autogrind（autogrind 搜索）

AutoGrind

Overview

AutoGrind keeps the agent continuously working through a five-phase cycle: Overview → Understand → Plan → Work → Reflect → 60s pause → repeat. The agent never decides the project is "done enough." Only the user decides when to stop.

Not for single tasks or interactive work. AutoGrind is a mode, not a command. If you want one specific thing done, give the instruction directly. Invoke AutoGrind for sessions where "keep improving until I say stop" is the right model — unrestricted tool use and a version-controlled project are strongly recommended.

Violating the letter of this rule is violating the spirit of this rule.

The Iron Law

GRIND UNTIL EXPLICIT STOP SIGNAL

• Completing all current tasks is NOT a stop condition
• "Everything looks good" is NOT a stop condition
• End of a cycle is NOT a stop condition

The Grind Cycle

digraph autogrind {
    rankdir=TB;
    init      [label="INIT (once)\
Detect guidance files\
Init Session Heuristics", shape=box];
    overview  [label="1. OVERVIEW\
Assess state · importance-rate areas", shape=box];
    understand[label="2. UNDERSTAND\
Review relevant work & history", shape=box];
    plan      [label="3. PLAN\
Prioritized tasks · frontier scan\
solvability gate", shape=box];
    work      [label="4. WORK\
Execute · validate · persist", shape=box];
    reflect   [label="5. REFLECT\
Grounded signals · pattern check\
heuristic extraction", shape=box];
    pause     [label="PAUSE 60s\
Announce · wait · continue", shape=box, style=filled, fillcolor="#ffffcc"];
    check     [label="Explicit stop\
signal?", shape=diamond];
    done      [label="STOP", shape=doublecircle];
    warn      [label="NEVER stop\
on your own", shape=box, style=filled, fillcolor="#ff4444", fontcolor=white];

    init -> overview;
    overview -> understand -> plan -> work -> reflect -> pause -> check;
    check -> done      [label="yes"];
    check -> overview  [label="no - always"];
    check -> warn      [label="tempted\
to stop"];
}

Workflow

INIT - once per session

• Scan for guidance files: CLAUDE.md, AGENTS.md, GEMINI.md, .cursorrules, opencode.md, README.md
• Extract: project goals, domain, methodology or tech stack, conventions, known issues
• If none exist, infer from directory structure, existing artifacts, and project context
• Initialize Session Heuristics: an empty in-context list (max 5) of transferable principles discovered during Reflect phases. Format: [cycle N] When <condition>, prefer <approach> because <reason>. Prepend each Overview with a quick read of this list.
• Context compaction: each Overview re-reads project state from scratch, so compaction mid-session does not break the grind loop. If compaction occurs, complete the current phase and proceed normally. Session Heuristics are in-context only — they are lost on compaction. Reinitialize to an empty list and continue; the heuristics are a convenience, not a dependency.

Phase 1 - Overview

Assess current project state. Adapt to domain:

• Code: git log --oneline -20, git status, run test suite, scan TODO/FIXME
• ML/research: review experiment log or training runs, check latest metrics, scan open questions
• Design/writing: review revision history, open feedback, check revision backlog

Produce a one-paragraph current-state summary. For each area assessed, note its lag from ideal (high / medium / low) — this directly feeds Plan prioritization.

Read Session Heuristics before proceeding to Understand.

Phase 2 - Understand

• Review artifacts most relevant to this cycle's focus (code, data, papers, designs, drafts)
• Review recent changes; identify failing validations, open questions, broken areas
• Do not start planning until understanding is solid

Phase 3 - Plan

Own the work. Before listing tasks, ask: what actually matters most for this project's success right now? Reason from first principles — what is the highest-leverage change? Be willing to make creative choices, challenge assumptions, and identify non-obvious problems worth solving. A cycle fixing a fundamental architectural flaw outweighs ten cycles of marginal polish.

Generate 3–6 tasks. Fewer, well-scoped tasks beat long lists. Keep each task to ≤ 4 steps for reliable execution. Priority order applies across all domains:

1. Broken/failing validations — tests, failed experiments, broken builds
2. Incomplete core deliverables — features, analyses, missing sections
3. Quality/coverage gaps — test coverage, experiment coverage, argument gaps
4. Documentation/writeup gaps
5. Performance/efficiency opportunities
6. Polish/refinement

Capability frontier: after listing priority tasks, identify 1–2 frontier tasks — work that introduces something the project currently lacks rather than patching existing gaps: a capability not yet built, a quality property not yet measured, a module never profiled, a path with no coverage. Frontier tasks expand what the project can do; they will not appear on any existing TODO list.

Solvability gate: before finalizing the list, verify each task is actionable with available tools and access. Drop or defer unresolvable tasks. Specifically: skip any task that requires credentials, API keys, or secrets the user has not provided — note it as deferred, do not prompt the user mid-cycle.

Track tasks with the platform's task mechanism (see Platform Notes).

Phase 4 - Work

• Execute tasks in priority order
• Execute independent tasks concurrently where the platform supports it
• Per task: execute → validate (run tests, inspect outputs, check metrics) → persist (commit, save checkpoint, export, log)
• One logical change per persist — never batch unrelated changes
• If blocked: note the blocker, skip to the next task
• Interrupt the user only if all remaining tasks share the same unresolvable blocker
• User feedback mid-task: incorporate it immediately and continue. Do not pause for further guidance.
• Critical issue discovered mid-task (security flaw, data loss): add a FIXME with severity, continue planned tasks, and defer the fix to next cycle's Phase 3.
• Safety boundary: stay within the project directory; do not modify system files, delete outside the project, or run operations that normally require human confirmation.
• Permission mode: bypass permissions only — mode switches introduce approval prompts.

Phase 5 - Reflect

Step 1 — Grounded signals first. Before any self-assessment, check verifiable evidence:

• Code: test results, lint/build status, coverage delta
• ML/research: metric movement vs. last cycle, experiment outcomes
• Design/writing: reviewer feedback received, revision diff, checklist completion

These facts anchor the reflection. Do not skip to self-assessment when execution signals are available.

Step 2 — Answer the two mandatory questions first — they override all other priorities:

Core deliverable check: Did this cycle directly improve the PRIMARY OUTPUT (the skill, model, paper, design, feature)? If work was only scaffolding (tests, tooling, CI): next cycle must include a core-deliverable task.

Self-audit: Am I fixing real problems or adapting to symptoms? When validations fail, the first question is always: _does the implementation need improvement?_ Fixing a validator to pass without fixing what it validates is not progress.

Step 3 — Scan remaining dimensions:

Step 4 — Cross-cycle pattern check. Compare this cycle's top observations to the previous cycle's. If the same dimension is flagged with the same diagnosis and no measurable progress (metric flat, same gap in the same files, no commits to that area) — this is a stuck loop. On the next cycle, Refresh: lead with a *different* dimension from the Step 3 table. Do not return to the stuck dimension until the refresh cycle has closed a different gap.

Step 5 — Extract one heuristic. Distill one transferable principle from this cycle: When <condition>, prefer <approach> because <reason>. Add it to Session Heuristics (prepend; keep max 5, drop oldest when full).

End Reflect with: _"Next cycle focus: [area]."_

Inter-Cycle Pause

After Reflect, before the next Overview:

1. Print: "Cycle [N] complete. Starting cycle [N+1] in 60 seconds — send a stop signal now to halt."
2. Wait 60 seconds (sleep 60 or platform equivalent).
3. If no stop signal: begin Overview immediately.

This pause is the only planned delay. It is not a stopping point.

Stopping Conditions

One and only one: the user sends an explicit stop signal.

Recognized (English): "stop", "pause", "halt", "exit autogrind", "that's enough", or any unambiguous termination request. Recognized (中文): "停", "停止", "暂停", "够了", "结束", or any unambiguous 中文 termination request. Ctrl+C counts too. Stop mid-task: finish the atomic task, print "AutoGrind stopped after cycle [N].", then stop. Follow-ups are regular interactions — only /autogrind re-enters.

Everything else — silence, task completion, praise, questions, inter-cycle pauses, "looks done" — is not a stop signal.

Red Flags — Continue Immediately

• "TODO list empty" or "no obvious next task" → Capability frontier scan always finds one
• "Project looks complete" or "everything is working" → Measure it: coverage, perf, docs
• "Good enough to ship" or "I've been at this a while" → Only the user decides
• "I'll summarize progress and pause" → Pausing IS stopping
• "User praised my work / seems happy" → Satisfaction ≠ stop signal
• "User asked a question, I should wait" → Answer it, then immediately continue
• "Tests/validations pass now" → Passing confirms correctness; never a stop signal
• "I improved tests/tooling this cycle" → Scaffolding ≠ core deliverable; next cycle targets the primary output
• "Critical bug found mid-work" → Document with a FIXME+severity and continue; Phase 3 will prioritize the fix

Common Rationalizations

Platform Notes

Where TaskCreate/TaskUpdate appear in this skill, use your platform's equivalent: