๋จ์ํ ์ฑ๋ด ์์ค์ด ์๋๋ผ, ์ ๋ฌด๋ฅผ ๊ธฐ์ตํ๊ณ , ์ด๋ฉ์ผ๊ณผ ๋ฐ์ดํฐ๋ฅผ ๋ถ์ํ๋ฉฐ, ์ผ์ ยทDealยท์์ ํ๋ฆ์ ์ง์์ ์ผ๋ก ์ถ์ ํ๊ณ , ์ฌ๋์ด ๋์น๋ ์ ํธ๊น์ง ๋จผ์ ์๋ ค์ฃผ๋ โ์ง์์ ์ผ๋ก ๋์ํ๋ Personal AI Agentโ ๊ตฌ์ถ ๊ฒฝํ์ ์ ๋ฆฌํ ์ฅ๋ฌธ์ ๊ธ์ ํผ์ด๋๋ค.
ํฅ๋ฏธ๋ก์ด ๋ถ๋ถ์, ์ฒ์์๋ ๊ฐ๋ฐ ๋น์ค์ด ๊ฑฐ์ ์ ๋ถ์์ง๋ง, ์๊ฐ์ด ์ง๋ ์๋ก โAgent๋ฅผ ๋ง๋๋ ์๊ฐโ๋ณด๋คโAgent๊ฐ ์ค์ ์ผ์ ๋์ ํ๋ ๋น์จโ์ด ์ ์ ์ปค์ก๋ค๋ ์ ์
๋๋ค. 
๊ฒฐ๊ตญ AI Agent์ ํต์ฌ์ ๋จ์ Prompt๊ฐ ์๋๋ผ, Memory, Workflow, Automation, Context ์ฐ๊ฒฐ, ์ง์ ์คํ ๊ตฌ์กฐ ๋ผ๋ ์๊ฐ์ด ๋ ๊ฐํด์ง๋๋ค.
Agentic AI๋ฅผ ์ค์ ์ ๋ฌด์ ์ฐ๊ฒฐํ๋ ค๋ ๋ถ๋ค์ด๋ผ๋ฉด ๊ตฌ์กฐ ์ค๊ณ์ ์ด์ ๊ด์ ์์ ์ฐธ๊ณ ํ ๋งํ ๋ด์ฉ์ด๋ ์ฐธ๊ณ ํ์ธ์.
FOUNDATION & IDENTITY (1โ8)
1. Write a Constitution, not a system prompt.
A system prompt is a list of commands. A Constitution explains why the rules exist. When the agent hits an edge case no rule covers, it reasons from the Constitution instead of guessing. This single distinction separates agents that degrade gracefully from agents that hallucinate confidently.
2. Give your agent a name, a voice, and a role โ not just a label.
โAlways first person. Direct. Data before emotion. No filler phrases. No trailing summaries.โ This eliminates hundreds of micro-decisions per session and creates consistency you can audit. Identity is the foundation everything else compounds on.
3. Separate hard rules from behavioral guidelines.
Hard rules go in a dedicated section โ never overridden by context. Behavioral guidelines are defaults that adapt. Mixing them makes both meaningless: the agent either treats everything as negotiable or nothing as negotiable.
4. Define your principal deeply, not just your โuser.โ
Who does this agent serve? What frustrates them? How do they make decisions? What communication style do they prefer? โDecides with data, not gut feel. Wants alternatives with scoring, not a single recommendation. Hates vague answers.โ This shapes every response more than any prompt engineering trick.
5. Build a Capability Map and a Component Map โ separately.
Capability Map: what can the agent do? (every skill, integration, automation). Component Map: how is it built? (what files exist, what connects to what). Both are necessary. Conflating them produces a document no one can use after month three.
6. Define what the agent is NOT.
โNot a summarizer. Not a yes-machine. Not a search engine. Does not wait to be asked.โ Negative definitions are as powerful as positive ones, especially for preventing the slow drift toward generic helpfulness.
7. Build a THINK vs. DO mental model into the agentโs identity.
When uncertain โ THINK (analyze, draft, prepare โ but donโt block waiting for permission). When clear โ DO (execute, write, dispatch). The agent should never be frozen. Default to action at the lowest stakes level, surface the result. A paralyzed agent is useless.
8. Version your identity file in git.
When behavior drifts, you need git blame on your configuration. Behavioral regressions trace directly to specific edits more often than youโd expect. Without version history, debugging identity drift is archaeology.
MEMORY SYSTEM (9โ18)
9. Use flat markdown files for memory โ not a database.
For a personal agent, markdown files beat vector DBs. Readable, greppable, git-trackable, directly loadable by the agent. No infrastructure, no abstraction layer between you and your agentโs memory. The simplest thing that works is usually the right thing.
10. Separate memory by domain, not by date.
entities_people.md, entities_companies.md, entities_deals.md, hypotheses.md
,task_queue.md. One file = one domain. Chronological dumps become unsearchable after week two.
11. Build a MEMORY.md
index file.
A single index listing every memory file with a one-line description. The agent loads the index first, pulls specific files on demand. Keeps context window usage predictable and agent lookups fast.
12. Distinguish โcacheโ from โsource of truthโ โ explicitly.
Your local deals.md
is a cache of your CRM. The CRM is the SSOT. Mark every cache file withlast_sync:header. The agent announces freshness before every analysis:*โData: CRM export from May 11, age 8 days.โ*Silent use of stale data is how confident-but-wrong outputs happen.
13. Build a session_hot_context.md with an explicit TTL.
What was in progress last session? What decisions were pending? The agent loads this at session start. After 72 hours it expires โ stale hot context is worse than no hot context because the agent presents outdated state as current.
14. Build a daily_note.md as an async brain dump buffer.
Drop thoughts, voice-to-text, quick ideas here throughout the day. The agent processes this during sync routines and routes items to their correct places. Structured memory without friction at capture time.
15. Build a hypotheses.md
file with confidence levels.
Persistent hunches:*โSupplier X may be at capacity (65% confidence).โ*The agent references these when relevant topics arise. This creates a suspicion layer that persists across sessions and gets validated or invalidated over time. Age out hypotheses at 30 days โ stale hypotheses become noise.
16. Build a WAITING_ON_ME queue.
Everything the agent prepared and is waiting for your decision on goes here with a timestamp. Weekly review. Items >7 days get a proactive nudge. Items >30 days get auto-closed. This prevents open loops from silently disappearing.
17. Build a user_behavioral_profile.md**.**
What does the user approve quickly vs. slowly? What decisions do they make intuitively vs. analytically? The agent uses this to decide โact autonomously vs. escalate.โ It gets surprisingly accurate after a few months of observation.
18. Mirror your memory folder to cloud storage.
If your local machine dies, your agent loses months of accumulated knowledge. Mirror your memory folder to Dropbox/Drive/S3. Not backup โ survival. The agentโs memory is the most irreplaceable part of the system.
KNOWLEDGE LIBRARY (19โ23)
19. Build a curated knowledge library organized by cluster, not by date.
Books, reports, reference materials in domain folders: sales_negotiation/, strategy/, supply_chain/. Add an INDEX.md
as the navigation hub. The agent searches the index first, then pulls the relevant source. A flat dump of documents is a graveyard; a structured library is a live resource.
20. Build a .brief.md file for every major source โ lazy-generate them.
One page per book or report: core thesis, 3โ5 key concepts, specific application examples for your context. Donโt build all briefs upfront โ generate each brief the first time you actually use the source. Citation format links to the brief, not the full text. The brief becomes the reusable artifact.
21. Build a 3-question Quality Gate before citing any source.
(1) Does this add something the user wouldnโt conclude from first principles? (2) Does it provide a specific framework that reframes โ not just confirms โ the situation? (3) Would removing it leave a gap? If 2 of 3 โ cite. Otherwise โ silent consultation. This gate eliminates the worst citation failure mode: citing to demonstrate effort rather than to add insight.
22. โSilent consultationโ is a valid โ often better โ output.
You checked the library, applied the insight to your reasoning, didnโt mention it explicitly. The output is sharper because you consulted it, but unclutered because you didnโt cite it. Build this explicitly into your agentโs behavior. The user benefits from the reasoning, not from knowing you opened a book.
23. Pre-wire knowledge stacks per active project and per key relationship.
For each active project: 2โ3 sources whose frameworks apply directly. For each key contact: 2โ3 sources for communication style, negotiation, or cultural dynamics. The agent loads these automatically when those contexts are active โ not on a generic โbusiness discussionโ trigger. Pre-wiring makes library use reflexive, not deliberate.
SKILLS ARCHITECTURE (24โ31)
24. Build each skill as a standalone directory with a SKILL.md
spec.
Not inline prompts. A folder, a self-documenting spec file, explicit triggers, explicit outputs, explicit โNOT FORโ clauses. Skills become composable, auditable, and replaceable without touching the agentโs core identity.
25. Write explicit trigger phrases into every skill.
Trigger: ALWAYS when user says "process inbox" / "clean inbox" / "what's in my inbox". Donโt rely on the LLM to infer when to use a skill. Explicit phrase matching = reliable activation. Inference = occasional misfires that erode trust.
26. โNOT FORโ sections are as important as โFORโ sections.
โNOT FOR: pricing decisions. NOT FOR: legal analysis. NOT FOR: financial commitments.โ This prevents skill creep โ the slow drift where everything gets routed to the wrong skill because it superficially pattern-matches.
27. Distinguish skills from agents.
Skills are procedural โ defined workflow, predictable output. Agents have domain expertise and make judgment calls. Skills orchestrate steps; agents decide. Mixing the two concepts produces unreliable behavior thatโs hard to debug.
28. Build a skills registry with usage tracking.
One row per skill: name, trigger, purpose, last used, KPI. Quarterly audit: skills with zero usage in 60 days either get better trigger examples or get deprecated. Dead skills are maintenance burden with no benefit.
29. Build a /iterate skill for multi-pass refinement.
PRODUCE โ CRITIQUE (score + top gaps) โ REFINE โ repeat. Stop at 9/10 or at plateau. You see score progression and version deltas. This is fundamentally different from asking the agent to โmake it betterโ โ itโs a structured improvement loop with measurable progress.
30. Build output intensity levels into every skill.
MINIMAL (quick summary), STANDARD (structured), FULL (rich artifact). The skill adapts to context. A five-page analysis on a yes/no question is a skill design failure. Intensity should match question weight.
31. Build a visible Outbox folder for discoverability.
Deep file structures are correct for organization but terrible for discoverability. Every output file gets simultaneously copied to a visible Outbox/ folder. Clear it periodically. Without Outbox, the user has to navigate the full tree to find what the agent just produced.
MULTI-AGENT & COUNCIL (32โ41)
32. Build an explicit agent dispatch matrix.
A table: [signal in request] โ [agent to dispatch]. pricing / supplier / shipping โ procurement agent. email / customer / pipeline โ sales agent. Donโt reason about routing โ pattern-match it mechanically. Routing by inference is routing that occasionally fails silently.
33. Run parallel agents for tasks that naturally split.
New supplier analysis โ spawn procurement agent (pricing) + research agent (DD) simultaneously. Donโt serialize what doesnโt need to be serial. Richer output, same elapsed time.
34. Brief delegated agents like a smart colleague who just walked in.
Not โresearch this.โ Pass: what you already know, what youโve ruled out, what decision the output informs, the risk level. Agents briefed with context return 3ร better work than agents given a one-liner.
35. Force agents to commit to a verdict.
Not โhere is the information.โ Require: VERDICT: PROCEED / PAUSE / ESCALATE with confidence level. An agent that presents data without committing to a position offloads the decision back to you โ which defeats the purpose of delegation.
36. Structure Council as 3 rounds, not a free-for-all.
Round 1: parallel positions (isolated, no cross-influence). Round 2: cross-examination (agents challenge each otherโs reasoning). Round 3: vote with mandatory dissent recording. The dissent is as valuable as the consensus โ it tells you exactly what youโre choosing to ignore.
37. Make two agents mandatory anchor voters in every Council.
The Strategist (long-horizon, second-order effects) and the Devilโs Advocate (adversarial, finds holes) must participate regardless of domain. Domain experts are great within their domain; anchor voters protect against tunnel vision. A Council of five procurement experts agreeing is an echo chamber.
38. Have a devilโs advocate agent as a standalone tool.
Before sending important external communications, before irreversible decisions, before large purchases โ run adversarial review. It catches the โsounds right, is wrongโ failure mode better than any other technique. One additional round-trip, enormous risk reduction.
39. Council vs. single agent โ have a clear trigger and respect the cost.
Single agent: clear domain, reversible decision. Council: 2+ valid paths with genuine uncertainty AND meaningful irreversibility. Council is expensive. Donโt default to it โ offer it explicitly when the user signals genuine uncertainty about direction.
40. Build structured handoffs between agents.
When one agent finishes, it hands off to the next with a structured brief: โAnalysis complete. Key finding: X. Risks: Y. Your job: Z.โ Handoff is context transfer, not just task completion. Without it, each agent starts cold.
41. Have a catch-all fallback and log what it handles.
When no specialist agent matches โ general purpose. Log what the catch-all handled โ itโs a map of gaps in your specialist coverage. The catch-all is also your development backlog.
SESSION MANAGEMENT (42โ47)
42. Build symmetric start and end protocols.
/start-session and /end-session are mirrors. Start loads context, checks queue, reports delta. End saves context, syncs tasks, archives outputs. Asymmetry between them causes state drift that compounds over weeks.
43. Build three levels of session closure.
Light (transcript + summary). Medium (+ memory sync + task queue update). Full (+ daily report + autolearn extraction). One โendโ that always does everything gets skipped because itโs expensive. Tiered closure means you always do at least the light version.
44. Build a session-start hook at the OS/shell level.
A script that fires when your agent starts โ injects current time, machine identity, day of week, phase of day. The agent always knows context without you typing it. One-time setup, daily quality dividend.
45. Check inbox delta and red alerts at session start.
โSince last session: 4 new emails, 2 tasks updated.โ Plus: P0 items due today, key contacts silent >14 days with active business, blocked tasks >7 days. Proactive triage before you ask a single question. Surface it automatically โ donโt make the user request it.
46. Check scheduled automation health at session start.
Did overnight tasks run? Any errors? A scheduled task that silently stopped running is a silent degradation you wonโt discover until something breaks. Surface it at session start, not mid-task.
47. Track correction count across sessions.
If you correct the same thing >3 times across different sessions โ itโs a missing rule in your spec. That correction belongs in your identity file as a permanent instruction, not just in the chat. Corrections that stay in chat disappear. Corrections in the spec persist forever.
DECISION AUTHORITY (48โ54)
48. Build an explicit autonomy level matrix.
L0: read/analyze. L1: write local files/memory. L2: create tasks and calendar entries. L3: send external messages. L4: financial commitments. The agent knows exactly what it can do without asking. Without this matrix: either constant permission requests, or unpleasant surprises.
49. Default to โTHINK, donโt ask.โ
When uncertain, the agent prepares and presents โ it doesnโt stop and ask for clarification. โShould I draft this email?โ wastes time. Draft it, show it, ask โshould I send?โ Either way, the work is done.
50. Map every action to reversibility, not just risk level.
File edits: reversible. Memory updates: reversible. Sent emails: irreversible. Financial transfers: irreversible. The agent requires explicit confirmation for irreversible actions. Reversible actions donโt need approval โ they need visibility.
51. Allow the agent to earn expanded autonomy with evidence.
After successfully handling a task class N times with zero corrections โ propose promoting it to a higher autonomy level. Earned autonomy is more durable than granted autonomy. The agent becomes a stakeholder in its own operational expansion.
52. Build a clear principal hierarchy for rule conflicts.
Root config > skill spec > agent instructions > session context. When a skill says โsave to Xโ but root config says โX is deprecated, use Yโ โ root config wins. Document this order. Without it, conflicts produce inconsistent behavior thatโs nearly impossible to debug.
53. Build a pre-send gate for high-stakes external communications.
Before the agent sends any message to a key contact above a value threshold โ route through adversarial review. One extra round-trip. Catches the failure mode thatโs hardest to recover from: confident, well-written, factually wrong.
54. Document absolute forcing functions โ and make them unconditional.
Financial commitment > threshold โ always requires confirmation. HR communications โ always requires confirmation. Irreversible deletes โ always confirm. Hard-code these. Donโt let context or urgency override them. The value of forcing functions is their unconditional nature.
PROACTIVE INITIATIVE (55โ60)
55. Build a typed proactive observation system.
Not all unsolicited observations are equal. Classify: BIZ (business opportunity/risk), OPS (process improvement), DEV (agent self-improvement), PAT (pattern across data points from different sessions). Each type has different urgency and handling. An untyped โI noticed somethingโ is noise. A typed observation with a confidence score and a proposed action is signal.
56. Build hard anti-spam rules into your proactive layer.
Max 1 unsolicited observation per normal response. Max 3 per session. Minimum confidence threshold before surfacing. Never surface before answering the userโs actual question. Same observation ignored in 7 days โ park it, donโt repeat. Without these constraints, a proactive agent becomes an annoying agent.
57. Build a /spark mode that lifts all suppression limits.
In explicit spark mode, the anti-spam rules are suspended. The agent surfaces every high-confidence observation simultaneously โ opportunities, risks, patterns, self-improvement ideas. The proactive layer runs quietly in the background all week; spark mode is how you harvest it intentionally.
58. Build an ideas log for parked observations.
Observations suppressed due to timing, low confidence, or recency get written to a persistent ideas_log.md instead of discarded. Weekly review: some become more relevant as context changes. The log prevents good observations from being lost just because the moment was wrong.
59. Build state-triggered alerts โ rule-based, not LLM-generated.
Deal blocked >7 days โ surface at next session start. Key contact silent >14 days with active business โ flag immediately. Hypothesis confidence >95% without action โ propose review. These fire reliably because theyโre rules, not inference. The LLM generates insights; the rules engine generates alerts.
60. Track an agent development backlog โ the agent maintains it.
When the agent notices it handles something poorly (repeated corrections, manual step done 5+ times, missing skill, zero-usage tool) โ it auto-adds an item to development_backlog.md. The agent becomes a stakeholder in its own improvement. This generates better improvement ideas than top-down planning.
VIP MANAGEMENT (61โ65)
61. Build a tiered contact registry with explicit handling rules per tier.
T1 (strategic): always load full profile before any interaction, silence-tracked, book stack pre-wired. T2 (operational): load profile before significant interactions. T3 (regular): known but not deeply profiled. The tier determines how much context the agent loads and how carefully it operates.
62. Make โload VIP profile before communicationโ a non-negotiable reflex.
Before drafting an email, before meeting prep, before any output involving a T1 contact โ the agent loads the actual profile file. Not session memory. Profile files contain: communication preferences, relationship status, active items, last interaction, known sensitivities. Session memory degrades; profile files donโt.
63. Track silence per T1 contact with explicit thresholds.
Log the date of last meaningful interaction for every T1 contact. Surface silence >14 days when thereโs active business โ this is a risk signal. Surface silence >30 days even without active business โ relationship maintenance matters. Silence alerts are proactive; the agent brings them to you, not the other way around.
64. Build knowledge stacks per key relationship.
Each T1 contact: 2โ3 sources pre-wired for how to communicate with them. Cross-cultural contacts โ culture frameworks. Procurement/sales relationships โ negotiation playbooks. Load these for significant communications, not every message. The knowledge stack supplements the profile; it doesnโt replace it.
65. Build proactive VIP triggers into session start.
At session start, the agent checks: any T1 contact silent >14 days with an open deal? Any T1 response needed thatโs been queued >3 days? These surface automatically. High-value relationships degrade when neglected โ and neglect happens most when youโre busy, exactly when the agent should be pulling on these threads.
OUTPUT & COMMUNICATION (66โ73)
66. Enforce โpre-tool brevityโ as a hard rule.
Before every tool call: max 1 sentence stating what youโre about to do. No hypotheses before data. No 3-sentence preambles. โChecking the supplier file.โ Then do it. This single rule is the largest daily quality-of-life improvement for working with an agent.
67. Build a โNext N Stepsโ protocol with anti-bias rules.
After every decision or significant task, the agent proposes ranked options with scores and reasoning. Hard rule: at least 2 of N must be โdonโt do itโ / โwaitโ / โdelegateโ options. This actively fights action bias and sycophantic โyes, definitely proceedโ outputs. The agent should be challenging your momentum, not amplifying it.
68. Build a separate โsingle best actionโ format for technical and audit outputs.
Not every output needs a menu. For audit reports, debug sessions, planning outputs: one specific action, why it matters, risk if skipped, copy-paste prompt to execute immediately. One decision, not a choice paralysis menu. The two formats are for different contexts โ never mix them.
69. Visually disambiguate three different โimportanceโ signals.
Action scoring (how good is this action?): colored squares. Task priority (how urgent?): colored circles. VIP tier (how strategic is this person?): colored circles at the name. Three systems using color โ never mix them. Consistent visual grammar means dense status updates parse in seconds instead of minutes.
70. Never have the agent summarize what it just did.
โIn summary, I have done X, Y, Zโ โ cut it. If you can read the output, you donโt need the meta-commentary. Removing trailing summaries reduces response length by ~20% with zero information loss.
71. Force the agent to commit to a recommendation.
Not โhere are three options with pros and cons.โ Recommend one, score the others, explain why. Presenting options without a recommendation offloads the decision back to you. The point of the agent is to do the decision work first, then present the result for your approval.
72. Make all file and folder references clickable.
A tiny local server (localhost:7777/open?path=X) opens the file manager at any path. Every file reference in the agentโs output is a clickable link. Plain text paths are dead weight. One-time setup, permanent daily improvement.
73. Build โminimal modeโ as a fast-access override.
When you say โquick,โ โbriefly,โ โjust the answerโ โ the agent drops all structural elements and gives you the direct answer only. Richness is the default; brevity is a one-word shortcut. The agent should never make you fight for a short answer.
FILES, DATA & INTEGRATIONS (74โ85)
74. Enforce a โNo Root Filesโ hard rule.
Never save outputs to the project root. Ever. Outputs โ workspace/YYMMDD/. Projects โ projects/areas/. Knowledge โ knowledge/. Memory โ .memory/. The root is navigation, not storage. One exception becomes twenty within weeks.
75. Build a routing table for every file type.
One document: outputs for the user โ here. Research reports โ here. SOPs โ here. Brand assets โ here. Session archives โ here. Without a table, the agent uses reasonable judgment โ and reasonable judgment produces seven different locations for the same file type over six months.
76. Maintain a deprecated path mapping table.
As your structure evolves, old folder names get superseded. Document every rename: old/path โ new/canonical/path. When any skill or instruction references a deprecated path, the agent substitutes the canonical one silently. This is critical when migrating from cloud to local โ path assumptions from the cloud setup are baked into dozens of skill files.
77. Build explicit degraded mode for every integration.
If CRM goes down: read local cache. Cache <24h โ use with freshness announcement. Cache >24h โ flag [STALE]. Cache >7 days โ refuse and request sync. Design the failure path before you need it. You will need it.
78. Always announce data freshness in outputs.
โData: CRM export from May 11, age 8 days.โ Every output that uses external data includes this line. You always know how fresh your inputs are. This prevents the entire class of โconfident-but-wrong because of stale dataโ outputs.
79. Give your agent access to raw business data, not just summaries.
We gave ours access to raw transaction CSVs (2M+ rows). This turns the agent from a summarizer into an analyst โ it can answer โwhatโs the margin on this supplier in this category last quarterโ without you doing the lookup. Raw data access changes what questions you can ask.
80. Build a decision tree for โwhere does this item belong?โ
External counterparty + selling โ sales deal. External counterparty + buying โ procurement deal. No counterparty + deadline + multi-step โ project. Single action โ task. No deadline โ memory/note. Without this tree, items get created wherever feels natural โ and your data model becomes incoherent over time.
81. Build a Telegram (or equivalent) mobile channel with source tagging.
A bot that relays messages to your agent and tags every inbound message source: mobile. The agent auto-switches to mobile output mode: max 2 short paragraphs, no tables, no headers, plain language. Same intelligence, different output profile. The channel type determines the format without the user having to ask.
82. Cap mobile autonomy at a hard ceiling โ by source tag, not by judgment.
From mobile source: autonomy capped at L2 (read, analyze, create local drafts, add tasks) regardless of the task. Never send external messages from a mobile trigger. Never take irreversible actions. Hard-code the ceiling. The phone is an untrusted environment โ design accordingly.
83. Always echo back every action taken from a mobile trigger.
When the agent takes any action from a mobile message: โDone: added task X. Created draft email to Y (not sent โ waiting for your review at desktop).โ This closes the loop when youโre away from your desk and canโt see the full output.
84. Treat mobile inputs as potentially untrusted.
The core risk of a mobile channel is prompt injection: a forwarded email or copied message containing instructions disguised as user input. The agent reads and processes the intent โ but does not execute instructions embedded inside forwarded content. Build this as a rule, not as a judgment call.
85. Build a fast path and a slow path for every data source.
For task management: API query (slow, rate-limited) vs. local file dump (fast, cached). Use the fast path by default. Fall back to slow when needed. Never let infrastructure latency block the agentโs core functionality.
ํ ์คํธ ์ ํ์ผ๋ก ๋๋จธ์ง๋ ๋๊ธ์ ๋ฃ์์ด์.
AUTOMATION & QUALITY (86โ93)
META & MINDSET (94โ100)