一键在 Manus 中运行任何 Skill

loft-plan-workflow

星标6

分支0

更新时间2026年7月23日 21:21

Procedures for plan-shaped work — opening a plan, closing/deferring one, promoting a doc to a plan, choosing the lightest workflow, applying value categories. Apply when the task involves plans, plan directories, a roadmap, or doc-to-plan promotion. The method is tree-agnostic; the concrete paths/templates/tracker that bind it to this repo live in one § at the bottom. Cites source docs for definitions; does not restate.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

loft-lang

loft-lang/loft

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

相关职业SOC

基于 SOC 职业分类

项目管理专家商业与金融运营类职业·SOC 13-1082

SKILL.md

readonly

同仓库更多 Skills

同仓库

draw

loft-lang/loft

Make a considered image or 3D scene from a described intent the iterative-craft way — block in, render, measure, critique against the intent, refine across passes — instead of trying to emit a finished result in one shot. Use this whenever the user wants to DRAW, sketch, illustrate, depict, compose, or paint a 2D picture, OR model / build / lay out a 3D scene, asset, or environment (e.g. a game prop, a room, a character, a landscape), ESPECIALLY when they give a subject, a mood, or a composition to hit ("an old woman at dusk", "a cosy tavern interior", "a low-poly pine tree", "make it feel ominous"). Also use when iterating on or critiquing such an image/scene. Reach for it even if the user doesn't say "iteratively" — the loop is the point, and one-shot guessing is what it replaces. Do NOT use for text-to-image prompts handed to a diffusion model, for charts / plots / diagrams / dashboards (data visualisation), or for UI / web layout.

2026-07-246

loft-write

loft-lang/loft

Reference for writing correct loft code. Apply whenever writing, editing, or reviewing .loft files. Covers types, syntax, known bugs, workarounds, naming rules, and error→fix table.

2026-07-246

engineering-rigor

loft-lang/loft

The shared discipline for any non-trivial change — fixing a bug OR designing/refactoring a load-bearing algorithm. USE THIS when about to fix a non-trivial bug (crash, silent corruption, wrong result), design or refactor a load-bearing algorithm (core representation, runtime, memory, codegen, a public contract or data format), when a change feels fragile or "longer than expected for what it does", or when reaching for an approximation in an exact-invariant domain (geometry, caching, hashing, round-trips) and can't pin the exact construction — ESPECIALLY when a fix already looks obvious (that's the cue, not the exception). It is a SPEED tool, not a brake: a coherent explanation is a hypothesis, not a conclusion, and the fastest route to a fix that sticks is the cheapest probe that could prove your idea wrong (a timing, a grep of a count, one --emit-and-look — seconds). So the moment you can state a cause or a fix, that fluency is the cue to GATHER first, not act — which also lets you fix the SOURCE and delete

2026-07-146

design-protocol

loft-lang/loft

Design Protocol 1 — a design is a testable HYPOTHESIS about an invariant, not a plan you execute. The DESIGN-mode counterpart to the engineering-rigor skill (which routes here for the design half). USE THIS whenever you are about to commit to a load-bearing design — a core representation, runtime, memory model, codegen, a public contract, a data format, or any algorithm that will carry weight: name the ONE invariant, count its re-assertion sites, build the cheapest probe that could FALSIFY each load-bearing claim, then build and validate against the written prediction. And for exact-invariant domains where you cannot even form the invariant (geometry, caching, hashing, store/memory lifetime, round-trips), flip to the constructive instrument — plot a concrete instance of the ANSWER and read the invariant off it. Reach for it when a design feels too clean, when you keep reaching for an approximation, or whenever you catch yourself about to build the first coherent design without probing it.

2026-07-076

loft-ship

loft-lang/loft

Write, bridge, verify, and publish a loft library so it behaves IDENTICALLY on all four targets — the interpreter, --native, --native-wasm, and --html (browser). Use this skill whenever the user is creating, packaging, cross-compiling, or publishing a loft library or package; mentions loft.toml, the target matrix, a wasm bridge, `loft new` / `loft install`, the registry, the sha256/signature, or making a library "work in the browser" / "work in wasm"; or is debugging native-vs-wasm-vs-interpret divergence in a library. Trigger it even when the user doesn't say "all four platforms" — any loft library that has to ship beyond the interpreter belongs here, because the cross-target parity gate and the wasm-bridge pitfalls are exactly where this work silently goes wrong.

2026-07-046

loft-codegen

loft-lang/loft

The discipline for ANY loft compiler/codegen work — fixing a store-lifetime or codegen bug, adding/altering bytecode, parser IR, the interpreter, or the native generator, or implementing a language feature that emits ops — OR restructuring any of those without meaning to change what they emit (extracting a selector, folding a special case, deleting a redundant condition). USE THIS the moment you are about to edit src/parser/, src/state/codegen.rs, src/fill.rs, src/generation/, src/compile.rs, or scopes.rs to change OR reshape what code is emitted — ESPECIALLY when a fix looks obvious and you're tempted to "just patch the generator and run the suite." The one load-bearing instrument is the same in both modes — `loft introspect` on BOTH backends BEFORE you touch the compiler — but the gate flips: a BUG FIX proves the WORKING bytecode standalone beside the broken one; a behaviour-PRESERVING REFACTOR proves the emitted IR + native Rust BYTE-IDENTICAL before/after. Skipping it is how codegen work flails and regres

2026-06-256

name	loft-plan-workflow
description	Procedures for plan-shaped work — opening a plan, closing/deferring one, promoting a doc to a plan, choosing the lightest workflow, applying value categories. Apply when the task involves plans, plan directories, a roadmap, or doc-to-plan promotion. The method is tree-agnostic; the concrete paths/templates/tracker that bind it to this repo live in one § at the bottom. Cites source docs for definitions; does not restate.
user-invocable	false

Plan workflow

What this skill is — and what's portable

The methodology below is tree-agnostic: the way you decide whether work is a plan, open it, run an investigation, and close it is identical in any project that organizes multi-phase work as plans. The concrete paths, templates, tracker, and naming that bind it to this repository live in one place — § Bind to your tree at the bottom. Carrying this skill into another tree means keeping the body unchanged and repointing that one section.

This skill is procedural-only. Definitions, rationale, and templates live in the source docs the bindings point at — not restated here.

Cross-cuts: branch / commit / push policy and bug-filing policy live in the repo's master instructions (loft: CLAUDE.md). This skill assumes you've already followed those.

Pick the lightest workflow that fits

The lightest path that holds the work is the right one. Promote up a tier only when the work genuinely needs it — most TODOs never become plans, even ones that span several sessions.

Work shape	Path
Bug fix (single root cause, fits one commit)	Fix + regression test + commit. Reference the tracker issue if one exists; otherwise the fix + test are the record. No plan, no open-work row, no archive entry.
Tiny deliverable (demo deploy, version bump)	One overview/roadmap row, or nothing. No plan.
Operational change (CI tweak, doc fix)	Direct commit.
Light TODO (the default) — fits in one row of a reference doc's table	An `## Open work` section in the relevant reference doc. Same lifecycle as a plan, just one row. Edit that doc's architecture content directly when you implement; the row and the design share a file.
Plan — multi-phase, explicit phasing, design-before-build, cross-arc deps, or its own document space	A full plan directory. Cap active plans at 2–3.

The light flow is the default. A plan earns its directory only when the work is genuinely multi-phase and benefits from its own document space.

A plan's identity is its tracker issue

One model, stated once:

Identity = the issue number in the shared plan tracker — not a local directory integer. File (or find) the issue FIRST, then name the directory after the number the tracker returns. Never pick the number by scanning existing plan directories: a sibling branch may already hold an unmerged <N>-<slug>/ for that number (e.g. a 15-debugger/ on another branch while main shows only up to 14), so scanning mints a duplicate that collides the moment that branch merges — and the issue number, once GitHub assigns it, is immutable, so the collision is expensive to unwind. The tracker's auto-incremented issue number is the one global, collision-free source of truth; the local directory tree is not.
The directory is flat: <id>-<slug>/. There are no future//finished//deferred/ subdirectories — lifecycle state lives on the issue (a label), not in the path.
There is no hand-maintained roadmap table. The overview is derived from the tracker. Don't curate a parallel roadmap of rows.

Everything below assumes this model. A tree mid-migration will still have legacy plans in the older layout — lifecycle subdirectories plus hand-maintained overview rows. Leave them where they are (or migrate opportunistically), but author new plans in the model above; don't extend the legacy layout. Your tree's legacy specifics are in § Bind to your tree.

Opening a plan (standard shape)

Use when the deliverable is a feature ship or a fix landing.

Identity — claim the issue FIRST. Create (or find) the tracker issue before the directory; the number it returns is the plan's id. Do not derive the number from the local directory tree — that misses unmerged plans on sibling branches and mints colliding duplicates (see § A plan's identity).
Flat directory <id>-<slug>/ named for the returned issue number, README copied from the standard template.
Fill Status + Goal first. Add Sub-arcs / Phase ordering / Open questions / Cross-arc dependencies / See-also as the design clarifies. Link the source issue(s) and carry the plan id in the body.
Label the issue — subject + status (future planned · active in progress · finished). (No plan type label: in a dedicated plans repo every issue is a plan, so it partitioned nothing — retired 2026-06-14.)
No roadmap row, no per-plan entry in the master-instructions index — the plan is discoverable via the tracker and the plans-overview doc. Add a master-index entry only if the plan introduces a genuinely new top-level reference concept (vanishingly rare).

Length budget: 100–300 lines. Longer means reference content is leaking into the plan — extract it to its own reference doc.

When NOT to use this shape: if the plan's first phase is characterize the problem space rather than design + build, use the investigation shape below.

Opening an investigation-style plan

Use when the deliverable is mechanism understanding before fix design — multiple failure clusters to catalogue, where source-reading alone won't converge and the fix-design decision can't be made without the catalogue.

Identity + flat directory as above, README from the investigation template; add a probes/ subdirectory.
Stage A first — write probes before reading source. Probes are the executable spec for what "understood" means: a hypothesis is confirmed when the probe-pair diff confirms it. Be liberal — missing a crucial shape is the worst failure; redundant variants cost nothing because they get attic-curated at the end of Stage A, not during. Extract at least one probe from a real consumer, not only synthetic cases — real extraction catches classes synthetic probes miss.
Run every probe on every execution mode the result can diverge across (e.g. interpreter vs compiler). Record the full results matrix.
Keep a flat probe table (one row per probe: file, shape, cluster, status). Curate into groups only when the suite is large and multiple people read it cold.
For each failure mode, write a cluster doc with a verified-vs-hypothesized accountability table — every mechanism statement is either VERIFIED (cited trace/code-line) or HYPOTHESIZED (marked). Without this column, hypotheses drift into the prose as if they were facts.
Track severity as two separate fields — corruption/panic/hang and leak — so "FIXED" can't conflate them (closing corruption while leaks persist is a false-fix trap).
Tools as needed, not upfront. Don't build a debugging framework first; add the one tool blocking progress, revert nice-to-haves, and list what you added in a Tool gaps section — tools added during the plan are part of its output.
Add a Status + next-session roadmap (per-cluster action items with effort estimates).

Probe → regression migration. Probes stay in probes/ during the investigation and graduate to the regression suite per cluster, as each cluster's fix lands — not all at once. The plan stays open during phased implementation and closes when the last cluster's regression is in the suite. A probe is graduation-ready only when it passes all of: assertions pass · clean process exit (no crash at teardown — "PASSED prints" is not enough, check the exit code) · no leak warning · bounded runtime (seconds, not a hang). A probe that passes assertions but fails any other gate stays in probes/ with a status note; graduate a representative sibling from the same cluster instead.

Closing or deferring a plan

Close the moment the work is FINISHED — do NOT gate on merging to main. "Finished" means design + build + tests + docs are all done (on the branch); at that moment the plan closes. A plan issue is a claim about the DESIGN being settled and delivered, not about the commit having reached the trunk — the code lands on main later, on its own clock, usually batched with other work. So:

The doc-side closing below (trim the README, move reference content, rewrite links, swap the label + close the issue) is doc-only work that does not need the code on main — do it when the work is finished.
Close plans in a batch, never a PR-per-plan. A ~30-min PR/CI cycle to close a single plan is waste; bundle the closing doc-changes (across several plans, even unrelated) into the next substantive push (CLAUDE.md § no-cycle-for-trivial-docs / bundle-subjects).
Because the close is a hand-close (issue closed before/independent of a Closes @PLN<n> PR merge), you MUST swap the status label yourself — see step 4's hand-close note.

Pick the outcome first:

All phases shipped → close.
Some/all phases paused with a concrete trigger → defer (Status table grows SHIPPED / DEFERRED rows; the deferred phases keep their full design content).
Paused with no concrete trigger → the design moves to the closed-by-decision register, not a deferred state.

For each shipped phase:

Tag sections REFERENCE / CLOSURE-RECORD / HISTORICAL.
Move reference content OUT to its home doc — either create-and-move (a phase that grew a whole subsystem → its own reference doc) or trim-only (content already has a home → just delete the duplicate).
Trim the README to a lead Status — DONE/SHIPPED <date> line + a cross-link to where the reference content now lives.

Common to close + defer:

Reclassify any overview rows (shipped parts leave; deferred parts stay if tracked). Set the lifecycle state on the issue (not a directory move): closing → swap status:active for status:finished and close the issue; deferring → swap for status:future, issue stays open. Don't rely on GitHub's Fixes #N — it's same-repo only and can't reach the plans repo. Instead the finishing PR carries a cross-repo close directive (Closes @PLN<n>); on merge to the trunk a close-on-merge workflow runs the repo's close-shipped-plans script to do the status:finished + close, with a stale-plan audit as the drift safety net. The manual gh issue edit is the fallback / out-of-band path.
- A CLOSED plan's status label must be status:finished (delivered) or status:declined (de-scoped) — NEVER a live status (active / future / next / closing). The swap is automatic ONLY when the finishing PR used Closes @PLN<n>; a hand-close (the issue closed directly) or a PR that used Refs instead of Closes closes the issue but leaves the label stale, so you must swap it yourself: gh issue edit <n> -R loft-lang/plans --remove-label status:<live> --add-label status:finished. This drifts silently — plans 107-110 were all closed-but-mislabeled (active/future/next) from Refs-only or hand closes — so when you touch a closed plan, verify the label matches the state; don't trust the audit to have caught it.
Grep + rewrite incoming links — THE most-skipped step. Reference content embedded in a finished plan gets linked to from other docs; those links rot when the content moves. Grep every doc for the plan's path, rewrite the links to the new home, then run the repo's drift checker to catch what grep missed.

Promoting a reference doc to a plan

Audit shipped status FIRST. A doc titled "design" often has shipped phases in its body. Grep the body for shipped|done|implemented|landed|completed and check the tree for matching code before choosing a destination — misroute risk is real, and a mid-promotion status question from the user is the brake: apply it.
Route by the audit: mostly shipped → close it directly · mostly trigger-deferred → defer · genuinely-future + multi-phase → a plan · genuinely-future + one row → leave the doc in place and add an ## Open work section (the light flow). Reference content with an open tail stays a doc + an ## Open work section — a thin "pointer-plan" that only links back to a doc is over-engineering; don't create one.
If promoting to a full plan: move the doc into the flat plan directory, apply the opening steps from step 4 onward, and rewrite incoming links to the old path (same grep as closing, step 5).

Transferable pitfalls

Hard-won and tree-independent:

Audit shipped status before routing. The commonest misroute is filing a partly-shipped doc as "future". Audit first; the user's mid-flight status question is the brake.
Prefer an ## Open work section in the reference doc over a pointer-plan. A thin plan that only links back to a doc adds a layer of indirection without saving content. One home, no pointer.
Split when broad; single-file when bounded. A doc with broad intended-to-finish scope → split into focused files. A doc with an explicit scope ceiling ("never going to ship past X") → one file with a status block.
Close = move reference out + rewrite incoming links. This is the most-skipped step and the source of link rot. Don't skip the grep + drift check.
Named value categories beat numbered tiers. Categories that name the kind of value are stable across sessions; numbered tiers (V1/V2/V3) get re-ranked constantly. Re-categorize only when scope actually changes.
No per-plan entry in the master-instructions index. It duplicates the plans-overview doc and costs tokens every session. Plans are discoverable via the overview.
Never calendar-time language in plans, roadmaps, or memory — "2–3 weeks" ships in 2 days and "quick fix" takes weeks. Use effort letters. Historical retrospectives that document the rule's validity are fine to keep.
Investigation plans get their own template and reading order (Status → Probes → Cluster docs → Roadmap). Forcing investigation work into the standard feature-ship template either bloats the README or loses the catalogue.
Probe-first beats source-first for failure-class work. Source-reading without a probe to ground it explores code paths without converging; the probe suite is the executable spec for "understood".
Liberal probes, attic-curate after. Make probes liberally; curate at the end of Stage A. Real-consumer extraction is non-negotiable — it surfaces classes synthetic probes never imagine.
Verified-vs-hypothesized accountability prevents drift. Mark every mechanism claim; the table answers "do we actually know?" honestly.
Tools as needed, not upfront. Add the one tool blocking progress; list it in Tool gaps; revert nice-to-haves.
Claim the issue before the directory — never number a plan by scanning local dirs. The issue number is the identity; file it first and name the directory after it. Scanning the local tree for "the next free number" misses unmerged plans on sibling branches (a 15-debugger/ on a feature branch invisible from main), so you mint a duplicate that collides on merge — and GitHub issue numbers are immutable, so it cannot be cleanly reassigned afterward.

Filing bugs found during plan work

A tracker issue is a claim about the mainline. So, working inside a plan: file a problem only when it reproduces outside the plan — already on the mainline. A pre-existing mainline bug you stumble on during plan work gets filed (and cross-linked to the plan); a breakage the plan's own in-progress work caused is branch-internal — it lives in the plan's docs and is fixed on the branch, never filed. Investigation plans are the strongest case: the probes + cluster docs already document every shape, so a separate issue would just double-document it.

Full policy in the master instructions (loft: CLAUDE.md § Bug-filing policy).

Bind to your tree

Everything above is tree-agnostic. This section is the only loft-specific part — repoint these to port the skill to another tree.

Generic term	loft binding
Plan tracker / issue id	`loft-lang/plans` issues; id = `@PLN<N>`; next free: `gh issue list -R loft-lang/plans --state all --limit 1`
Plan directory (new model)	`doc/claude/plans/<N>-<slug>/` (core language) · `doc/claude/lib_plans/<N>-<slug>/` (libraries) — flat
Legacy layout (mid-migration — most existing plans)	`doc/claude/plans/{future,finished,deferred}/<N>-<slug>/` with rows in `doc/claude/ROADMAP.md` (still maintained — being retired in favor of the tracker). Plans 51 (finished/) and 54 (future/) live here; new plans use the flat/tracker model above.
Standard plan template	`doc/claude/plans/_TEMPLATE.md`
Investigation template	`doc/claude/plans/_INVESTIGATION_TEMPLATE.md` — canonical example `plans/finished/51-hidden-buffer-aliasing/` (5 clusters, 39 probes)
Close / defer procedure (full)	`doc/claude/plans/_LIFECYCLE.md`
Docs-vs-plans rule, three workflows, lifecycle, value categories	`doc/claude/plans/README.md`
Reference-doc `## Open work` homes	`NATIVE.md` / `PERFORMANCE.md` / `PACKAGES.md` / `QUALITY.md`
Value-category labels	`S/R/G/F/U/C/Q/N` (issue labels — definitions in `plans/README.md § Value categories`)
Drift checker	`scripts/check_doc_drift.sh`
Incoming-link grep (close/promote)	`grep -rn "plans/<NN>-<slug>" CLAUDE.md doc/claude/ --include="*.md"`
Investigation regression suite	`tests/scripts/NN-<slug>.loft`
Probe gates (leak / exit)	`LOFT_STORES=warn`; the `loft_suite` leak gate; both backends = interpreter + native
Execution modes to verify across	`--interpret` and `--native`
Canonical examples	partial defer: plan-28 / plan-12 · create-and-move close: `31-html-export → HTML_EXPORT.md` · trim-only close: `04-slot-assignment-redesign → SLOTS.md`
Branch / commit / bug-filing policy	`CLAUDE.md`
Full dev procedures (rebase, commit hygiene)	`doc/claude/DEVELOPMENT.md`