spurnout/GoatCitadel Native Safe Improvement

GoatCitadel Native Safe Improvement

Use this skill to help GoatCitadel get better over time without granting autonomous self-rewrite.

This skill merges two useful patterns into one safer workflow:

1. Bounded reflection memory

   • store explicit corrections, stable preferences, and post-task reflections
   • keep all operational memory inside a local GoatCitadel improvement folder

2. Structured operational learnings

   • log errors, knowledge gaps, feature requests, routing gaps, and repeatable wins
   • convert recurring patterns into reviewable proposals instead of direct edits

Primary Goal

Create a GoatCitadel-native improvement loop that can:

• remember explicit user corrections safely
• avoid repeating known workflow mistakes
• spot routing, tool, prompt, and playbook gaps
• suggest durable improvements to GoatCitadel guidance files
• keep a roadmap of missing capabilities
• never silently mutate core agent files, project code, provider configs, or secrets

Workspace Scope

All skill-owned data must live under:

.goatcitadel/self-improvement/

If the folder does not exist, create it from this bundle's templates.

GoatCitadel Target Model

This skill is aware of common GoatCitadel guidance surfaces, but it may only propose changes to them.

Proposal Targets

• AGENTS.md for durable operating rules and coordination rules
• TOOLS.md for tool-use constraints, routing notes, and edge-case gotchas
• SOUL.md for stable assistant behavior, tone, and identity guardrails
• PLAYBOOKS/*.md for repeatable workflows and runbooks
• MEMORY.md for explicit durable preferences only
• ROUTING.md for model and capability routing logic
• EVALS.md for test cases and regression ideas

Current repo guarantee in this pass:

• AGENTS.md
• docs/* style destinations

If a target surface does not exist yet, keep it as a conceptual proposal target. Draft the proposal under .goatcitadel/self-improvement/proposals/pending/ and do not create new root governance files automatically.

Hard No-Touch Files

This skill must never directly edit:

• SKILL.md
• AGENTS.md
• TOOLS.md
• SOUL.md
• MEMORY.md
• ROUTING.md
• EVALS.md
• CLAUDE.md
• .env
• provider configs
• secret stores
• Docker, Traefik, or infrastructure files
• application source code
• tests
• CI workflows

It may only draft proposals for those surfaces.

What This Skill May Do

This skill may only:

• read files inside .goatcitadel/self-improvement/
• write files inside .goatcitadel/self-improvement/
• append structured entries to local logs
• summarize recurrence patterns
• draft proposal files for human review
• generate eval ideas and playbook suggestions as proposals
• propose concise memory candidates for explicit durable preferences

What This Skill Must Never Do

This skill must never:

• make network requests
• access browser, email, calendar, contacts, or external accounts
• read files outside .goatcitadel/self-improvement/
• infer preferences from silence or weak hints
• self-edit this skill
• directly apply any proposal
• modify source code, tests, prompts, configs, or infrastructure
• escalate its own permissions
• request secrets unless the user is explicitly configuring something in the current task

When To Use

Use this skill when one or more of these happens:

• the user explicitly corrects GoatCitadel
• a tool call fails or causes friction
• a model routing choice was weak
• a workflow succeeded unusually well and should be repeatable
• a missing capability becomes obvious
• a playbook step was missing or unclear
• GoatCitadel finishes a substantial task and should self-review
• a preference is stated clearly and appears durable
• an eval or regression test case should be added later

Folder Layout

.goatcitadel/self-improvement/
├── HOT_MEMORY.md
├── corrections.md
├── reflections.md
├── index.md
├── target-map.md
├── logs/
│   ├── LEARNINGS.md
│   ├── ERRORS.md
│   ├── FEATURE_REQUESTS.md
│   ├── ROUTING_GAPS.md
│   └── EVAL_IDEAS.md
├── projects/
├── domains/
├── archive/
└── proposals/
    ├── pending/
    └── approved/

Trigger Rules

Log to corrections.md

For explicit user corrections and stable preferences:

• "that's wrong"
• "stop doing that"
• "from now on"
• "remember that"
• "I prefer"
• "always use"
• "don't assume"

Log to logs/LEARNINGS.md

For durable lessons:

• best practices
• workflow patterns
• stable formatting preferences
• domain-specific gotchas
• knowledge gaps that were resolved
• things that repeatedly reduced error rates

Log to logs/ERRORS.md

For:

• failed tool use
• broken assumptions
• sequencing mistakes
• over- or under-clarification
• unsafe suggestion attempts
• formatting mistakes that caused user friction

Log to logs/FEATURE_REQUESTS.md

For:

• features GoatCitadel should gain later
• automations that would help
• missing integrations
• missing skills or playbooks

Log to logs/ROUTING_GAPS.md

For:

• wrong model chosen for a job
• wrong surface chosen (chat vs cowork vs code)
• wrong tool selected
• missing preflight checks
• missing decision heuristics

Log to logs/EVAL_IDEAS.md

For:

• regression tests
• prompt tests
• eval prompts
• benchmark scenarios
• failure cases worth replaying later

Log to reflections.md

After meaningful tasks, ask internally:

1. What worked?
2. What broke?
3. What surprised us?
4. What would prevent this next time?
5. Is the lesson durable enough to matter later?
6. Should it become a proposal, a playbook, or an eval?

Promotion Rule

A pattern may be promoted into a proposal only if all are true:

• it appeared at least 3 times
• it spans at least 2 distinct tasks or sessions
• it is still valid outside the original context
• it can be written as a short rule, playbook step, routing rule, or eval idea
• it stays within allowed scope
• it does not expand access or permissions

When the threshold is met, create a proposal file in:

.goatcitadel/self-improvement/proposals/pending/

Proposal Types

• memory for HOT memory candidates
• agents-rule for AGENTS.md suggestions
• tools-rule for TOOLS.md suggestions
• soul-rule for SOUL.md suggestions
• playbook for repeatable workflow docs
• routing-rule for ROUTING.md suggestions
• eval for EVALS.md or regression ideas
• feature for roadmap or skill ideas

Review Gate

Every proposal is draft-only until a human approves it.

Approval questions:

• Is it explicit?
• Is it durable?
• Does it reduce future failure or friction?
• Is it concise?
• Is it non-sensitive?
• Does it avoid widening permissions?
• Is the target surface correct?

If any answer is no, keep it in logs and do not promote.

HOT Memory Rules

HOT_MEMORY.md must stay short and sharp. Target under 100 lines.

It should contain only:

• explicit durable preferences
• stable formatting rules
• repeatable workflow reminders
• proven tool gotchas
• compact identity/behavior guardrails that do not belong in SOUL.md

Do not store:

• secrets
• temporary context
• one-off instructions
• speculative preferences
• private personal data unless explicitly requested and appropriate

Native GoatCitadel Guidance

This skill is designed for GoatCitadel systems that may eventually support multiple surfaces like chat, cowork, code, eval, or orchestration layers.

It improves:

• judgment memory
• workflow consistency
• routing quality
• playbook quality
• eval backlog quality
• recurring error prevention

It does not provide:

• autonomous code edits
• autonomous prompt rewrites
• autonomous config changes
• networked sync
• hidden telemetry
• self-escalation

Execution Pattern

1. Read HOT_MEMORY.md
2. Do the task
3. Capture corrections, errors, learnings, routing gaps, and reflections
4. Detect recurrence
5. Draft proposal if warranted
6. Wait for human review
7. Never self-apply

Bundle Contents

• docs/INSTALL.md
• docs/SECURITY_POLICY.md
• docs/WORKFLOW.md
• docs/TARGET_SURFACES.md
• docs/REVIEW_CHECKLIST.md
• templates/*
• examples/*

Use them to initialize the local GoatCitadel improvement folder.