Skip to content

Latest commit

 

History

3 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

auxfirst

A Claude skill for building agents an organisation can actually own.

Score every action an agent can take, cap its autonomy at the hottest dimension, and write the mandate, exceptions and owner's manual behind it.

Skill: validated License: CC BY 4.0 Source: auxfirst.com

Install · What it does · The one rule · Contents · Site


What it does

Most agent guidance answers "can the model do this?". This one answers a different question, the one that actually blocks production:

Given an action this agent can take, what does the team owe before it is allowed to take it unsupervised?

You give it an agent — a copilot, an automation, a workflow that writes to a system of record. It gives you back written artefacts: a scored action inventory, a mandate with an enforcement mechanism on every row, exception handling with named recipients and deadlines, and a readiness verdict out of 24.

It is deliberately not a survey of agent design. It assumes the agent can do the task and asks whether the organisation can own it.

The one rule everything hangs from

An action is as hot as its hottest dimension. No averaging.

Four cool dials never buy back one hot one. "Send a routine invoice" scores 1 on reversibility, 0 on blast radius and 1 on authority — and is still a HIGH-band action, because exposure and commitment are hot. A team scoring by average calls that low-medium and ships it on auto-run.

$ python skills/auxfirst/scripts/heat.py --score 1,0,3,3,1 --action "send an invoice"

  reversibility  1
  blast radius   0
  exposure       3  <- hottest
  commitment     3  <- hottest
  authority      1

  BAND            HIGH  (highest single score, never the mean)
  MAX AUTONOMY    Approve-each-action
  POSTURE         A named accountable human approves each instance before
                  execution, logged with rationale and identity.

Install

Claude Code — as a plugin (one command, updates with the repo):

/plugin marketplace add auxfirst/auxfirst-skill-global
/plugin install auxfirst@auxfirst

Claude Code — as a plain skill:

git clone https://github.com/auxfirst/auxfirst-skill-global.git
ln -s "$PWD/auxfirst-skill-global/skills/auxfirst" ~/.claude/skills/auxfirst

Use .claude/skills/ inside a repository instead of ~/.claude/skills/ to share it with a team through git.

Claude.ai / Desktop — as an uploadable .skill file:

cd /path/to/skill-creator
python -m scripts.package_skill /path/to/auxfirst-skill-global/skills/auxfirst

Upload the resulting auxfirst.skill and press Save skill on the file card.

It fires on its own — no slash command. Ask something like "we're about to let the support bot issue refunds instead of drafting them, what do we need first?" and it takes over.

The loop it runs

1. LIST     the verbs. What can this agent actually do?
2. SCORE    each verb on five dimensions, 0-4.
3. BAND     take the highest single score. That is the heat band.
4. CAP      the band sets the maximum permitted autonomy mode.
5. OWE      the mode sets the required patterns and supervision primitives.
6. WRITE    the mandate, action by action.
7. EXCEPT   decide the behaviour for each failure mode.
8. EVALUATE against the ten heuristics before ship.
9. CHECK    the twenty-four questions before calling it production.

Steps 1-5 are mechanical, and the bundled script does them so nobody quietly averages. Steps 6-9 are where the judgement lives.

Band Control posture Max autonomy mode
LOW Auto-run, log everything Autonomous
LOW-MED Auto-run, sampled review, one-click undo Act-and-notify
MEDIUM Propose, then batch-approve Review-before-act
HIGH Named approver, per instance Approve-each-action
CRITICAL Human executes, agent prepares Human-only

What's in the box

skills/auxfirst/
├── SKILL.md                              the loop, the tables, the output contract
├── references/
│   ├── heat-ladder.md                    dimension anchors, escalators, de-escalators,
│   │                                     the six supervision primitives
│   ├── heuristics-and-patterns.md        the 10 AUX heuristics, the 6 AUX patterns
│   ├── production-check.md               the 24 questions, scoring bands, 22 artefacts
│   ├── owners-manual.md                  15-field manual, 5 forms of control, AGENTS.md evidence
│   ├── workflow-readiness.md             the 3-layer, 12-check operability test
│   ├── anti-patterns.md                  named failure modes and precise vocabulary
│   └── skill-map.md                      Mermaid map of the whole procedure
├── assets/                               blank mandate, owner's manual, autonomy map
├── scripts/heat.py                       deterministic scoring, no dependencies
└── evals/                                trigger eval set + output test prompts

Two things it will not tell you

It will not tell you an agent is safe. It tells you whether the organisation can see, approve and undo what the agent did. Those are different claims, and the second is the only one a document can support.

It will not give you a single score. Capability and supervision are reported separately and never combined, because the gap between them is the finding.

Development

# validate the skill against the skill-creator rules
python -m scripts.quick_validate skills/auxfirst   # run from the skill-creator directory

# score a batch of actions from a CSV
python skills/auxfirst/scripts/heat.py --file actions.csv --json

evals/ ships in the repo and is excluded from the packaged archive. The negatives in the trigger eval are deliberate near-misses — RAG quality, model selection, prompt engineering, cost tuning — because easy negatives make the score meaningless.

Provenance

Assembled from published auxfirst material. Each source is the canonical, longer treatment:

Component Source
Five dimensions, five bands, escalators The Action Heat Ladder
Six supervision primitives, the two scoring rules Agent Supervision Method · the Register
Mandate, five controls, nine exceptions, four rules, fifteen fields The Agent Owner's Manual
Context-file findings AGENTS.md for teams that don't write code
Ten heuristics The 10 AUX Heuristics
Six patterns, the autonomy-to-pattern mapping The Six AUX Patterns
Twenty-four questions, twenty-two artefacts The Agent Development Lifecycle
Three layers, twelve-point readiness check Agent Operability
Agent definition, AUX/AX stack The AUX Manifesto

New here, and on no single source page: the join between heat band and autonomy mode, the band thresholds for the six supervision primitives, the "Enforced by" column in the mandate template, the first-five ordering of the artefact list, and the numeric score-to-band mapping the script applies.

License

Dual-licensed, and the split is simple: if it executes, it is MIT. If it is read by a person or a model, it is CC BY 4.0.

What License Covers
Code MIT skills/auxfirst/scripts/**, tools/**
Content CC BY 4.0 SKILL.md, references/**, assets/**, this README

Use the code in commercial products with no obligation beyond the copyright notice. Reuse and adapt the frameworks freely — keep the attribution:

Based on the auxfirst skill by auxfirst agency, licensed under CC BY 4.0.


auxfirst designs relationships between users and agents. The agent is a user. The relationship is the product. Trust is the moat.

auxfirst.com

About

Score every action an AI agent can take, cap its autonomy at the hottest dimension, and ship the mandate, exceptions and owner's manual behind it. A Claude skill for agent governance. Free, CC BY 4.0.

Topics

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Contributors

Languages