Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when repository docs, README, AGENTS.md, rules, plans, chronicles, or ATLAS are stale or current work changed their lifecycle; archives task history, with --clean for corpus-wide cleanup.
.claude/skills/hashgraph-online-align-docs/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-16 | ✗→✓ | ▲ Improved | 26% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -16% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 15% | 0% |
| case-13 | ✗→✓ | ▲ Improved | -8% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 1% | 0% |
Make the repository's existing knowledge easy for humans and agents to find and trust. Run it inside the development-skills workflow, after a conversation, or standalone. Inside an active workflow, continue that workflow instead of starting another. When prior context exists, include lasting knowledge gathered there; otherwise follow the default steps.
Always leave this shared entry layer:
textREADME.md # evidence-based human entry point AGENTS.md # under 70 lines; critical facts and rules index CLAUDE.md # exactly @AGENTS.md .agents/rules/ # scoped team/project rules; .gitkeep if empty .claude/rules -> ../.agents/rules docs/ATLAS.md # curated routing map of the repo's knowledge
Resolve the git root. Inspect the entry layer, manifests, their referenced paths, agent memory, and documents created, changed, or invalidated by the current task. Without prior task context, inspect the entry layer and its references only. Read enough reliable evidence to verify claims. Exclude dependencies, caches, build output, and vendored content.
Report missing, stale, contradictory, duplicated, orphaned, or oversized knowledge before editing. A normal run may make safe in-scope fixes but never merges or deletes first-party documents.
When prior task context proves that the current task changed a document's lifecycle, also move the resulting history with git mv:
docs/plans/archive/;docs/chronicles/archive/, preserving decision prose and lifecycle links;archive/ beside its active collection.Before archiving, move durable agent instructions to the right .agents/rules/ file. Keep current owners and uncertain documents active, and make archive metadata match the path. This bounded normal run does not scan unrelated history for archive candidates.
With --clean, follow clean-mode.md; it applies these lifecycle rules to the full corpus, and its proposal gate replaces normal apply behavior.
Give each fact one owner: brief always-read facts in AGENTS.md, scoped detail in .agents/rules/, execution in plans, decisions in chronicles, and other existing knowledge in its clearest current home. Prefer links over copies and preserve useful project-specific structure.
Resolve claims by what they describe: code, tests, and configuration for implementation; telemetry or live data for runtime state; approved contracts for promises; chronicles for rationale; and executable commands or CI for procedures. Report conflicts between authorities instead of guessing.
Apply the documentation contract within its stated scope.
Build docs/ATLAS.md without frontmatter, as a routing map, not an inventory: the filesystem and git already enumerate files. Route every recurring question or area to its authoritative home with a link and one when-to-open line. Be exhaustive only inside small curated scopes (entry points, rules, reference shelf); give self-indexing directories such as plans and chronicles one routing entry that explains their naming convention. Exclude fixtures, implementation-support Markdown, and ATLAS itself. No lifecycle columns — validity and work state live in each document's frontmatter.
An existing AGENTS.md is authoritative: report issues, never overwrite it or prepend the template. Create AGENTS.md from agents-template.md only when the file is missing or the user explicitly asks to migrate. Delete a leftover align-docs:principles-customized marker. A rule owns its topic; AGENTS.md only links to it.
Keep personal machine facts in .claude/CLAUDE.md or the user's global Codex AGENTS file. Ensure .gitignore contains .claude/CLAUDE.md and AGENTS.override.md.
Capture only lasting session facts that are absent from disk. A run without prior session context reports CAPTURE: NONE. Move useful memory to its owner; leave empty or generic memory for clean mode.
Finish by checking the AGENTS line budget, rule scopes, symlink, ignored personal files, README/manifests against disk, ATLAS routing (every area reachable, no excluded content), required frontmatter on changed docs, and Markdown links. Apply the reduce-gate to every changed instruction file.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 9,475 | 9,904 | +5% | 1 | 1 | 0% | 278 | 1,296 | +366% | 0 | 0 | — |
case-16 | fail→pass | 16,770 | 11,985 | -29% | 1 | 1 | 0% | 2,421 | 3,049 | +26% | 0 | 0 | — |
case-22 | fail→fail | 9,603 | 16,813 | +75% | 1 | 1 | 0% | 267 | 1,569 | +488% | 0 | 0 | — |
case-02 | fail→fail | 3,986 | 12,651 | +217% | 1 | 1 | 0% | 321 | 1,576 | +391% | 0 | 0 | — |
case-03 | fail→fail | 9,168 | 15,135 | +65% | 1 | 1 | 0% | 227 | 1,270 | +459% | 0 | 0 | — |
case-04 | pass→fail | 12,567 | 12,491 | -1% | 1 | 1 | 0% | 1,343 | 1,523 | +13% | 0 | 0 | — |
case-05 | fail→fail | 16,346 | 6,853 | -58% | 1 | 1 | 0% | 1,660 | 1,598 | -4% | 0 | 0 | — |
case-06 | pass→pass | 13,627 | 11,862 | -13% | 1 | 1 | 0% | 1,409 | 1,642 | +17% | 0 | 0 | — |
case-07 | pass→pass | 10,180 | 9,498 | -7% | 1 | 1 | 0% | 1,685 | 1,735 | +3% | 0 | 0 | — |
case-08 | fail→pass | 12,843 | 2,795 | -78% | 1 | 1 | 0% | 1,735 | 1,450 | -16% | 0 | 0 | — |
case-09 | fail→pass | 7,649 | 8,015 | +5% | 1 | 1 | 0% | 1,350 | 1,557 | +15% | 0 | 0 | — |
case-10 | pass→pass | 20,173 | 14,240 | -29% | 1 | 1 | 0% | 2,635 | 2,728 | +4% | 0 | 0 | — |
case-11 | pass→pass | 13,253 | 6,995 | -47% | 1 | 1 | 0% | 2,187 | 2,040 | -7% | 0 | 0 | — |
case-12 | pass→pass | 8,554 | 4,483 | -48% | 1 | 1 | 0% | 1,390 | 1,835 | +32% | 0 | 0 | — |
case-13 | fail→pass | 16,705 | 9,796 | -41% | 1 | 1 | 0% | 1,928 | 1,776 | -8% | 0 | 0 | — |
case-14 | fail→pass | 9,483 | 1,777 | -81% | 1 | 1 | 0% | 1,255 | 1,266 | +1% | 0 | 0 | — |
case-15 | pass→pass | 10,866 | 11,173 | +3% | 1 | 1 | 0% | 1,822 | 2,042 | +12% | 0 | 0 | — |
case-17 | fail→fail | 15,569 | 10,800 | -31% | 1 | 1 | 0% | 2,507 | 1,355 | -46% | 0 | 0 | — |
case-18 | pass→pass | 10,498 | 10,571 | +1% | 1 | 1 | 0% | 1,631 | 1,876 | +15% | 0 | 0 | — |
case-19 | fail→pass | 10,739 | 12,253 | +14% | 1 | 1 | 0% | 977 | 2,281 | +133% | 0 | 0 | — |
case-20 | pass→fail | 43,888 | 15,782 | -64% | 1 | 1 | 0% | 8,230 | 1,265 | -85% | 0 | 0 | — |
case-21 | pass→fail | 15,940 | 5,351 | -66% | 1 | 1 | 0% | 2,224 | 1,299 | -42% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 13 counted toward the lift figure. The other 9 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +14 percentage points is the difference between those two pass rates over the 13 comparable cases. 3 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.