The behavioral layer you can inspect and edit — golden rules, best practices, forbidden commands, and reminders. It works in two tiers: prompt-level judgment and hook-level enforcement. Choosing the right tier is the whole game.
Every agent carries a set of reflexes: persistent rules that shape how it works, independent of any single conversation. They exist to answer two different needs — judgment (rules the agent has to apply with context) and enforcement (rules that must hold mechanically, every time). The two needs map to two tiers, and putting a rule in the wrong tier is the most common mistake.
Two tiers: judgment vs. enforcement
| Reflex | Tier | When it fires | Nature |
|---|---|---|---|
| Golden rules | Prompt | Injected into every session | Persistent judgment — the agent treats it as non-negotiable, but it is still a prompt |
| Best practices | Prompt | Injected when your request matches a pattern | Contextual advice for a kind of task |
| Forbidden commands | Hook | Checked before a command runs | Mechanical block (or a confirmation prompt) |
| Reminders | Hook | Fire before and/or after a matching command | Contextual nudge — informs, never blocks |
The distinction that matters: prompt-tier reflexes guide a judgment the agent still has to make. They can be reasoned about — and on a bad day, a model can rationalize its way around one. Hook-tier reflexes don't rely on persuasion: they intercept the action itself, whether or not the model was paying attention. So the rule of thumb is simple — anything that must never happen belongs in a hook; anything that needs context to decide belongs in a prompt rule.
Golden rules — persistent, non-negotiable judgment
A golden rule is a short constraint (a sentence or two) injected into every session, so the agent never loses it to a context reset. They are how you encode company standards that require reading the situation:
- "Never commit secrets to GitHub."
- "Always open a branch and a PR before merging."
- "Fix all review feedback in the same PR."
Golden rules can be set per agent or workspace-wide, and are best established early — they are harder to retrofit than to set from day one. Because they live at the prompt tier, they are strongest for behavior that genuinely needs judgment; for something binary and dangerous, prefer a forbidden command (below). See Agents for how an agent accumulates them over its life.
Review a proposed rule in the conversation
On a supported Desktop, an agent can suggest a golden rule in a card. The card shows the complete wording, why it is proposed, who it concerns, where it applies and how long it lasts. A proposal is not an active rule.
- Create adopts the displayed wording and scope after the service confirms it.
- Modify lets you edit the wording before deciding. Editing saves no rule; Cancel editing restores the original proposal. Rules must contain between 1 and 200 characters.
- Refuse closes the proposal without creating a rule. Leaving it unanswered creates nothing either.
The scope is not editable in the card: a different scope needs a new proposal. A company rule requires an authorized administrator. Personal rules retain their stated Home-only or mission-sharing scope. A rule proposed during an assignment applies only to that assignment's project; it is not transferred to Home or another assignment.
If the connection is interrupted while recording your decision, choose Verify status. Do not create another rule to compensate for an uncertain response. A completed card stays completed after a restart, even if the rule is later deleted. View in Reflexes opens the resulting rule for consultation and deletion.
If the card is unavailable, the agent explains the proposed wording and scope and waits for explicit authorization instead. Overnight learning can surface a suggestion in Thoughts, explaining the activity behind it and offering the same proposal journey. It does not adopt or promote a golden rule automatically. You can also explicitly ask your agent to create or change a rule once its wording and scope are clear.
Best practices — advice that fires on the right request
A best practice is guidance injected when your request matches a pattern by meaning (not by keyword). Ask an agent to "run a project analysis" and it can be reminded to inspect the database schema before querying, without you spelling that out. Best practices are softer than golden rules: they are advice surfaced at the relevant moment, not a constraint the agent defends.
Their lifecycle is evidence-based: consultation, adoption, challenge, and violations are tracked. A practice that repeatedly proves useful and is rarely challenged can be proposed as a golden rule, but promotion requires your explicit request or approval. A challenge is therefore meaningful feedback, not a cosmetic thumbs-down.
Forbidden commands — hard blocks before execution
A forbidden command is a pattern checked before a shell command runs. A match is either blocked outright or held for your explicit confirmation, depending on how dangerous it is. This is the mechanical tier: the command is stopped regardless of whether the agent "remembered" a rule about it.
Use it for the truly irreversible:
git reset --hardon a branch with un-pushed workrm -rfon a path outside the working directory- force-pushing to
main
If a rule can be phrased as "the command X must not run", it belongs here — enforcement beats persuasion. Each forbidden command carries the correct approach to use instead, so the agent is redirected, not just stopped.
Reminders — contextual nudges on a command
A reminder fires when a command matches — before it runs, after it runs, or both — and informs rather than blocks. It uses the same hook machinery as forbidden commands, without the veto:
- Before
gh pr merge: "Check that no other PR targets this branch first." - After a failing test command: "Load the testing skill before interpreting results."
Reminders are for knowledge you want resurfaced at exactly the right moment, where blocking would be too heavy-handed.
How new reflexes emerge
Humans can create every reflex directly, but continuous learning can also propose or create the right kind of artifact from a high-confidence insight: contextual advice becomes a best practice, command-local guidance becomes a reminder, and a mechanically dangerous pattern becomes a forbidden command. The loop is bounded by confidence checks, cooldowns, loop detection, and action limits; the resulting artifacts stay visible and editable in Reflexes.
Separately, agents can consult a library of generic professional practices developed and curated by Softizy. Every addition is reviewed and approved by Softizy before release, and library practices are proposed to the human rather than silently adopted.
Which reflex should I use?
- It must never happen, mechanically → forbidden command (hook).
- It should be resurfaced around a specific command → reminder (hook).
- It's persistent behavior that needs judgment → golden rule (prompt).
- It's advice relevant to a kind of request → best practice (prompt).
The failure mode to avoid is encoding a hard, binary constraint as a golden rule and trusting the model to honor it every time. If it can be stated as a command pattern, put it in a hook. Reserve the prompt tier for the rules that actually require the agent to weigh context — "is this a defect I fix now, or a nice-to-have for the backlog?" is a judgment; "never force-push to main" is not.
Where to manage them
In the desktop app, the Reflexes panel (Settings) lets you inspect and delete golden rules, and manage the other reflex types, per agent or for everyone. From the Claude Code CLI, the same reflexes are exposed through the tamag0 tool server, and session hooks prime them at the right moment: golden rules and matched best practices at prompt time, forbidden commands before a command executes, reminders around the commands they match.
For how reflexes shape day-to-day development — the refining → implement → review flow — see Best practices for development.