achieve-goal

name

achieve-goal

description

Generic skill for tasks that require planning, debugging, or iteration before they can be completed. Three modes — execute / optimization / orchestrate — selected by the request shape. Discriminator: does the task require planning? If the request is unambiguous and mechanical ("run this notebook with python3"), do NOT use this skill — direct execution is correct. If achieving the goal requires planning, recovery from failures, or iterating toward a target, this skill applies. Triggers: "use achieve-goal to X", "iterate to achieve metric Y", "execute this multi-milestone plan", or any goal that needs planning + recovery.

Achieve Goal

Generic goal-oriented skill that plans, executes, debugs, and (when needed) iterates until the goal is met or stop conditions fire.

Modes

Mode	Trigger style	Behavior
execute	"use achieve-goal to run X"	Plan how to achieve the goal, execute, debug failures using debugging-protocol.md, recover, retry until goal is met. Differs from direct execution: this loop continues through failures rather than reporting and stopping.
optimization	"run the experiment and iterate to achieve metric < target"	Loop: try → measure → propose new approach → repeat until metric goal met or stop condition. Absorbs the previous iterative-experiment use case. References repo-specific strategy docs (e.g. recsys-experiment-strategies.md) surfaced via fetch-repo-skill.
orchestrate	"execute this multi-milestone plan" (after plan confirmation)	Create one milestone_N.md per milestone per milestone-execution-protocol.md, embed autonomous prompt, run end-to-end with re-read guard at each step.

Modes are not strictly mutually exclusive. Orchestrate may dispatch to itself in execute or optimization mode for individual milestones.

When NOT to use

• Request is unambiguous and mechanical ("run this notebook with python3", "rename function X to Y")
• Task is exploratory ("look at this and tell me what you see")
• User explicitly waives planning ("just do it")

References

• kb/context/agent-design/clarification-protocol.md — used in the planning phase of all modes
• kb/context/agent-design/debugging-protocol.md — used on failure in any mode
• kb/context/agent-design/milestone-execution-protocol.md — used in orchestrate mode
• kb/context/agent-design/autonomous-behavior.md — embedded at the top of milestone files and any execution document
• Repo-specific context docs (e.g. kb/context/experiments/experiment-policies.md, kb/context/experiments/recsys-experiment-strategies.md) — surfaced via fetch-repo-skill when working on a registered repo
• ~/.claude/skills/dispatching-parallel-agents/SKILL.md (obra/superpowers) — methodology for concurrent subagent workflows; consult before scaling out parallel work in execute or optimization mode
• ~/.claude/skills/verification-before-completion/SKILL.md (obra/superpowers) — evidence gating before declaring a milestone or goal met
• ~/.claude/skills/cli-review/SKILL.md — CLI-based review gates via Gemini CLI (plans, large corpora) and Codex CLI (code); consult for routing and invocation patterns
• For ML benchmarks: kb/context/experiments/dataset-sources.md + scripts at ~/.claude/tools/experiment-tools/ (download_dataset.py, train_mlp.py)

Steps

Phase 1 — Clarify (all modes)

1. Apply kb/context/agent-design/clarification-protocol.md — close ambiguity in goal, success criteria, scope, constraints, failure handling, interaction expectations
2. Detect mode from clarified goal:
   • Single-stage goal with potential failures → execute
   • Iteration toward a metric target → optimization
   • Multi-milestone with checkpoints → orchestrate
3. State the chosen mode and the plan back to the user. Interactive invocation: wait for confirmation before proceeding to Phase 2. Autonomous invocation (dreaming mission, NEVER-STOP context, scheduled/headless run): do NOT wait — state the plan, proceed, and flag the plan in the mission log / REVIEW_INBOX for retroactive review

Phase 2 — Plan

• For execute: lay out the steps and known risk points; identify dispatch targets (which subagents/skills will run each step)
• For optimization: define the metric, the iteration unit, the proposal strategy (reference recsys-experiment-strategies.md if recsys), and stop conditions
• For orchestrate: enumerate milestones, success criteria for each, and any cross-milestone state

CLI Review Routing (apply to every plan before confirming with user):

Is the plan large, cross-cutting, or uncertain?
  → add a Gemini CLI review gate before Phase 3 (see cli-review skill)

Will implementation touch non-trivial code across multiple files?
  → add a Codex CLI review gate after Phase 3 execution (see cli-review skill)

Reviews are optional gates — include them when task risk or complexity justifies the overhead. Reference ~/.claude/skills/cli-review/SKILL.md for routing rules and invocation patterns.

Phase 3 — Execute

• Dispatch to subagents (no single-agent execution loops)
• Embed the autonomous-behavior.md NEVER STOP + HEADLESS-SAFE prompt verbatim at the start of execution AND in every subagent brief — briefs are the only channel subagents receive (memory files never reach them)
• On failure → apply debugging-protocol.md (reproduce → minimize → validate single → scale)
• For orchestrate specifically: create milestone_N.md files per milestone-execution-protocol.md; write to the active project folder

Phase 4 — Stop conditions

Condition	Action
Goal met	Report; close milestone files if any; summarize
3 consecutive failures with no path forward	Escalate to user
Hard iteration cap reached (configurable, default 10)	Surface to user
A step would need a confirmation-gated command (rm/mv/kill/pkill/sudo/destructive git)	Never issue it in autonomous mode — use the HEADLESS-SAFE substitute from autonomous-behavior.md; if no substitute fits, escalate and log via document-agent-failure
User interrupts	Pause, summarize state

Output

• Goal achieved → summary of what was done, key decisions, any open follow-ups
• Or: structured escalation to user with current state, blockers, and proposed next steps