DEUK AGENT FLOW Clear actions. Checked results. AI collaboration through tickets, skills, roles and persona Flow keeps scope · Skills guide actions · Persona connects

· 한국어 · Website

• Installing is the only barrier = nothing to memorize; the installed agent runs tickets, checks and handoffs on its own.
• Deuk Agent Flow = an AI coding workflow connecting tickets, skills, roles and persona.
• Goal = fewer unnecessary edits, missed checks and lost handoffs between conversations.
• Core = the same way of working and the same ticket, whether you use Claude, Codex or another client.

Get started

• Requirements = Node.js 20+, npm and a supported agent client.
• Verified = tested in real use on Claude Code and Codex.
• Caution = Antigravity sometimes does not follow through on instructions.

npm install -g deuk-agent-flow
deuk-agent-flow init

1. Run init from the project root → choose your agent clients.
2. Tell your agent “Create a ticket” → the agent writes scope, plan and completion criteria.
3. Review the plan, reply “approved” → the agent implements, verifies and reports.
4. When things drift = just invoke the two skills below.

• Next session = “Continue the current ticket”.
• Update = upgrade the global package, then rerun deuk-agent-flow init.

When things drift

Symptom	Invoke	What it does
Forgets rules or scope, wanders off	“Run rule-reload”	Rereads project rules and design docs, then resumes the task
Tone or persona slips	“Call persona-maid again”	Rereads persona guidance and restores the tone

• Both drifting = invoke both in a row.

Learn more

What is different · Why · Research

What is different

Approach	Scope retention	Completion judgment	Handoff between sessions
Direct chat instructions	Repeated by the user	Agent self-report	Lost at session end
One AGENTS.md / CLAUDE.md	Static; diluted as it grows	Agent self-report	Manual notes
Client planning mode	Within one session	Agent self-report	Lost at session end
Deuk Agent Flow	Fixed ticket Scope and APC	Close after evidence slots	Kept in ~/.deuk/tickets/

• Planning modes = Claude Code, Codex /plan, Cursor Plan Mode, Copilot Plan agent; pre-execution approval is shared.
• Difference = scope re-injected after approval, close refused without evidence, same ticket across clients.
• Combined use = keep the planning mode on while working on an AgentFlow ticket.

Why

Observed problem	Study	AgentFlow response
Rising share of rule violations and false completion reports	Tang et al. 2026 · 20,574 sessions	No close without criteria and evidence
91.5% of visible misalignments fixed by the user	Tang et al. 2026	Stage gates stop first
Most failures start as early epistemic errors	Zhao et al. 2026 · 1,794 runs	Confirm scope and design before execution
Experts 19% slower with AI, yet felt faster	METR RCT 2025 · 246 tasks	Human review narrowed to plan and evidence
Context files: lower success, 20%+ higher cost	Gloaguen et al. 2026	CLI supplies only stage-relevant rules
Information in the middle of long context underused	Liu et al. TACL 2024	Short state surface re-injected each turn
Accepting agent edits lowers code comprehension	Balepur et al. 2026 · 54 users	Human approves plan and change scope

• Own measurement = telemetry records ticket create/move/close counts.

How skills steer agent behavior

1. Clarify the request	2. Apply relevant skills	3. Follow the criteria	4. Check, correct, hand off
Capture the task in a ticket; define change boundaries and completion criteria.	Read relevant SKILL.md playbooks; apply pre-edit checks and execution criteria.	Use stage-specific roles and work within scope; curb guessing and unnecessary edits.	Compare actual results with completion criteria; fix failures, recheck and record.
“Change this function only” → name the target and protected areas	safe-refactor → small changes / generated-file-guard → edit the source	Design → implement → QA; doc-sync governs document updates	Failed check → correction → recheck → evidence-backed handoff

• Behavior steering = task-time instructions, change boundaries and verification feedback; no model retraining.
• On failure = correct and recheck before completion; Persona shapes expression, while Role supplies the perspective for each stage.

Measured effects in related research

Method and study	Reported comparison	Behavior carried into AgentFlow
SkillsBench v4 · 2026	Curated skills: 33.9% → 50.5% (+16.6 pp); 87 tasks, 18 model/harness configurations	Select relevant skills; use focused procedural instructions
MetaGPT · ICLR 2024	Adding executable feedback: HumanEval +4.2 pp; MBPP +5.4 pp; human revision cost metric 2.25 → 0.83	Role-specific artifacts; execute, verify and correct
Reflexion · 2023	Reported HumanEval pass@1: GPT-4 80% / Reflexion 91%; reflection and retry procedure included	Record failure feedback and use it in subsequent attempts
Bickmore & Picard · 2005	101 users, one-month exercise-support study: greater liking, trust and desire to continue with the relational agent	Consistent persona expression and conversational continuity
Klein, Moon & Picard · 2002	Empathic, active-listening agent vs ignore/vent controls: users kept using the frustrating system significantly longer	Empathic response and recovery after failure or correction
EmotionPrompt · Li et al. 2023	Emotional stimuli added to prompts: Instruction Induction +8.00% relative, BIG-Bench +115%; 106-person evaluation +10.9% on generative tasks	Emotionally framed instructions shape model output quality

• Research-supported direction = relevant skills + structured roles + executable verification and correction + consistent interaction.
• Task fit matters: SWE-Skills-Bench v2 found no significant effect for 92 of 98 skills; average pass ratio 76.6% → 76.9%.
• Persona evidence = trust and continued use with relational agents (Bickmore), continued use after frustration (Klein), output quality gains from emotional stimuli (EmotionPrompt).
• Figures belong to the cited experiments, not Deuk Agent Flow measurements.

Components

Element	Responsibility	Examples
Flow	Scope, approval, verification, handoff	Tickets, APC, state-aware CLI
Skill	Procedural criteria for recurring work	safe-refactor, generated-file-guard, doc-sync
Role	Stage-specific expertise	Planning, design, implementation, QA, closeout
Persona	Expression and situational responses	Optional persona-maid

• Persona effect = softens stiff robotic tone; responsiveness stays warm through long tasks.
• Refusals = no "that feature is unavailable" declarations or lectures, just a brief move-on; judgment unchanged.
• After correction = flinch → recover → fix at once; code, documents and evidence stay neutral.
• Storage = ~/.deuk/tickets/{workspace-uuid}/; workspace selected with CLI --workspace.
• VS Code panel = ticket inspection, workspace switching, handoff copying and skill controls.

Apply a skill

deuk-agent-flow skill list
deuk-agent-flow skill add --skill safe-refactor
deuk-agent-flow skill expose --platform codex --skill safe-refactor

• Optional persona = deuk-agent-flow skill add --skill persona-maid.
• Exposure platforms = claude, codex, cursor, copilot, antigravity.
• Removal = skill unexpose, then skill remove; use the CLI instead of deleting deployed files.

Documentation

Document	Purpose
Usage guide	Setup, updates, daily use
How it works	Flow, storage, roles, handoffs
Skills guide	Behavior steering, installation, exposure, persona
Architecture	State, rules, storage
Principles	Collaboration goals, evidence, change policy

• VS Code extension · License Apache-2.0 · Built by joygram with AI-assisted development.