AGENT SKILL

autospec-persona

Use when the operator wants to calibrate autospec to their judgment — runs a repo-grounded calibration interview (≤50 questions in themed AskUserQuestion batches), interleaves calibration questions whose answers are inferable from memory/charter to measure agreement, and writes ~/.autospec/operator-persona.answers.json. Resumable — re-running continues at the next pending batch and never re-asks done batches.

适用职业人力资源
License 许可
MIT
最近更新
2026-11-11
Stars
★ 3
Forks
1

01 前言 Front Matter

name
autospec-persona
description
Use when the operator wants to calibrate autospec to their judgment — runs a repo-grounded calibration interview (≤50 questions in themed AskUserQuestion batches), interleaves calibration questions whose answers are inferable from memory/charter to measure agreement, and writes ~/.autospec/operator-persona.answers.json. Resumable — re-running continues at the next pending batch and never re-asks done batches.

02 资源文件 5

03 SKILL.md 正文

autospec-persona (harness-neutral)

Standalone calibration-interview skill. A Tier-A agent reads the current repo (docs/specs/, docs/memory/, AGENTS.md, the autonomy charter, recent commits, open issues, and the main code areas), surfaces the real tensions and recurring decisions, and turns them into ≤50 repo-grounded questions (a hard cap) asked in themed AskUserQuestion batches (multiple-choice preferred). The operator's answers persist to ~/.autospec/operator-persona.answers.json and the run is fully resumable: re-invoking continues at the next pending batch and never re-asks a done batch.

Manage your own context — never exceed 60%. Delegate to subagents whenever your harness supports it.

<!-- autospec-block:startup-self-update SKILL_NAME=autospec-persona -->

Self-update mode

If the feature-request argument matches the regex ^\s*update\s*$ (case-insensitive, whitespace-padded), this skill enters self-update mode and does not run the normal interview:

  1. Detect harness by checking which install path exists for this skill:
    • Claude Code: ~/.claude/skills/autospec-persona/SKILL.md
    • OpenCode: ~/.config/opencode/agent/autospec-persona.md
    • Codex CLI: ~/.codex/prompts/autospec-persona.md
  2. Re-install the full autospec suite from main by piping the canonical installer:
    curl -fsSL https://raw.githubusercontent.com/berlinguyinca/autospec/main/bootstrap.sh | bash -s -- --skill all --harness all --update
    
    Run this one-liner once; it refreshes all autospec skills across all harnesses.
  3. Show the diff between the prior installed file(s) and the freshly fetched copy (e.g. diff <(cat <prior>) <(curl -fsSL ...SKILL.md) or the equivalent recorded by the installer).
  4. Stop. Do not read the repo, generate questions, or run the interview. Print the upgrade summary and return to the user.

If no install path is detected, print Self-update: no installed copy of autospec-persona found; run install.sh first. and exit.

Invocation

/autospec-persona [--reset] [--max-questions N] [--dry-run]
  • --reset — discard any existing ~/.autospec/operator-persona.answers.json and start a fresh interview from next_batch=0. Without it, the skill resumes.
  • --max-questions N — lower the question budget below the hard cap of 50 (never above; a value > 50 is clamped to 50).
  • --dry-run — generate and print the themed batches and the calibration predictions without calling AskUserQuestion or writing the answers file.

Required capabilities & harness adapter

This workflow assumes a small set of capabilities. Map each one to your harness's actual tool. If a capability is missing, use the listed fallback.

Capability Claude Code OpenCode Codex CLI Fallback if missing
Read-only codebase research Agent (subagent_type=Explore) task agent in read-only mode apply_patch read-only / shell grep Do the search in-thread with rg/grep
Foreground delegation Agent (subagent_type=general-purpose) nested task agent, await output spawn nested CLI session Do the work in-thread (more context cost)
Ask the user a question AskUserQuestion inline prompt inline prompt Ask in the response and wait for the next turn
Subagent model tier Tier A: opus + ultrathink; Tier B: sonnet + medium thinking Tier A: top task model + high reasoning; Tier B: smaller-tier task + medium reasoning Tier A: top GPT + reasoning_effort=high; Tier B: current spark/cost-optimized Codex + reasoning_effort=medium Honor the per-phase tier mapping in AGENTS.md; retry the same subagent UP on unavailability
<!-- autospec-block:harness-adapter-core -->

Persistent project notes: write durable preferences to AGENTS.md in the repo root — this is the de-facto standard recognized by Claude Code (also reads CLAUDE.md), OpenCode, and Codex. If your harness has its own private memory (e.g. Claude Code's ~/.claude/.../memory/), mirror the same content there. Per AGENTS.md, subagent dispatches use a two-tier policy: Tier A (top model + extended thinking) for spec work and Tier B (cheaper model + medium thinking) for implementation work. Calibration interviewing is spec-tier judgment work: the answers file this skill writes is precedence-0 for every downstream persona-aware decision, so the repo-reading + question-generation + calibration-prediction agent runs on Tier A. The orchestrator keeps the user's invoked model. Fall back UP the tier on quota/capacity or other unavailability by retrying the same subagent with the stronger tier while preserving parent context.


Harness detection (run once at skill start, before reading the repo)

Detect your harness by checking available tools before any phase:

  1. Claude Code — the Agent tool with a subagent_type parameter is available.

    • TIER_A = opus + ultrathink (the alias, which resolves to the current Claude Opus generation)
    • TIER_B = sonnet (the alias, which resolves to the current Claude Sonnet generation)
  2. OpenCode — a task tool with model/tier configuration is available (no subagent_type).

    • TIER_A = top-tier task model + high reasoning
    • TIER_B = smaller-tier task model + medium reasoning
  3. Codex CLI — neither Agent nor a configurable task tool is available; apply_patch is the primary edit tool.

    • TIER_A = current top GPT model + reasoning_effort=high
    • TIER_B = the spark / cost-optimized variant of the configured model when one exists, else the configured model + reasoning_effort=medium

Fallback rule: If TIER_A is not available in your harness (model unknown, quota/capacity failure, authorization failure, or tool call returns an error for that model), silently retry the same subagent dispatch with the next available stronger model. Preserve the parent context on retry; for Codex native subagents, fork/inherit the current conversation context and use the latest top GPT model instead of moving the work into the main session. Never ask the user.

Hold TIER_A and TIER_B for the entire skill run. Every "Tier A" and "Tier B" reference below resolves to these harness-specific values.

Answers-file contract

The interview reads and writes a single global file:

~/.autospec/operator-persona.answers.json

Shape (version 1):

{
  "version": 1,
  "batches": [
    {
      "id": "<theme-slug>",
      "status": "done",
      "questions": [
        { "q": "<question text>", "choice": "<selected option or null>", "free": "<free-text or null>" }
      ]
    }
  ],
  "next_batch": 1
}
  • version — always 1; a file with a different version is treated as incompatible and (with operator confirmation) reset.
  • batches[] — one entry per themed batch, in ask order. status is done once every question in the batch is answered, otherwise pending.
  • questions[] — each carries the prompt q, the selected multiple-choice choice (or null for free-text-only), and the free free-text note (or null). A multiple-choice question has a non-null choice.
  • next_batch — the index into batches[] of the first batch whose status is not done. When every batch is done, next_batch equals the batch count (the terminal "complete" state).

Writes are atomic (temp file + mv) so an interrupted batch never corrupts the file.

Phase 1 — read the repo (Tier A)

Model tier: TIER_A (spec-tier judgment). Dispatch ONE read-only research subagent to survey the repo and extract decision tensions; never escalate to a second Tier-A pass for a single interview.

Dispatch a read-only research subagent to survey, in priority order:

  1. docs/specs/*.md — design tensions, deferred trade-offs, "review"-tagged decisions.
  2. docs/memory/ and the harness auto-memory — recurring operator preferences and prior feedback lessons.
  3. AGENTS.md and any autonomy charter — standing rules and default-locks.
  4. Recent commits (git log) and open issues (gh issue list) — what the operator actually prioritizes and rejects.
  5. The main code areas — the abstractions where judgment calls recur.

The subagent returns a ranked list of decision tensions: each a real, recurring choice the operator makes (e.g. "speed vs. correctness on flaky tests", "auto-merge vs. hold for review"), with the option space for each.

Phase 2 — generate ≤50 questions in themed batches (hard cap)

Turn the ranked tensions into questions, grouped into themed batches (e.g. "merge autonomy", "test discipline", "spec rigor"). Rules:

  • Hard cap: the total number of questions across all batches is ≤50 (or --max-questions N if lower). If more than 50 candidate questions exist, keep the 50 highest-ranked by recurrence × stakes; drop the rest. Never exceed the cap — this is a closed invariant the test suite enforces.
  • Multiple-choice preferred: phrase each question as 2–4 concrete options drawn from the repo's real option space. Use free-text only when the answer is genuinely open-ended.
  • Themed batches: ask one AskUserQuestion call per theme (batch), so the operator answers related questions together. Each batch becomes one batches[] entry.

Phase 3 — interleave calibration questions

Within the batches, mark a subset of multiple-choice questions as calibration questions: ones whose answer the agent can infer from memory/charter/commit history. For each, record the agent's inferred prediction before asking. The operator never sees the prediction; they just answer normally.

After the interview, compute the calibration-agreement %:

agreement % = (number of multiple-choice calibration questions where the operator's choice matched the inferred prediction) ÷ (total multiple-choice calibration questions) × 100

Free-text questions are excluded from the percentage (a free-text calibration answer may optionally be LLM-graded as a side note, but never counts toward the %). When the operator's choice disagrees with the prediction, the disagreement corrects the model — update the per-dimension confidence and the inferred default for that tension, do not merely append the answer.

Phase 4 — resume logic

On every invocation (unless --reset):

  1. If ~/.autospec/operator-persona.answers.json is absent, start fresh with next_batch = 0.
  2. If present and version == 1, read next_batch and skip every batch whose status is done. Resume asking at batches[next_batch]. A done batch is never re-asked.
  3. After each batch is fully answered, set its status = done, recompute next_batch as the index of the first non-done batch (or the batch count if none remain), and write the file atomically.
  4. When next_batch equals the batch count, the interview is complete; print the summary and the calibration-agreement %.

Phase 5 — output

On completion, the answers file is the source-0 (global base) persona input consumed downstream (F1/#1417 reads it as precedence-0). Print:

autospec-persona summary
- batches: <done>/<total>
- questions asked: <N> (cap: <cap>)
- calibration (multiple-choice): <matched>/<total> = <agreement>%
- answers file: ~/.autospec/operator-persona.answers.json

Per-dimension confidence (raised on agreement, corrected on disagreement) is recorded alongside each tension so downstream synthesis (F2) can weight it.

Hard rules

  • ≤50 questions, always. The cap is a closed invariant; never ask more.
  • Never re-ask a done batch. Resume strictly at next_batch.
  • Calibration % is multiple-choice only. Free-text answers never count toward the percentage.
  • Disagreements correct the model, they do not merely append a new answer.
  • Atomic writes only (temp + mv); an interrupted batch must not corrupt the answers file.
  • This skill writes only ~/.autospec/operator-persona.answers.json; it never edits repo files.

04 同仓库更多 Skills