AGENT SKILL

manifest

Use when starting a non-trivial feature, migration, refactor, or multi-step task in ANY repo. Creates and continuously maintains a living markdown manifest at `<repo-root>/tasks/<slug>/manifest.md` (inside the repo being worked on — NEVER the skill dir or ~/) capturing the problem, decisions, exploration results, plan, open questions, test + rollout plans, and final review. Trigger on "build X", "add X", "migrate X", "refactor X", tasks needing 3+ steps, or whenever a plan-before-code loop would help. NOT for single-line fixes, obvious bugs, or one-shot questions.

License 许可
MIT
所属仓库
auth-02/hub
最近更新
2026-08-12
Stars
★ 3
Forks
0

01 前言 Front Matter

name
manifest
description
Use when starting a non-trivial feature, migration, refactor, or multi-step task in ANY repo. Creates and continuously maintains a living markdown manifest at `<repo-root>/tasks/<slug>/manifest.md` (inside the repo being worked on — NEVER the skill dir or ~/) capturing the problem, decisions, exploration results, plan, open questions, test + rollout plans, and final review. Trigger on "build X", "add X", "migrate X", "refactor X", tasks needing 3+ steps, or whenever a plan-before-code loop would help. NOT for single-line fixes, obvious bugs, or one-shot questions.

02 资源文件 4

03 SKILL.md 正文

manifest

Every feature gets a manifest — a living markdown doc that declares the problem, the plan, the findings, and the eventual outcome. It lives inside the repo you are working on, right next to the code, so reviewers and future sessions read it alongside the diff. The deliverable is not just the code — it is the manifest that explains what/why/how.

Where the manifest lives

Always inside the repo being worked on, at <repo-root>/tasks/<slug>/manifest.md. Resolve <repo-root> from the current working directory: the nearest ancestor that is a repo root (contains .git / pyproject.toml / package.json / similar). In a multi-repo workspace, that is the specific repo the work belongs to — not the workspace root above it.

Never write to:

  • The workspace root above any repo. Manifest-collection tooling scans each repo's tasks/, so a manifest placed above a repo would be invisible.
  • The skills directory (~/skills) or the skill directory itself.
  • Anywhere outside a repo.

If the current working directory is a multi-repo workspace root (not inside a specific repo) and the user hasn't named a target repo, ask which repo this work belongs to before creating the file. Don't guess.

Why this matters: tooling that collects manifest docs looks for manifest.md files under each repo's tasks/**/manifest.md. A manifest in the wrong location (or named anything other than manifest.md) won't be picked up, defeating the proactive-update workflow below.

Create the tasks/ directory if it does not exist. Every task gets its own subfolder. At init, create only tasks/<slug>/manifest.md — do not pre-create data/, artifacts/, or runs/ upfront. Each subdir is created lazily, only when something actually needs to go in it (the first data file → data/, the first probe note/script → artifacts/, the first validation run → runs/<date>/). Empty scaffolding directories are noise; a fresh task is just its manifest.md.

<repo-root>/tasks/
└── <slug>/
    ├── manifest.md      # the living manifest          (template: templates/manifest.md)
    ├── data/            # referenced/source inputs only: PDFs, Excel, CSVs, datasets, JSON
    ├── artifacts/       # inspection notes, scripts,   (template: templates/inspection.md)
    │                    # probe outputs
    └── runs/            # dated outputs from actual test/validation runs
        └── YYYY-MM-DD/  # one folder per run date; SSE logs, tables, run CSVs go here

Rules:

  • The manifest is always named manifest.md (never <slug>.md) — this makes it predictable for tooling (standup collection, grep).
  • data/ holds referenced / source inputs only — files given to or consumed by the task (example workbooks, sample PDFs/Excel/CSVs, datasets, JSON inputs). Create it on demand when such a file first enters the task (subfolders OK, e.g. data/results/). Keep generated virtualenvs, node_modules, and caches out of the task tree entirely.
  • artifacts/ holds everything I generate — inspection notes, one-off scripts/probes, AND generated output deliverables (PDFs, guides, derived spreadsheets). Generated files are never put in data/; if I produced it, it goes here (dated run outputs go in runs/).
  • runs/ holds time-stamped outputs from actually firing queries or running validation: SSE logs, parsed tables, benchmark results. One subfolder per date (YYYY-MM-DD/). Run outputs stay under their dated folder — they are not moved into data/.
  • Never put inspection files, data files, or run outputs at the tasks/<slug>/ root — keep the root clean (just manifest.md + the subdirs).

Copy from ~/skills/manifest/templates/<name>.md when creating the manifest — fill in the title, ask, and any decisions already captured, then leave the rest as placeholders.

Slug: lowercase, hyphenated, imperative (ndjson-to-parquet-migration, add-sso-login, refactor-payment-flow).

The tasks/ directory is gitignored — manifests, their data/, artifacts/, and runs/ are local-only and never committed or pushed. Don't try to git add them. (When setting up a brand-new repo, add tasks/ to its .gitignore first.)

Core principle

Write the manifest before the code, then keep it honest. Every user answer, every surprise from exploration, every change of plan lands in the manifest in the same session. If the manifest drifts from reality, the workflow has failed.

When to trigger

Invoke at the start of:

  • A feature build (e.g. "add X", "support Y", "integrate with Z").
  • A migration (format, schema, dependency, infrastructure).
  • A refactor touching 3+ files.
  • Anything where the user says "plan" / "first figure out" / "explore" / "let's design".
  • Any task requiring exploration before implementation is sensible.

Skip for: typo fixes, single-line bug fixes, renames, answering a question, trivial one-shot edits.

The manifest structure

Use these sections in this order. Sections can be empty with placeholder text; don't skip them.

# <Feature / Task Title>

## Context
One paragraph on the ask as you understood it from the user, plus any business or technical background. Not the plan — the problem.

### Decisions
Numbered list of decisions the user has already made (or made during the session). Each line should stand alone.
1. <decision> — captured <date or "from session">.
2. ...

## Exploration Results
What you found in the codebase. Cross-link `file:line` for every claim. Sections suggested (adapt as needed):
### Consumers / callers
### Data shape / schema
### Related subsystems
### Out of scope
### Not found
(Keep this section factual. No plan verbs like "we will" or "let's".)

## Plan
Checkbox list of steps. Each step: one line heading + a few sub-bullets describing the edit. Update checkbox state (`[x]`) as you go.

- [ ] **Step 1 — <name>** `<path/to/file>`
  - What to change
  - Why / what it unblocks

## Open questions
Numbered list. When the user answers, append `-> <answer>` inline on the same line; do not delete the question.
1. <question>?
2. <question>? -> <user's answer, captured inline>

## Confirmed findings (after real exploration)
Populate AFTER running exploration code / probes. Distinguish "assumed" from "confirmed". Includes:
- Tables of concrete observations (blob names, schemas, counts).
- Callouts for surprises that invalidate the initial plan — use ⚠️.

## Testing plan
Concrete steps. Day-by-day if multi-day. Include: parity baselines, canary, fallback tests, scheduled runs, exit gate.

## Rollout plan
Stages, rollback procedure, coordination notes, gotchas.

## Review
Fill in at the end. Files changed (with one-line per-file summary), how to cut over, known follow-ups, what was flagged out-of-scope.

Workflow stages

1. Capture the ask (before any tool calls other than creating the manifest)

  • Read the user's message carefully. Paraphrase the problem back in one sentence before doing anything.
  • Create tasks/<slug>/manifest.md populated from the ask — and nothing else; data/, artifacts/, runs/ come lazily per the rule above. "Exploration Results" and "Confirmed findings" are empty placeholders. "Plan" has rough shape only — don't overcommit before exploring.
  • If the user gave a spec/list/table, paste it verbatim under Context so it's not lost.
  • After creating the manifest, write the slug to tasks/.current-task (one line, no newline padding) so the shell prompt can display it: echo -n "<slug>" > tasks/.current-task. This file is the prompt's source of truth — update it whenever the active task changes mid-session (e.g. user switches to a different task on the same branch).

2. Explore

  • Use grep / read / file-listing to find consumers, patterns, and current behaviour. Delegate to subagents (Explore / general-purpose) for broad sweeps. Limit direct reads to files you've identified.
  • Write findings to Exploration Results as you go — file:line references, not summaries.
  • For anything that requires real-world data (bucket contents, DB schema, external API shape), write a read-only probe script/asset/query, run it, and save the output under artifacts/ (e.g. artifacts/inspection.md, artifacts/<probe>.csv). Never guess what remote state looks like.
  • Flag surprises immediately in Confirmed findings with ⚠️ — e.g. "field X is dropped from the new schema", "file is unsharded where docs said sharded".

3. Plan + open questions

  • Convert exploration into a step-by-step Plan with checkboxes. Keep each step minimal — bug fix-sized if possible.
  • Extract decisions the user must make into Open questions. Ask them all at once, numbered, so the user can answer inline.
  • When the user answers, edit the manifest: append -> <answer> to each question. Then merge the answer into Decisions.

4. Implement

  • Update the plan's checkboxes as steps complete. Don't batch — mark each done when it's actually done.
  • When a step reveals something unexpected (a missing field, a broken assumption), stop, update Confirmed findings, update the plan if scope changed, then continue.
  • For risky writes (DBs, buckets, external systems shared with other envs), design a local/dev-only gate (e.g. a local_csv_output_dir config, a dry_run flag, a dev_only toggle) — then document it in a dedicated subsection of the manifest.

5. Tests + rollout

  • Write a Testing plan with explicit parity baselines ("diff row counts vs main"), canary (one asset / one file at a time), fallback verification, and an exit gate.
  • Write a Rollout plan with stages, a rollback procedure (ideally config-only, no redeploy), and coordination notes (who needs to know before flipping production).
  • If the user has constraints like "don't touch demo", "dev only", "keep fallback" — surface them at the top of the rollout plan.

6. Finalize

  • Populate the Review section: files changed (with a one-line summary each), how to cut over, follow-ups, out-of-scope items.
  • Offer the next natural step (branch + push + MR, schedule a cleanup agent, etc.) but don't take destructive/shared actions without confirmation.
  • On git commit / branch creation: the commit message should mirror the Review summary. Keep the manifest out of the commit if it's gitignored — otherwise include it.

Operating rules

  • One manifest per task. Don't create tiers/sub-docs unless the user asks for it.
  • Append-only for decisions and open questions. Never delete history; strike-through or annotate if a decision reverses.
  • Real data over assumptions. If the manifest makes a schema claim, it must be backed by a probe saved alongside.
  • Update in the same session. Don't defer updates with "I'll record this later".
  • Cross-link everything. file:line for code claims; relative paths to inspection files; commit SHAs in the Review.
  • Defer code until the plan is agreed. If the user hasn't confirmed direction, don't start edits — unless they explicitly said "just do it".
  • Protect shared infra. When the work touches DBs, buckets, queues, or external services used by demo/prod, add an isolation knob early and document it.
  • Never invent skill / subagent / tool names. Only use ones that exist in the environment.

Proactive updating — do this without being asked

Updating the manifest is part of the task, not an extra chore the user has to request. Once a manifest exists for the current work, treat every one of the events below as an automatic trigger to edit it in the same turn the event happens — before moving on, before the next reply, before running more tools. The user should never have to say "write this down" / "log the decision" / "update the task file".

Triggers that automatically update the manifest (in the same turn):

  1. User answers an Open Question → append -> <answer> inline on that question, then promote the answer to Decisions (numbered, dated).
  2. User reverses or refines an earlier decision → strike-through the old decision, add the new one with a date and a Superseded (YYYY-MM-DD) tag; never delete the old line.
  3. User approves a course of action ("ok", "do it", "that's fine", "whatever you choose") → record the concrete choice that was on the table as a Decision before executing.
  4. Exploration reveals something that invalidates the plan → add a ⚠️ entry to Confirmed findings and update the affected Plan step, in the same turn.
  5. A plan step completes → flip its checkbox [ ] → [x] immediately; if the step split into sub-steps, list them inline.
  6. New scope added mid-session → either add a new Plan step or (for a genuinely separate task) spin up a second manifest file; do not silently grow an existing step.
  7. A follow-up iteration after "Review" (user asks for a tweak after you declared done) → add a Follow-up increments subsection under Review with what changed and why, in the same turn. Don't reopen the main Plan unless the change is large.
  8. Files are edited, created, or deleted → update Files changed in Review the same turn (don't wait until the user asks for a summary).
  9. gitignore / tasks/ visibility → if tasks/ is not in .gitignore, note it; if the user wants it local, add it to .gitignore without asking again after the first confirmation.

Rule of thumb: if you're about to say something in chat that is a decision, a finding, a change, or a reversal, the manifest gets the same content first (or at the latest, in the same turn). Chat is ephemeral; the manifest is the record.

When you can't tell whether something warrants a manifest update, err toward writing it down — a slightly over-documented manifest costs nothing; a silently-drifted one costs trust.

Anti-patterns to avoid

  • Writing a grand plan before any exploration. Plans written without grepping are fiction.
  • Letting the manifest and the code diverge. If you updated code, update the manifest in the same turn.
  • Mixing "the ask" and "the plan" in one section. Separate what the user wanted from what you intend to do.
  • Dropping the fallback / rollback section because "we won't need it". Keep it — future-you will.
  • Hiding surprises. If exploration invalidates a decision, say so with ⚠️ and update Decisions with the revised stance.
  • Committing the manifest without checking gitignore.
  • Narrating every step in chat while the manifest stays sparse. The chat is ephemeral; the manifest is the record.
  • Waiting to be told, batching updates, or leaving reversed decisions undocumented — all violations of Proactive updating above: update in the same turn, unprompted, keeping superseded decisions struck-through with a date.

Templates

Copy from ~/skills/manifest/templates/ — but only when each is actually needed (see the lazy-creation rule above), not all upfront:

  • manifest.md — the main manifest scaffold (Context, Decisions, Exploration Results, Plan, Open questions, Confirmed findings, Testing plan, Rollout plan, Review). Copied at init — lands at tasks/<slug>/manifest.md.
  • inspection.md — for raw exploration probes. Copied only when you start probing — lands at tasks/<slug>/artifacts/inspection.md (creating artifacts/ then).
  • notes.md — free-form scratch. Copied only when you need it — lands at tasks/<slug>/artifacts/notes.md.

For the manifest, fill in the title and any context you already have from the user's ask, then keep the other sections as placeholders to populate as the work progresses.

Example invocations

  • "Let's add support for parquet in the pipeline" → create tasks/add-parquet-support/manifest.md, start at stage 1.
  • "Refactor the auth middleware to use JWT" → create tasks/refactor-auth-to-jwt/manifest.md.
  • "Migrate sector_exposure to a normalized table" → create tasks/normalize-sector-exposure/manifest.md.
  • "Integrate with Acme Payments" → create tasks/integrate-acme-payments/manifest.md.

Example

examples/migrate-uploads-to-object-storage.md is a comprehensive manifest to model yours on: a design-first migration showing a full Context paragraph, a numbered Decisions log (with dates), exploration findings with file:line cross-links, a Stack section, a per-layer Plan, testing + rollout plans, and the Review placeholder. Use it to calibrate altitude and how much to capture at each stage.

Not for

  • "Fix the typo on line 12" — too small.
  • "Why is this function slow?" — it's a question, not a feature.
  • "Rename foo to bar across the repo" — rename, no design surface.
  • Multi-turn conversations where the user has already started implementing without a plan and just wants help finishing — if the manifest would be retrofit, ask first.

04 同仓库更多 Skills

auth-02
auth-02/hub

stacked

Decompose LARGE, complex, or risky changes into a stack of small, dependent, independently reviewable units. Use only when the change is genuinely big: multiple architectural layers with functionally distinct concerns (models + storage + logic + API + UI), or a diff too large to review as one unit. Trigger on "stacked" / "stacked PRs" / "stack this" / how to break down a large change — or when a "build X" / "refactor W" request is clearly multi-layer and large. Do NOT use for small or moderate changes: bug fixes, tweaks, config, single-file or single-layer changes belong on ONE branch as ONE change — a stack of tiny branches adds overhead without value.

★ 3
auth-02
auth-02/hub

kagaz

Kagaz (कागज़) — a personal design system for all frontend and document work — editorial-technical-archival, serif display + mono metadata on warm paper tones with a single rust accent. Use for ANY frontend, UI, web page, landing page, dashboard, artifact, component, mockup, poster, slide deck, report, or PDF/document design, for HTML→PDF conversion, and for any "make this look good" / "style this" / "make slides" / "turn this into a PDF" request — even when no design preference is mentioned. Overrides generic AI defaults (Inter-on-white, purple gradients, cool grays). Ships an HTML-to-PDF converter with visual QA and on-brand SVG charts.

★ 3
auth-02
auth-02/hub

dak

暂无简介

★ 3
auth-02
auth-02/hub

change-log

Map a task's changes as an editable draw that renders an interactive change-map on Save — nodes are functional CHANGES (a capability that moved), arrows are dependencies, and clicking a change deep-dives into its file / function / test detail in a side inspector. Use when someone asks "what changed", "map this change", "diagram this PR/branch", "show how the changes connect", or as the closing step of a task once code has landed. Reads the diff + the task's manifest (the *why*) + Hub's MCP task context, then writes ONE `.excalidraw` into `tasks/<slug>/change-log/`; the user edits it and Saving renders the sibling HTML. The agent does the reading — Hub has no model. `/change-log <task-slug> [--since <ref>] [--doc]`.

★ 3
auth-02
auth-02/hub

auth

暂无简介

★ 3
auth-02
auth-02/hub

rate_limiting

暂无简介

★ 3