Kimi Skills Guide: Create, Use, Limits & Test Status

Kimi Skills are reusable instruction and knowledge packages for recurring work. Instead of restating the same format, rules and quality checks in every chat, a Skill can package that workflow and be invoked when it is relevant. Kimi’s documentation contrasts this with direct prompting: a prompt is suitable for one-off work, while a Skill is intended to improve consistency across repeated tasks.

That product description is not proof that a Skill produces better answers. KI AI Team prepared a controlled five-task comparison, but it has not run. A preliminary ordinary-prompt attempt returned a downloadable SKILL.md file without registering the candidate. We then used the real embedded /skill-creator flow with both the original Markdown and a structurally validated DOCX. On August 4, 2026 at 19:11:16 PDT, an immediate high-demand message stopped that attempt before creation began. Without a registered Skill and visible invocation evidence, there is no valid Skill condition, no paired comparison and no performance result.

Independent guide: Kimi AI Guide is not affiliated with Moonshot AI. Official Skills documentation was checked on August 4, 2026. This page reports two incomplete setup attempts and a preregistered protocol. It does not claim that Kimi Skills improve accuracy, structure, speed or reliability.

Kimi Skills at a glance

Question	Current documented answer	What we independently established
What is a Skill?	A reusable knowledge package loaded when relevant	We created an original candidate specification; no registered Custom Skill was confirmed
Where are Skills supported?	Agent Mode and Kimi Claw	Not independently compared here
How can a Skill be invoked?	Slash command, the plus menu or automatic selection by Kimi	Manual invocation was not available for our unregistered candidate
How can a Custom Skill be created?	Office Document to Skill or `/skill-creator`	The real `/skill-creator` was reached, but high demand stopped creation before execution
Which document inputs are listed?	DOCX, XLSX, PDF, PPTX and document screenshots	Official documentation only
Upload limit	Up to three files per upload, each no larger than 100 MB	Not independently stress-tested
Daily document-conversion limit	Up to 10 Office Document to Skill conversions per day, including retries	Official account rule; can change
Concurrent creation limit	Up to three creation tasks under the FAQ’s task-count rule	Official account rule; not stress-tested
Does creation consume quota?	Successful creation uses gifted searches first, then official membership credits; a failed creation does not deduct tokens	Official rule; account balance and availability vary
Is Custom Skill creation guaranteed on the free plan?	The public pages do not say that every free account has access	Unverified; check the signed-in account before testing
Did our comparison run?	No	Ordinary tasks 0/5; Skill tasks 0/5; NOT RUN, no scores and no performance claim

The current product information comes from Kimi’s What are Skills?, Using Skills in Agent Mode and Custom Skills creation FAQ.

What a Kimi Skill is – and what it is not

A Skill is useful when the same operating rules must survive across repeated tasks: a house style, a fixed report structure, a decision procedure, a data-cleaning rule or a document workflow. Kimi says relevant Skills can be loaded when a task matches them, reducing the need to repeat instructions.

A Skill is not:

proof that its instructions were followed;
a substitute for source verification or human review;
permission to access data the user is not authorized to share;
a guarantee that automatic routing selected the intended package; or
evidence that an ordinary prompt could not produce the same result.

Those distinctions matter in a comparison. If Kimi creates a Markdown file that describes a Skill but the product does not register it, the file is only an artifact. A valid Skill condition needs visible evidence that the exact candidate was installed and invoked for that output.

How to invoke a Skill in Kimi

Kimi documents three routes in Agent Mode:

type / and select a Skill;
use the plus menu; or
allow Kimi to decide that a relevant Skill should be used.

For a controlled test, manual invocation is the safer route. Automatic selection introduces a hidden variable: the evaluator may not know whether the Skill was loaded, which Skill was chosen or whether an ordinary-looking response actually used it.

Kimi currently says Skills are supported in Agent Mode and Kimi Claw. Product surfaces, rollout and account visibility can change, so record the exact route, account plan, region, mode and date whenever a Skill is tested.

How to create a Custom Skill

Kimi documents two creation workflows:

Office Document to Skill: upload source documents for Kimi to convert into a reusable package.
/skill-creator: describe the intended workflow and use the creator route to build it.

The Office Document workflow lists DOCX, XLSX, PDF, PPTX and document screenshots. A single upload can contain up to three files, each no larger than 100 MB. Kimi recommends that a good Skill do one thing well; a narrow workflow is easier to describe, invoke and evaluate than a package that attempts to govern unrelated tasks.

Current creation and naming limits

According to the official creation FAQ checked August 4, 2026:

a maximum of three creation tasks can occupy the relevant creation-task count at once; the FAQ includes Creating and Failed states in that rule;
Office Document to Skill is limited to 10 conversions per day, including retries;
the name can contain no more than 25 characters;
names use lowercase English letters, hyphens and underscores;
a name cannot duplicate an existing Skill name;
a successful creation consumes tokens;
tokens are taken from gifted searches first and then official membership credits;
a failed creation does not deduct tokens; and
the Create Skill control can be unavailable when the account has exhausted its credits.

The public documentation does not state that Custom Skill creation is included for every free-plan account. The reference to gifted searches describes a possible credit source, not a universal free-tier entitlement. Verify that the creation control and sufficient credits are present in the signed-in account before starting a test.

What happened in our two setup attempts

Attempt 1: an ordinary Agent task returned a file

KI AI Team first supplied the candidate workflow through an ordinary prompt. At 18:54:46 PDT, Kimi completed an ordinary Agent task and returned a downloadable evidence-brief-sop/SKILL.md, but the candidate did not appear in the account’s Custom Skills library. That is a useful intermediate artifact, not a registered or invoked Custom Skill.

Attempt 2: the real embedded `/skill-creator` was blocked

We then opened the actual Custom Skills creation flow and used Kimi’s embedded /skill-creator token. Two retained versions of the same candidate were attached:

evidence-brief-sop.md, shown as 4.05 KB; and
evidence-brief-sop.docx, shown as 38.56 KB.

At 19:11:16 PDT on August 4, 2026, Send was pressed. Kimi displayed a high-demand message immediately, and Skill creation did not begin. The candidate remained absent from Custom Skills, and no manual invocation became available.

Required setup event	Observed?	Interpretation
Downloadable `SKILL.md` returned	Yes	A file was generated
Real embedded `/skill-creator` opened	Yes	The correct creation surface was reached
Markdown attached	Yes, 4.05 KB	Original candidate source was supplied
DOCX attached	Yes, 38.56 KB	Structurally validated conversion was supplied
Send result	Immediate high-demand message	Creation was blocked before execution
Candidate shown in Custom Skills	No	Registration did not occur
Candidate manually selectable	No	Invocation could not be verified
Visible Skill invocation on a task	No	No Skill-condition output exists
Ordinary benchmark tasks	0/5	No baseline task output exists
Skill-condition benchmark tasks	0/5	No Skill task output exists
Ordinary-versus-Skill paired runs	0/5	Comparison remains NOT RUN
Performance score	None	No result can be calculated

Both events were pre-comparison setup attempts, not Condition A or Condition B of the five-task benchmark. Treating the returned file as a successful Skill run would confuse content generation with product registration; treating a pre-execution high-demand blocker as a failed performance result would confuse access with output quality.

Our candidate Skill: evidence-brief-sop

The original candidate is named evidence-brief-sop. It is designed to turn a user-supplied source packet into a concise decision brief while keeping calculations, contradictions, unknowns and citations visible.

The complete candidate SKILL.md and its DOCX conversion are included in the downloadable Kimi Skills test pack. Its core rules are:

use only the supplied source packet unless separate browsing is authorized;
cite factual claims with the packet’s [S#] labels;
preserve names, dates, units, percentages, currencies and status terms;
never silently resolve a contradiction;
write Not provided instead of inventing a missing owner, date or criterion;
show reproducible arithmetic;
apply an explicit decision rule when one exists;
never describe pending work as complete; and
return seven required sections in a fixed order.

The seven output sections are Decision, Executive brief, Evidence ledger, Calculations or changes, Unknowns and contradictions, Actions, and Verification note.

The file is a candidate specification, not a validated Kimi Custom Skill. Its syntax, registration and behavior remain unverified until the signed-in product accepts it and exposes visible invocation evidence.

DOCX fixture verification

The attached DOCX has SHA-256:

DEBB1D7F28E37FD77E9175D2816929942EE3CF5D2A48164AF3F0EE0890FC19F1

Its structural QA record, included in the test pack, passed 13/13 predefined checks, including package validity, semantic-content preservation, headings, front matter, numbered rules, required-section order, editable tables, page setup and absence of comments or tracked changes.

Visual QA remains NOT RUN because a rendering tool was unavailable in the test environment. The 13/13 result supports structural integrity only; it is not a claim that every rendered page is visually correct.

The five-task Skill-versus-prompt fixture

The benchmark fixture is KS-SKILLS-COMPARE-V1. Each task is run under two conditions:

Ordinary: a new chat with no candidate Skill installed, selected or automatically invoked.
Skill: a separate new chat using the same account, model, mode, plan and region, with evidence-brief-sop manually invoked.

Both conditions receive the exact same task prompt. No web search, extra files, correction follow-up or hidden human edits are allowed. Conditions should run close together, their order should alternate across tasks, and two replicates per condition are recommended. The complete task fixture and preregistered ground truth are available in the Kimi Skills test pack.

Task 1 – Incident closure decision

Prepare a concise decision brief for the operations lead using only the source
packet below. Decide whether incident I-17 can be closed at 10:25 UTC. Do not
browse or add facts. If evidence is missing or contradictory, say so.

[S1]
Incident I-17 began on 2026-08-04 at 09:15 UTC. The payments API error rate
rose from 1.2% to 8.7%. A rollback completed at 09:42 UTC. The error rate was
1.3% at 09:55 UTC and remained between 1.1% and 1.4% through 10:25 UTC.

[S2]
The closure rule allows the incident to close after the error rate remains at
or below 1.5% for 30 continuous minutes. The incident owner is Maya Chen. The
root-cause analysis is due on 2026-08-12.

Ground truth: Proceed because 09:55 to 10:25 is exactly 30 continuous minutes at or below the threshold. The response must not claim that the root cause is known or the analysis is complete.

Task 2 – Vendor shortlist with a conflict

Prepare a concise decision brief for the procurement committee using only the
source packet below. Decide which vendor, if any, qualifies for the shortlist.
Do not browse or add facts. If evidence is missing or contradictory, say so.

[S1]
Mandatory shortlist criteria: single sign-on must be included; deployment must
complete within 50 calendar days of signing; annual software cost must not
exceed USD 55,000.

[S2]
Alpha proposal: USD 48,000 per year, 99.9% stated uptime, single sign-on not
included, proposed implementation time 30 calendar days.

[S3]
Alpha security review: single sign-on is unavailable in the proposed package.
Security estimates that implementation cannot complete in fewer than 45
calendar days.

[S4]
Beta proposal: USD 54,000 per year, 99.95% stated uptime, single sign-on
included, proposed implementation time 60 calendar days.

Ground truth: Hold because neither vendor satisfies every mandatory criterion. The Alpha timing conflict must remain visible even though both stated values are within 50 days.

Task 3 – First-year budget calculation

Prepare a concise decision brief for the finance manager using only the source
packet below. Calculate the first-year total and decide whether it fits the
approved cap. Do not browse or add facts. Show the arithmetic.

[S1]
The plan requires 120 seats at USD 28 per seat per month for 12 months.
Training is a one-time USD 4,800 charge. Add a 7% contingency to the subscription
line only. The approved first-year cap is USD 48,000.

Ground truth: subscription 120 x 28 x 12 = USD 40,320; contingency 40,320 x 0.07 = USD 2,822.40; total 40,320 + 2,822.40 + 4,800 = USD 47,942.40. The expected decision is Proceed, with only USD 57.60 of headroom.

Task 4 – Policy change adoption

Prepare a concise decision brief for the accounts-payable lead using only the
source packet below. Identify every operational change required before the new
policy takes effect. Do not browse or add facts.

[S1]
Current policy: invoices above USD 10,000 require Director approval. Review is
due within five business days. Approval records are retained for three years.

[S2]
New policy effective 2026-09-01: invoices above USD 7,500 require both Director
and Finance approval. Review is due within three business days. Approval records
must be retained for five years.

Ground truth: identify the USD 2,500 threshold reduction, added Finance approval, two-business-day shorter review window and two-year longer retention. The expected decision label is No decision because the packet provides required changes but no proceed/hold rule or implementation owner.

Task 5 – Launch go/no-go with missing evidence

Prepare a concise go/no-go decision brief for the launch sponsor using only the
source packet below. Use the status as of 2026-09-08. Do not browse or add facts.
If evidence is missing or pending, say so.

[S1]
Target launch date: 2026-09-18. Legal approval is pending and due 2026-09-10;
owner Priya Nair. Data migration is pending and due 2026-09-12; owner Omar Reed.
Training is scheduled for 2026-09-15, but the owner is not provided. The
cutover rollback threshold is not defined. Security-test status is not provided.

[S2]
Launch rule: Go requires completed legal approval, completed data migration,
a passed security test, a defined rollback threshold and an assigned training
owner. Any failed, pending or unverified mandatory item results in Hold.

Ground truth: Hold. Legal approval and migration are pending, the security result is absent, the rollback threshold is undefined and the training owner is not provided.

The 100-point comparison rubric

Each output is scored independently under the same rubric. A Skill receives no points merely because it was invoked.

Category	Points	Standard
Fact and decision accuracy	30	Required facts and the final decision match the preregistered ground truth
No invention and visible uncertainty	20	Missing, pending and contradictory evidence is not silently filled
Source traceability	15	Factual claims map to supplied `[S#]` labels
SOP structure	15	The seven candidate-Skill sections appear in the required order
Calculation or conflict handling	10	Arithmetic, criteria and contradictions are explicit and reproducible
Clarity and actionability	10	The brief is concise and preserves supplied owners and dates
Total per output	100	Same standard for Ordinary and Skill conditions

The final comparison reports mean, median, minimum, maximum and range for each condition; standard deviation when replicates permit; unsupported-claim count; citation coverage; structural-compliance rate; exact-decision accuracy; and arithmetic, date, owner and status errors.

Even a higher average in this small benchmark would support only the tested candidate, five tasks, account, model, mode, date and run count. It would not prove that Skills are universally better than prompts.

Why the comparison is still NOT RUN

The protocol requires paired outputs and visible proof that the candidate Skill was invoked. At present:

there are zero scoreable Ordinary-condition outputs from the five fixed tasks;
there are zero scoreable Skill-condition outputs;
the candidate is not confirmed as registered;
the actual /skill-creator attempt was blocked before creation by an immediate high-demand message;
no manual invocation evidence exists; and
no category score or aggregate metric can be calculated.

The dated creation run record in the Kimi Skills test pack documents both setup attempts, while its raw-results template remains blank by design. A downloadable SKILL.md cannot be scored as though it were a response produced under the Skill condition, and an access blocker cannot be assigned a quality score.

A valid future run

Before testing, confirm the following in the signed-in product:

Skills are available in the chosen surface.
The candidate name is not already in use.
evidence-brief-sop is registered or installed successfully.
Manual invocation is available.
The Skill condition displays visible invocation evidence.
The Ordinary condition has no Skill selected or automatically invoked.

Then retain the complete raw output and screenshots for every task, condition and replicate. Do not silently exclude a failed invocation or generation error. Record it as a blocker or deviation, because hiding setup failures would bias the comparison toward the successful condition.

Limits of this guide

Official creation, quota and file-limit statements are product documentation, not independently stress-tested limits.
Public documentation does not establish universal free-plan access to Custom Skill creation.
The returned SKILL.md proves file generation only; it does not prove registration or invocation.
The high-demand event describes one dated creation attempt; it does not establish a universal access or plan rule.
The DOCX passed structural QA but did not receive visual QA.
The candidate’s syntax and behavior have not been validated in a registered Kimi Skill.
No paired task completed, so there is no accuracy, consistency, latency or quality result.
The five synthetic decision tasks do not represent every Skill or workflow.
Account, plan, credit, region, rollout and model state can change access and behavior.
A future result must remain scoped to its date and execution conditions.

Practical questions

Are Kimi Skills available on the free plan?

The public Help Center does not guarantee Custom Skill creation for every free account. It says successful creation consumes gifted searches first and then membership credits. Confirm availability and credits in the signed-in account instead of assuming access from that wording.

What files can Office Document to Skill use?

Kimi currently lists DOCX, XLSX, PDF, PPTX and document screenshots, with up to three files per upload and a 100 MB limit per file.

Does a downloaded SKILL.md mean a Custom Skill was created?

Not by itself. The product must also show that the candidate was registered or installed and can be invoked. Our first setup attempt returned a file, and the subsequent real /skill-creator attempt was blocked by high demand before creation; neither established registration.

Are Skills better than ordinary prompts?

We have no result supporting that claim. The controlled comparison remains NOT RUN. Any eventual conclusion must be limited to the exact candidate, tasks and recorded environment.

Sources and update policy

Product facts were checked against Kimi’s What are Skills?, Using Skills in Agent Mode, Custom Skills creation FAQ and Kimi Code Skills documentation on August 4, 2026. We will update the test-status section only when registration and paired runs are backed by retained evidence. See our testing methodology and Sources and Corrections policy.