Decide what to say. Then decide how to write it


Minto is a Claude Code and Codex skill that turns notes and drafts into a clear argument: it identifies the reader’s question, surfaces the main answer, groups supporting points and shows where evidence is missing.

The main point comes first The reader sees where you are going
Arguments grouped by kind Reasons with reasons, steps with steps
Missing evidence is marked Unsupported claims stay visibly separate

Before / after: how Minto changes the structure


Minto moves the decision to the top and organises the evidence around it. Compare the original email with the result, then see how the skill checks logic, draws the structure and combines sources.

A message about the defect backlog

The same facts are organised around the decision instead of placed before it.

Source · four paragraphs of evidence

Hi Alex,

I am increasingly worried about the defect backlog. The attached chart compares new defects with the number we close and projects the backlog one year ahead.

Customer Experience and Research also report that defects are contributing to churn. A second chart places churn beside defect volume, and an interview illustrates the same issue.

There are signs of brand damage too: user posts, a complaint from a major account, and Sales reports that these concerns are making deals harder to close.

I recommend using the next sprint to reduce the defect backlog. Let me know what you think.

Alex reads four paragraphs of evidence before finding out what he is being asked to decide.

With Minto · write · scores: structure 8/8, quality 10/10

Hi Alex,

Let’s spend the next sprint on the defect backlog — we’re losing customers and deals to it faster than we’re closing it.

  • The backlog will keep growing on its own. We’re opening defects faster than we close them; the attached chart projects where that leaves us in a year if we don’t intervene now.
  • Defects are costing us customers. Customer Experience and Research tie them directly to churn — the second chart puts churn alongside defect volume, and the attached interview shows the mechanism up close.
  • They’re now costing us new business too. User posts and a complaint from one of our largest accounts have made this visible outside the product, and Sales reports it’s slowing close rates.

Can you confirm the sprint focus so we can shape the backlog before planning?

The skill moved the ask to the top and organised the evidence around it. The scores did not change: the same model’s control run also scored 8/8 and 10/10, and both texts contained one unsupported fact. This example shows the restructuring; the overall effect appears across the 172 judgements below.

A short note to a colleague

The short note stays short while the conclusion moves to the top.

Source · a list of constraints

Collins cannot attend at 3:00. Johnson can meet later today or tomorrow, but not before 10:30. Clifford will not return from Frankfurt until late tomorrow. The conference room is occupied tomorrow but available on Thursday. Thursday at 11:00 appears possible. Is that acceptable?

The reader assembles the conclusion himself, out of five constraints.

With Minto

Could we move today’s meeting to Thursday at 11:00?

Collins and Clifford can’t make 3:00 today — Clifford is in Frankfurt until late tomorrow — and Johnson is only free from 10:30 tomorrow, when the conference room is already booked. Thursday at 11:00 works for all three, and the room is free.

Let me know if that suits you and I’ll send the invitations.

A memorandum in six sections

The diagnosis uses the setout without rewriting it; unknown body content stays outside the assessment.

Source · the setout

“This memorandum describes how project teams identify and deliver significant profit improvements. It is organized in six sections:”

  1. Background
  2. Principles of the project-team approach
  3. Definition of project work
  4. Organization of the program
  5. Distinctive benefits and results
  6. Conditions required for success

Six topics of six different kinds. What the reader must decide is never stated.

With Minto · audit 2/10 structure score
critical

No answer anywhere in the visible text — the document has a subject, not a controlling idea.

critical

One level mixes context, a definition, process, benefits and conditions. It is not one logical list.

medium

“Benefits and results” is the answer to the reader’s question, buried as the fifth of six items.

medium

Principles, definition and organization overlap: all three describe what the approach is.

The three changes with the greatest effect

State the answer in the first sentence. Promote “benefits and results” to the first level. Rewrite the six topic labels as three or four same-kind assertions.

A memo on composing cost

The argument structure becomes visible separately from the prose.

Source · a page of prose

“Composition represents about 40% of hardback cost and 50–55% of paperback cost. TTW does not know whether the cost is excessive, but customers regard it as uncompetitive on simple jobs. Our preliminary work indicates that TTW can cut composing cost substantially…”

Then two all-caps sections and ten facts inside paragraphs: a 20–50% productivity gap, identical checks for every title, below-market pay, two compositors just gone, overtime more than 50% over budget.

With Minto · viz

TTW can cut composing cost substantially

Remove unnecessary process stages

  • Productivity trails the benchmark by 20–50%
  • Every title gets the same checks regardless of complexity
  • Test next week: fewer checks on simple jobs
  • Possible saving up to 10% of composing cost

Raise wages to the local market level

  • Pay is below nearby printers
  • Two compositors have just left
  • Overtime exceeds budget by more than 50%
  • A new union claim may force higher pay anyway

The skill marks its own gap: the saving is quantified for the first branch only — “substantially” is unmeasured for the second.

Three compliance sources

Several sources become one decision instead of a sequence of summaries.

Sources · 3 files

gate-review.xlsx

Seven questions answered by the bank’s compliance office. One is declined outright: “cannot be answered without an architecture diagram”.

sync-0709.md

The package list “is being prepared”. A new contact is named. The remark about two engineers in Lisbon draws no response.

Risk register

Three open rows: the price cannot be fixed, the list is 98 days late, the diagram has not been sent.

With Minto · digest

The one thing to do today: send the architecture diagram.

It has been drafted since 11 August and it is the only compliance blocker Fernway owns — the 9 July sync agreed it must arrive before the next sync so the storage, encryption and logging questions can get a real answer. Everything else is waiting on Aldergate, and two items have been waiting long enough to need escalating by name.

Already answered, treat as settled

A security review before production is mandatory. An internal hardening framework exists, though conforming to it does not remove the review. An open-source policy exists. Work from Ireland is acceptable.

Waiting on Aldergate, and overdue

The approved-package list was promised on 12 May and again on 9 July, both times with no date — 98 days open. The named information-security owner has never joined a joint meeting. Ask them for the list directly, or for a date in writing.

Two answers they will not give in advance — and they hit the price

Whether a penetration test is required is decided by the reviewer during the review. Dependency approval has no standard turnaround. So the build-phase price cannot be a single number: either carve both out as separately-priced items, or hold the price until the review is scheduled.

One open risk to raise ourselves

On 9 July it was said that two engineers are based in Lisbon, and nobody from the bank responded. Silence in a meeting is not approval — put Lisbon in writing alongside the diagram, before those engineers touch bank data.

Tested in 172 blind judgements


Minto’s effect is measured on real texts, showing where the structure improved and where the skill still needs work.

Five jobs for Minto


Choose the result you need and invoke the command beside it: intent, audit, write, digest or viz.

The flow the skill checks in every document

  1. Situation what the reader already agrees with
  2. Complication what changed or gets in the way
  3. Question what the reader has to decide
  4. Answer what they should do or understand

How the benchmark works


The method, all six measures and model-level scores.

How we tested it

11 different jobs

Emails, memos, an audit, a digest and a pasteable list — from a two-line note to an eight-page review.

Structure against instruction

Minto fixes the reader’s question, the answer and its supports; the control prompt only asks for the Pyramid Principle.

One rubric for every text

Shuffled outputs are scored against the same criteria without revealing the model or preparation method.

Every figure is reproducible

A script calculates every figure on this page from the published verdicts.

What changed in the texts

without the skill with Minto the difference
  • The reader’s question is written down

    13%
    35%

    +22 percentage points

  • First-level points are all the same kind

    54%
    66%

    +12 percentage points

  • Structure, share of the judge’s 8 points

    59%
    66%

    +7 percentage points

  • Writing quality, share of the judge’s 10 points

    72%
    79%

    +7 percentage points

  • No source material lost without saying so

    62%
    65%

    +3 percentage points

  • Answer in the first sentence

    got worse
    84%
    74%

    −10 percentage points

View the exact scores by model
Share of the judge’s maximum, control then with Minto. Eleven documents per model (nine for qwen3.8-27b — two timed out in both arms).
Model Structure /8 Quality /10
Claude Opus 5 63%84% 74%90%
Claude Sonnet 5 66%61% 78%80%
Claude Haiku 4.5 53%63% 67%76%
Codex gpt-5.6-terra 60%67% 75%82%
gpt-oss-120b 50%52% 53%64%
qwen3.6-35b-a3b 58%67% 76%79%
qwen3.6-fp8 64%69% 75%80%
qwen3.8-27b 63%63% 79%81%

In raw points, pooled across all eight models: structure 4.76 out of 8 becomes 5.27, quality 7.17 out of 10 becomes 7.90. Applying the rubric’s own penalty for invented or dropped facts gives 6.34 against 7.08.

Read the full methodology and limitations

Install Minto


Add the marketplace, install the plugin, then start a new session.

Claude Code

CLI & desktop app
  1. /plugin marketplace add welltraum/minto
  2. /plugin install minto@minto

Then invoke /minto:minto — or just describe the writing problem and let Claude route it.

Codex

CLI & desktop app
  1. codex plugin marketplace add welltraum/minto
  2. codex plugin add minto@minto

Then invoke $minto, or describe the problem in words.

Test it as soon as it is installed

Paste the text and say: “Check this proposal against the Pyramid Principle. Name the structural problems first. Do not rewrite the text.”