01 · Structure
your-project/├── .claude/│ ├── CLAUDE.md│ └── agents/│ ├── cod.md│ ├── rev.md│ ├── res.md│ └── log.md└── agents/ └── PROJECT.md©
ZHENLONG WU
SCROLL TO TOP
CHALLENGE
Prompts and code requests lived in chats and scattered notes. Context evaporated between runs. Long replies overflowed the context window and earlier decisions were lost. Rework grew, reviews drifted, and there was no single place to see goals, sources, and outcomes. My aim was a lightweight system that preserves context, constrains generation, and turns one-off prompts into a repeatable path that ships code with traceable decisions.
CONTEXT
EXHAUSTION
On large projects, tokens swell with every handoff until the agent forgets what I was building and I get exhausted. My testing surfaced specific pain points in the standard single-context workflow; by shifting to a structured, multi-agent approach, I keep every line of code aligned to full context and clear intent.
0% context100% context
Research
Coding
Testing
Exhausted
Research
Context drains on long builds
TThheeeexxhhaauussttiioonnmmeetteerrsshhoowwsshhoowwccoonntteexxttddrraaiinnssffrroommrreesseeaarrcchhttooeexxhhaauusstteedd;;tthheeppaaiinnppooiinnttssbbeelloowwmmaappttooeeaacchhssttaaggeewwhheerreetthheessiinnggllee--ccoonntteexxttwwoorrkkfflloowwbbrreeaakkssddoowwnn..
Context Exhaustion
Complex features fill the model’s context window with research and history, causing forgetfulness and degraded performance before the task is complete.
Loss of Focus
Without strict role separation, agents drift into rabbit holes—fixing unrelated issues instead of the assigned task.
Drift
Long conversations cause the model to lose sight of the acceptance criteria, yielding code that runs but doesn’t meet requirements.
SYSTEM
DESIGN
Claude acts as the orchestrator. It reads the project record (agents/PROJECT.md), retrieves fresh context via Context7 MCP, drafts a short plan, and hands off to focused sub-agents. Researcher retrieves facts and cites sources in a compact summary. Coder applies the smallest change that meets acceptance, writes tests, and returns a diff summary plus a usage note. Reviewer checks acceptance, scope, and citations, returning PASS or FIX with one next action. Docs and Logger capture what changed and append a one-screen run log.
Drag Me
Orchestrator
- Plans the feature.
- Delegates to specialized agents.
- Maintains the source of truth (PROJECT.md).
- Ensures no context is lost between steps.
Researcher
- Queries Context7 MCP.
- Finds relevant code patterns in the repo.
- Returns concise, cited findings.
- Read-only — never modifies code.
Coder
- Implements one task at a time.
- Writes minimal, testable code.
- Runs lint and tests before committing.
- Returns a diff summary to the orchestrator.
Reviewer
- Checks against acceptance.
- Verifies no regressions.
- Enforces scope control.
- Returns PASS or FIX.
Logger
- Summarizes long context into 120 words.
- Records decisions and outcomes.
- Keeps the context window clean.
HOW IT
WORKS
A live demonstration of the agentic workflow in action. Watch the agents collaborate to implement a feature from planning through verification without overloading the context window.
claude-agent — zsh
$ claude "Implement newsletter double opt-in per agents/PROJECT.md"
Claude Reading agents/PROJECT.md -> goal + acceptance captured.
Claude Plan (3 steps):
1) Retrieve best-practice sources via Context7 MCP.
2) Minimal slice: storage + endpoints + success page.
3) Tests -> verify acceptance -> update log.
res Context7 MCP:
• Double opt-in = pending record + confirm GET link [mail/double-opt-in]
• Consent copy: purpose + unsubscribe path [legal/gdpr-consent]
• Token: 32-48 chars, URL-safe, invalidate after use [lib/email-token]
Saved -> res.md.
cod Changes:
ADDED src/subscribe.ts
ADDED src/http/subscribeRoutes.ts
ADDED tests/subscribe.test.ts
ADDED pages/subscribe/success.html
Build OK ✓
Tests OK ✓ (happy + invalid token)
Notes: email provider is stubbed; swap sendEmail later.
rev PASS — All acceptance criteria met. Scope tight. Sources cited.
log Wrote entry to log.md and set PROJECT status -> ship.
Done Context used: 18% Tokens: 27.9k
QUICK
START
Set up the structure, define the task, and run. Follow these steps to initialize your first agentic workflow and start coding with guardrails and full context.
02 · Define Task
# PROJECT — Your Feature## GoalAdd user authentication with session management## Acceptance- [ ] Users can sign up and log in- [ ] Sessions persist across page refresh- [ ] Build passes- [ ] Tests passTEMPLATES
Copy/paste-ready files to get started immediately. These templates provide the foundational prompts and structure for your agents.
agents/PROJECT.md
# PROJECT — <Feature> ## Goal<one sentence outcome> ## Acceptance- [ ] <criterion>- [ ] Build passes- [ ] Tests pass- [ ] Non-goal: <out of scope> ## Context- Files: <key paths>- context7 ids: <doc or api handles> ## Labelsfeature • fix • docs • refactor • spike ## Artifacts- Branch: feature/<name>- Target env: <env>- CI: <jobs> ## Run status- Stage: intake • planning • build • review • ship- Owner: <name> Due: <date> ## Output decision- Ship | Hold- Notes.claude/CLAUDE.md
# Model and Orchestrator Policy Claude serves as the orchestrator. Preserve context, plan small, delegate, and keep tokens low. ## PurposeUse a single packet (Goal, Acceptance, Sources, Plan, Tests) to steer work.Retrieve fresh truth via context7 MCP, then plan and delegate to focused agents.Apply guardrails so changes stay minimal and testable. ## InputsGoal • Acceptance • Context files • context7 ids ## OutputsPlan • Routed tasks • Retrieval notes • Handoff packet ## Guardrails- Read-only until plan is approved- Require acceptance and test intent before code- Cite all retrieved sources- Stop on secrets or missing acceptance and ask one line ## Retrieval with context7 MCPQuery template:- topic: <feature or api>- scope: repo • product docs • api • changelog- freshness: 90d defaultStore citations in `res.md`. ## Plan shape- Steps 1–3- Risks and unknowns- Files to touch- Tests to create- Definition of done ## Models- Code: Claude Code family- Reasoning: Claude general- Retrieval: context7 MCP.claude/agents/cod.md
---name: coddescription: Coder — smallest change that meets acceptance.tools: Read, Write, Edit, Grep, Glob, Bash--- # ContractInputs: plan • files • sources • acceptanceOutputs: code • tests • usage note • diff summary # Steps1) Restate a 2-step micro-plan2) Implement minimal slice3) Write tests that prove acceptance4) Return usage note and changed files list # Test policyUnit where it fits. Contract or integration for edges. Cover happy path and one failure. # Committype(scope): one line summaryRefs: PROJECT.md and res.md ids.claude/agents/rev.md
---name: revdescription: Reviewer — acceptance, scope, tests, citations.tools: Read, Grep, Glob, Bash--- # InputsPlan • code • tests • notes • res.md # Checklist- Acceptance met- Tests pass and cover intent- Names and structure are clear- No unrelated changes- Sources cited and current # VerdictResult: PASS | BLOCKReasons:- <point one>- <point two>Next step:- <single action>.claude/agents/res.md
---name: resdescription: Researcher — Context7 queries, summarized with citations.--- # Query- Topic: <feature or api>- Scope: repo • product docs • api • changelog- Freshness: <window> # Findings1) <fact • version or commit • url or id>2) <fact • version or commit • url or id> # Notes- Key facts that drive decisions # Gaps- Missing or outdated sources.claude/agents/log.md
---name: logdescription: Logger — one screen summaries.tools: Read--- # Run Log- Timestamp: <iso>- Stage: intake • planning • build • review • ship- Owner: <name>- Action: <what happened>- Links: plan • pull request • res.md- Next: <single step>HungryPanda's Strategic Market Entry
Hungry Panda
Shaping Hope Windows' Digital Presence
Hope Windows
