Install
$ agentstack add skill-ohong-agent-skills-start ✓ scanned · ✓ verified — works with Claude Code, Cursor, and more.
Security review
✓ PassedNo issues found. Passed automated security review. · v0.1.0 How review works →
- ✓ Prompt-injection patterns
- ✓ Secret / credential exfiltration
- ✓ Dangerous shell & filesystem operations
- ✓ Untrusted network calls
- ✓ Known-malicious package signatures
What it can access
- ✓ Network access No
- ✓ Filesystem access No
- ✓ Shell / process execution No
- ✓ Environment & secrets No
- ✓ Dynamic code execution No
From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.
About
Mission Start — Enter the OODA Cycle
You are beginning execution. Each milestone is its own OODA cycle: orient to the milestone, act on it, observe the results, reorient for the next.
Pre-flight: Orient to the mission
- Check that
.mission/plan.mdexists. If not: "No mission plan found. Run/mission:planfirst." Stop.
- Check that
.mission/progress.mdexists. If not: "No progress file found. Run/mission:planfirst." Stop.
- Read
.mission/plan.md— load the full plan into your orientation.
- Read
.mission/progress.md— if milestones are already done, ask: "Previous progress detected. Resume with/mission:continue, or restart from milestone 1?"
- Mismatch check: Read the plan's build/test/lint commands and run them once. Do they work? If not, fix the environment before starting. Starting with a broken environment means every observation will be contaminated.
Execution: The Milestone OODA Loop
Follow the execution protocol in [../../references/execution-protocol.md](../../references/execution-protocol.md) starting from Milestone 1.
Each milestone follows the OODA cycle:
Orient
- Re-read
.mission/plan.mdfor this milestone's goal and acceptance criteria - Re-read
.mission/progress.mdfor current state - Read the source files you'll be modifying — never edit from memory
- If
.mission/learnings.mdexists, check for relevant patterns
Decide (or act implicitly)
- Familiar pattern? Act directly. Standard CRUD, common test setup, well-known library usage — move with implicit guidance.
- Novel situation? Deliberate explicitly. Unfamiliar architecture, unexpected file structure, complex integration — slow down, read the relevant code, and think before acting.
Act
- Write code, staying scoped to THIS milestone only
- Commit after each meaningful change:
mission(N/total): {description} - Every action is a hypothesis: "I believe this change will make the acceptance criteria pass"
Observe
- Run ALL acceptance criteria. This is reality speaking to you — listen.
- If they pass: update progress, proceed to next milestone
- If they fail: you have a mismatch. Reorient before retrying. See failure recovery below.
Failure Recovery: Destruction and Creation
When acceptance criteria fail, Boyd's principle applies: don't retry blindly. Reorient.
Level 1 — Fix (attempt 1-3): Read the error. It's observation data. Orient to what it's telling you. Fix the specific issue. Re-run.
Level 2 — Destroy and create (after 3 failures): Your mental model is wrong. Stop coding.
- Re-read the relevant source files FROM DISK (not from memory)
- Re-read the error messages — what are they actually saying?
- What assumption did you make that was wrong?
- Destroy the old approach. Build a new one from the pieces + new information.
- If the plan itself is wrong, tell the user: "This milestone needs rethinking. Run
/mission:pivot."
Level 3 — Escalate: Write what you tried to .mission/progress.md. State the goal, what failed, your best guess at root cause. Ask the user for direction.
Context Management
- After completing a milestone, if the conversation is long (20+ minutes of work, many file reads): recommend
/clear→/mission:continue. Fresh context is forced reorientation — a feature, not a bug. - Use subagents sparingly. Delegate only when the user requests it or independent parallel work clearly repays the coordination cost.
- Keep the current milestone freshest in context. Read background files first, then the current milestone details last. LLMs attend most strongly to the end of context (recency anchor).
Key Rules
- Never skip validation. Every action is a hypothesis. Test it.
- Never self-assess. "It looks right" is not observation. Run the command.
- Never expand scope. Only touch files in THIS milestone. New work =
/mission:pivot. - Commit per milestone. Each passing milestone gets a checkpoint commit.
Source & license
This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.
- Author: ohong
- Source: ohong/agent-skills
- License: MIT
Install and usage instructions live in the source repository linked above.
Reviews
No reviews yet — be the first.
Write a review
Versions
- v0.1.0 Imported from the upstream source.