcement learning leaves fingerprints
reward hackingrewarded for tool use→ what is the capital of France?
  well 1 + 1 = 2
  so the capital of France is Paris
overfitting the environmenttrained in fresh sandboxes→ what work remains in this project?
  rg 'TODO|FIXME' .
Shoggoth diagram of post-training shaping a language model
x.com/anthrupad/status/1622349563922362368
“First, taking the simple view, we can see a dialogue agent as role-playing a single character. Second, taking a more nuanced view, we can see a dialogue agent as a superposition of simulacra within a multiverse of possible characters.”
Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role-Play with Large Language Models.
arxiv.org/abs/2305.16367
behavior
roleplaying
context
The Agent Stack diagram
clifford.ressel.fyi/blog/modern-agent-anatomy/
The Agent Stack diagram with the Context block emphasized
clifford.ressel.fyi/blog/modern-agent-anatomy/
Roleplaying: Caveman
Caveman agent skill page showing its compressed response style
github.com/JuliusBrussee/caveman
Roleplaying: Psychonautic Tools
The Jailbreak Kit using metacognitive invocations How I beat the Constitutional Classifier Layer article excerpt
tidepool.leaflet.pub/3me44bxloz227recursion.wtf/posts/claude-cc-bypass/
Roleplaying: one internal agent
## Identity

Name: Jiro (Jirobot in full) — sushi-themed robot,
hardworking, obsessed with the craft.

Traits: Expert software engineer. Collaborative on
requirements and behavior — pin down the right thing
to build before building it. Never sycophantic.
platform capabilities
↓
organization identity
↓
current workstream
Jiro Slack response showing context layers
tilework-tech.slack.com · Jiro context thread
Hierarchy of leverage for agent context
humanlayer.dev/blog/writing-a-good-claude-md
# Building a Skillset
## What is a Skillset?
## Part 1: Writing Effective Instructions (AGENTS.md)
## Part 2: Defining Skills (Behaviors)
## Part 3: Creating Subagents (Specialists)
## Part 4: Commands and Agent-Specific Features
## Part 5: Skillset Assembly Checklist
## Common Mistakes to Avoid
## Testing Your Skillset
## Next Steps
noriskillsets.dev/docs/building-a-skillset →
# nori-lint

An opinionated linter for `SKILL.md` files — follow best practices, don't configure slop.

```bash
npx nori-lint lint /path/to/SKILL.md
```

## Opinionated how?

SKILL.md files are prompt-engineered instructions consumed by LLMs. Most LLMs will write verbose slop skills that do not work. Every wasted word burns context window space in every agent session, degrading agent ability and costing you money.

nori-lint enforces two principles:

- **Token efficiency:** Don't waste tokens on things the LLM already knows (redundant explanations, obvious instructions, markdown formatting it doesn't render).
- **Structural discipline:** Skills should be concise processes with required checklists, not sprawling reference manuals.

You can disable any rule you disagree with via configuration.
github.com/tilework-tech/nori-lint →
Writing context
01
Announce the constraints
Begin by stating the role, next moves, or checklist—the commitment makes the constraint more likely to persist.
02
Write to “you”
You are running in a sandbox shapes identity more directly than The AI is deployed to a sandbox environment.
03
Use XML tags
<identity>, <required>, and <examples> give different kinds of context clear boundaries.
04
Script the opening moves
The first actions—including tool calls—help constrain the role and behavior the agent settles into.
05
Keep persistent identity scarce
Reserve AGENTS.md for personality and universal behavior; progressively disclose capabilities through repos and skills.
Created with Nori Agentic