Skip to main content
QFlow.aiQFlow.ai blog

QFlow Recipes: Consistent, Maintainable, and Efficient AI Analysis

When multiple people use MCP to access systems of record, each can build their own analysis. Without a shared method, three problems follow:

  • Consistency breaks. Different definitions, filters, and joins produce competing answers to the same question.

  • Maintenance gets harder. Rules scatter across separate workflows. Change a definition, and someone must find and update every copy.

  • Costs pile up. People spend tokens rebuilding calculations others have already solved. Next month, they may do it again.

A shared skill can already address much of this. Finance and RevOps teams can publish skills that call tested SQL or Python, apply agreed definitions, and return compact results.

An analytical service behind your skills

QFlow supplies the analytical service those skills can use. Recipes capture business definitions and transformations as reusable operations. QFlow runs them against business data, checks the outputs, and keeps the execution results available for review.

The value is in maintaining the workflow around the calculation: connected data, saved operations, scheduled runs, checks, and execution history. QFlow packages that work in a shared product. Your team still owns the business definitions and their approval.

A board-reporting skill, for example, can use a Recipe’s bookings result to build the team’s deck. QFlow’s Organization Skills package templates, instructions, and scripts for that work in Claude or Codex.

A good harness can reuse the work

In The Harness Margin Opportunity, Tomasz Tunguz argues that the software around a model can set the cost of its answers. He cites research where changing the harness cut the cost per resolved coding task by 71%, using the same model. The comparisons did not establish a statistically significant quality difference.1

The harness gathers context, calls tools, and keeps work moving. Its efficiency also depends on what those tools do.

A capable general-purpose harness becomes more efficient when it can call proven, business-specific operations. Shared skills can direct it to those operations. QFlow supplies and runs the analytical workflows.

Same harness, fewer tokens

In one internal test, we compared raw-data analysis with a prepared QFlow Recipe result for an 18-month opportunity win-rate task. Both final executions used the same model through Codex. The Recipe already existed. The raw-data workflow included preparation, debugging, and an unsuccessful attempt.

Customer-model tokens in this test:

Comparison Raw-data workflow Prepared Recipe result
Including observed baseline preparation 14,123,603 tokens 15,821 tokens
Final execution only 76,613 tokens 15,821 tokens

Both final answers reproduced the expected counts and win rate. In that final execution, consuming the Recipe result used 79.35% fewer customer-model tokens.2

These are token counts, not a total-cost comparison. The test measures the benefit of a prepared result. It does not compare Recipes with a skill running an equivalent saved calculation.

Make the method shared and inspectable

Consider a recurring request:

Explain how new-business bookings changed last quarter, by region and product.

1. Agree on the business rules

A few choices determine the answer: what counts as new business, which fiscal dates apply, how products share a deal, and which region owns it.

Teams must agree on the business definition and recipe revision. Using the same recipe and configuration then gives them the same calculation.

2. Execute and inspect the operations

AI writes structured commands. Deterministic commands filter records, join datasets, group results, calculate comparisons, and so on.3

QFlow saves proven commands and replays them against fresh data. We call them recipes. Here, “proven” means the saved command has run successfully during verification. It does not certify that the business definition is right for every future use.

On the normal transformation path, a step without a valid saved command stops and requires verification. A reviewer still needs to inspect choices such as joins, exclusions, and fiscal periods.

3. Maintain the calculation where it runs

Suppose Finance moves a report from calendar quarters to fiscal quarters. The team updates the Recipe’s date rule, checks deals around each cutoff, and reconciles the new totals. The reporting skill then uses the revised Recipe result.

That keeps the calculation in one maintained workflow. Skills that consume its result need no separate copy of the date rule. A team can build this discipline with its own code; QFlow supplies the shared execution and review workflow. The team still decides when a revision is ready to use.

Keep review focused on the evidence

An incorrect method can be repeated perfectly. QFlow checks for known failure patterns and reviews actual outputs. Large datasets stay in its execution environment. Its correctness reviewer asks structured questions and gets focused answers.

Recorded gaps prevent a clean review verdict. The analyst must resolve gaps that matter to the report. A clean review still cannot establish that every business assumption is correct.

Anthropic describes the same efficiency pattern: filter, join, and aggregate data outside the model’s context, then return the relevant results.4

The roles stay clear. As OpenAI describes it, the harness runs the model and tool loop; applications supply business context, rules, and tools.5 QFlow supplies analytical operations and checks. The team’s skills guide how Claude or Codex uses the results.

Build the improvement loop

QFlow has evaluations for recipe generation and repair, regression checks, and model-usage tracking. These are the foundation for Tunguz’s improvement factory. Connecting them into a systematic loop is the next step:

  • Turn observed failures into trusted test cases.
  • Test changes to generation, execution, and review.
  • Keep changes that improve results against those tests.

The measure should be the full cost of a correct, useful report: preparation, execution, checking, repair, and maintenance. The token comparison above measures only part of that work.

For the team publishing skills today, the practical value starts sooner. Put the shared calculation in QFlow. Have the skills use its results. Maintain the method as the business changes, and spend the next reporting cycle explaining what happened.

  1. Tomasz Tunguz, “The Harness Margin Opportunity”, September 17, 2026. The cited reduction is from $1.540 to $0.441 per resolved task for GPT-5.6 Sol across two coding harnesses. As Tunguz’s footnote explains, the study’s sample does not establish that quality is identical. These are coding-benchmark results, not measurements of QFlow’s savings. ↩︎

  2. QFlow internal win-rate benchmark, July 2026. Both final executions used GPT-5.6 Sol through Codex with high reasoning and the same business-rule contract. The first-run baseline totaled 14,123,603 tokens: 13,667,386 allocated to task-specific planning, data work, and debugging in the benchmark session; 379,604 from a separate unsuccessful attempt; and 76,613 from the successful final execution. The model consuming the prepared Recipe result used 15,821 tokens, yielding a 99.89% first-run reduction under that accounting. Recipe creation used zero model tokens. Excluding preparation and the unsuccessful attempt gives the 79.35% execution-only reduction. Both final outputs reproduced the expected counts and win rate. This was one task and one repetition, with exact numerical validation rather than a broader blinded quality study. Counts include cached input tokens; token reduction is not the same as cost reduction. QFlow-internal execution tokens were unavailable, so these results describe customer-model usage and do not establish an end-to-end or platform-wide savings rate. ↩︎

  3. “Deterministic” refers to the computed transformations: the same operations applied to the same inputs and configuration produce the same computed results. Business assumptions still require validation, and AI-generated explanations can vary. ↩︎

  4. Adam Jones and Conor Kelly, Anthropic, “Code execution with MCP: Building more efficient agents”, November 4, 2025. The article describes processing large datasets outside the model’s context and returning selected results, including examples involving filtering, aggregation, and joins. It provides an independent explanation of this architectural principle; its reported savings are specific to its examples. ↩︎

  5. Nicolas Bonamy and Derrick Choi, OpenAI, “Codex as a platform: build on the open agent harness”, August 19, 2026. The article distinguishes the Codex harness’s model and tool loop, conversation state, and execution controls from application-owned business context, rules, and tools. This supports the architectural distinction; it does not describe or evaluate QFlow’s implementation. ↩︎

2026-11-09
QFlow.ai Engineering