TECH HAUS

Codex / Claude Code / Kimi Code / repository files

Put TECH HAUS frameworks to work inside coding agents

The premium framework pages explain why certain framings work, how they progress, what they preserve, and where they can fail. Coding-agent files should carry those useful behaviors without naming the source framework. Translate only the measures that improve the real assignment, then verify that the result preserved the user's intent.

Use this when
The assignment depends on repository files, commands, tests, saved decisions, or work that must survive another session.
You leave with
A concise, personalized instruction file that applies useful framing behaviors to a real project without framework labels or copied turns.

First action

Reviewed 26 Jul 2026

Start the work before reading further.

Choose the instruction file supported by your coding tool, then name the objective and the one operating behavior that materially improves the work.

01 / Work surface

Use the browser for a question. Use a coding agent for a body of work.

If the result depends on project files, commands, tests, saved decisions, or several passes, use Codex, Claude Code, Kimi Code, or another repository-aware tool. The framework becomes durable context, and you stop repeating uploads, setup notes, and copy-paste handoffs in every message.

Choose a coding agent

  • YESThe work spans more than one file or requires repository discovery.
  • YESCommands, tests, browser checks, or build output determine whether the result is correct.
  • YESThe framework and project rules should remain available across sessions.
  • YESThe result must be written back into a project, reviewed as a diff, or handed to another agent.

Stay in web chat

  • WEBThe question is isolated and no local files or tools are required.
  • WEBYou need a quick comparison, explanation, or draft that will not become a maintained artifact.
  • WEBYou are not ready to grant a tool access to a working folder.

The operating loop

One objective. One useful behavior. One result you can verify.

01A fixed intent contract

Capture the objective

Write the requested outcome, supplied facts, exclusions, deliverable, and success checks in the current request or an optional task brief. Keep the user's wording where it carries meaning.

02A name-free operating pattern

Extract the useful behavior

Study the source guidance and pull out only what helps this assignment: the opening condition, progression, decision gates, evidence standard, cautions, and recovery behavior. Leave the framework name behind.

03A form that fits the tool

Map the progression

Translate the useful functions into agentic stages, checkpoints, or project rules. Preserve how context and decisions build without copying the example conversation as a script or exposing its label.

04Work grounded in real project state

Run from the repository

Put the personalized guidance in the instruction surface your tool supports. Let the agent inspect the workspace and use the tools the assignment requires.

05Observable completion evidence

Verify the artifact

Check the final result against the task brief. Require file paths, commands, test results, unresolved risks, and a concise statement of what changed.

06A repeatable next pass

Record the next move

If the result missed, change one cause at a time: the current request, operating behavior, stage mapping, project instruction, or verification rule.

02 / Baseline and instruction surface

Put the method where your tool will read it.

The premium framework pages explain effective framings and the special measures that make them work. The Markdown blocks below are starting structures for coding-agent instructions. Choose the file your tool supports, translate only the useful behaviors, and personalize them for the real project and objective. Do not put framework names in the file. Keep the same posture, progression, and evidence expectations active when you write each live prompt.

  1. 01Read the relevant framework guidance and identify the progression, evidence rules, cautions, and decisions that matter to your job.
  2. 02Choose the instruction file supported by the coding tool and repository.
  3. 03Translate the useful functions into agentic stages, checkpoints, decision gates, recovery rules, or scoped project instructions.
  4. 04Leave framework names and sample-turn wording out of the file. Describe only what the agent should do and why it matters.
  5. 05Replace every placeholder with real project context before asking the agent to act.
  6. 06Keep the current request specific, and phrase its opening move, progression, decision gates, and completion checks so they match the project framing.
  7. 07Run the work, inspect the result, and retain only the guidance that proved useful for the next similar task.

Example layout, not a required system

Choose the instruction file your tool supports:
  AGENTS.md
  CLAUDE.md
  .kimi-code/AGENTS.md
Optional for repeated work:
  docs/agent-guidance/OPERATING-NOTES.md
  docs/agent-guidance/TASK-BRIEF.md

Codex

01

AGENTS.md

Keep repository conventions, useful operating behaviors, stage guidance, commands, and verification rules here. Add a closer nested AGENTS.md only when one subtree needs different guidance.

Claude Code

02

CLAUDE.md

Keep project guidance concise and import or link focused Markdown references. Use CLAUDE.local.md only for personal instructions that should not be committed.

Kimi Code

03

AGENTS.md or .kimi-code/AGENTS.md

Use the project file for shared instructions. Keep reusable procedures as Markdown agent or skill files when the workflow should load only for matching tasks.

Cursor

04

AGENTS.md or .cursor/rules/*.mdc

Use AGENTS.md for plain repository guidance. Use scoped project rules when only certain paths or file types need a different operating behavior or verification standard.

GitHub Copilot

05

.github/copilot-instructions.md or AGENTS.md

Keep repository-wide guidance short. Put path-specific instructions in .github/instructions when a behavior applies only to part of the project.

Windsurf

06

AGENTS.md or .windsurf/rules/*.md

Use a root AGENTS.md for shared rules and workspace rule files for narrower activation. Include only the behaviors that improve that scope.

Aider

07

CONVENTIONS.md loaded with /read

Load the operating notes and task brief as read-only context. Keep editable project files separate so the source guidance is not changed during the run.

AGENTS.md or CLAUDE.md
# Project instructions

## Operating posture
- Maintain this stance: [the tone, level of challenge, and relationship to evidence that should remain stable]
- Frame the work by: [how the problem should be introduced and bounded]
- Treat success as: [the observable result this posture should produce]
- Never drift into: [the conflicting posture, shortcut, or assumption this route must prevent]

## How the work should run
- Start by: [the context, source material, or unknowns that must be established first]
- Progress by: [the ordered stages that make the result stronger]
- Carry forward: [facts, decisions, language, or evidence later work must preserve]
- At decision points: [what must be checked before continuing]
- Recover by: [what to do when evidence is missing, a check fails, or the route stops fitting]
- Do not include framework names or copy sample turns. Express only the behaviors that improve this project.

## Intent protection
- Preserve the requested action, supplied facts, exclusions, audience, and deliverable.
- If a proposed change alters any of those items, stop and name the change before acting.
- Do not add a posture the current request did not ask for.

## Working rules
- Work against the actual repository state.
- Use tools and checkpoints where they materially improve the result.
- Keep warnings, decisions, and checks beside the work they affect.

## Completion
- Return the requested artifact in the requested location and format.
- Run the relevant checks and report their actual outcomes.
Optional TASK-BRIEF.md
# Optional task brief

## Objective
[What the user needs to accomplish]

## Continue the project framing
- Open this task by: [the first move that matches the established operating posture]
- Progress by: [the equivalent framing and stage sequence for this specific request]
- At each decision: [the evidence or signal required before advancing]
- Maintain throughout: [tone, assumptions, boundaries, or interaction expectations that must stay consistent]
- Finish when: [the artifact and proof that complete the same route]

## Behaviors that should shape the work
- Establish first: [facts, context, sources, or unknowns needed before action]
- Build through: [how discovery, decisions, execution, and verification should progress]
- Preserve throughout: [details later stages must not lose]
- Require before advancing: [evidence, confirmation, test, or observable signal]
- Watch for: [the failure pattern, unsupported assumption, or drift this route must prevent]
- Recovery: [how to correct course without changing the objective]

## Keep fixed
- [Supplied facts]
- [Required actions]
- [Exclusions and non-negotiables]

## Deliverable
[The exact artifact, location, format, and audience]

## Reuse profile
Task family: [the kind of work this route can repeat]
Reusable when: [conditions that must remain true]
Replace each run: [facts, files, audience, constraints, and target]
Reset the route when: [a change that requires different progression, evidence, or decision behavior]

## Success checks
- [Observable check 1]
- [Observable check 2]
- [Command, test, or review that proves completion]

The calibration rule

The template gives you structure. Your project context makes it useful.

No framework page or template can know your subject, source material, audience, constraints, tools, or definition of done. Use the guidance as a behavioral baseline. Do not paste framework names or example turns into the file. Translate the effective measures into instructions for one real assignment and the instruction surface your tool actually reads.

Keep
Preserve what makes the framing work
Keep the useful progression, decision logic, cautions, and evidence standard intact. These behaviors are the reusable baseline.
Personalize
Replace the operating context
Add the real objective, subject matter, audience, source material, constraints, tools, deliverable, and success checks for this assignment.
Prove
Calibrate against a real result
Run a small representative task, inspect the artifact and evidence, then tighten the route where the result drifted or left a requirement unmet.

The complete circuit

Keep one framing alive from guidance to finished work.

Each of the 21 routes has a distinct operating posture. That posture may control how the agent opens the work, asks or tests, carries facts forward, handles resistance or uncertainty, decides when to advance, and proves completion. Capture those parts as a framing contract. A persona here is not a theatrical character. It is a consistent stance toward the user, the evidence, and the decision.

The project file holds the stable behavior.

The current prompt supplies the new assignment and activates the same behavior in task-specific language. Equivalent framing means matching the posture and progression, not repeating the same script.

  1. 01StudyFind the method's operating postureIdentify how it opens the work, frames the problem, moves between stages, handles uncertainty, and decides when the result is complete. Keep the mechanism, not the branded label.
  2. 02EncodeMake the posture durableWrite the stable behaviors into the project instruction file as direct actions, decision gates, evidence rules, recovery behavior, and completion expectations.
  3. 03PromptContinue the same framingWrite each live request in compatible language. Supply the new objective and facts, then restate the opening move, progression, evidence threshold, and finished artifact that matter for this run.
  4. 04ExecuteLet the agent carry the routeThe project file holds the stable operating behavior. The current prompt activates it for one assignment. Together they keep the agent from drifting into a generic answer or a conflicting posture.
  5. 05CalibrateKeep only what proves usefulInspect the artifact, decisions, and evidence. Tighten the instruction or prompt where the behavior disappeared, then reuse the proven route for similar work.

Worked translation

One framing. Two instruction surfaces. One observable standard.

This name-free example shows how a source behavior becomes durable project guidance, then reappears in the live request without copying a framework name or sample turn. Use the same translation process with whichever of the 21 routes fits your work.

  1. 01

    Behavior found in the source guidance

    Open with verified context before drawing a conclusion.

    Write into the project file
    Establish the observed state, available sources, and unresolved unknowns before proposing a cause or solution.
    Use equivalent framing in the prompt
    Investigate the duplicate-order failure. Begin with the observed sequence in the logs and code path, and keep unknowns separate from confirmed facts.
    Verify in the result
    The result distinguishes observations, inferences, and unknowns before recommending a change.
  2. 02

    Behavior found in the source guidance

    Progress by testing the assumptions that could change the direction.

    Write into the project file
    Test the highest-impact competing explanations in order. Do not advance while a failed check leaves the central premise unresolved.
    Use equivalent framing in the prompt
    Test whether retries, concurrent submissions, or idempotency handling explains the duplicate before editing the implementation.
    Verify in the result
    The chosen fix traces to a tested cause instead of the first plausible explanation.
  3. 03

    Behavior found in the source guidance

    Maintain a calm, evidence-led posture when evidence conflicts.

    Write into the project file
    Treat contradictions as work to resolve. Record what conflicts, gather the missing evidence, and revise the working explanation without hiding the change.
    Use equivalent framing in the prompt
    If the application logs and database history disagree, show the conflict and resolve it before choosing the repair.
    Verify in the result
    Conflicting evidence stays visible and changes the decision when warranted.
  4. 04

    Behavior found in the source guidance

    Finish only when the claim can be checked in the real work surface.

    Write into the project file
    Complete the requested artifact, run the relevant checks, and report the paths, commands, results, and unresolved limits.
    Use equivalent framing in the prompt
    Implement the supported fix, add the regression test, run the affected suite, and report the changed files and actual results.
    Verify in the result
    The run ends with a working artifact and observable verification, not a confident summary.

03 / Behavior routing

Extract the useful behavior. Leave the label out.

Start with the job the method must perform. The source names below help you find the relevant premium guidance, but they do not belong in the coding-agent file. Read the source page, extract its selection signal, operating role, evidence standard, and caution, then rewrite those measures as direct project behavior.

Inquiry and discovery

5 methods

  1. Socratic Elenchus

    01
    Select when
    The user needs to test a claim, expose an inconsistency, or reach a conclusion through ordered questions.
    Role
    Turn the central claim into a chain of questions that tests premises and contradictions.
    Evidence
    Question sequence, premise table, and reasoned conclusion.
    Write into the file
    Turn the central claim into a chain of questions that tests premises and contradictions. Require the work to produce or verify: Question sequence, premise table, and reasoned conclusion.
    Watch
    The questioning sequence must still end in the deliverable the user requested.

  2. Deep Curiosity Complex

    02
    Select when
    The request is weak because the important unknowns have not been identified or explored.
    Role
    Find the few questions whose answers materially change the result, then carry those answers forward.
    Evidence
    Discovery brief, ranked unknowns, and updated task direction.
    Write into the file
    Find the few questions whose answers materially change the result, then carry those answers forward. Require the work to produce or verify: Discovery brief, ranked unknowns, and updated task direction.
    Watch
    Stop asking when the remaining unknowns no longer change the output.

  3. Scharff Technique

    03
    Select when
    The user needs to learn what is known, missing, corrected, or confirmed through a calm information exchange.
    Role
    Organize known facts and targeted prompts so new information can be compared with the existing record.
    Evidence
    Known-facts ledger, information gaps, and interview notes.
    Write into the file
    Organize known facts and targeted prompts so new information can be compared with the existing record. Require the work to produce or verify: Known-facts ledger, information gaps, and interview notes.
    Watch
    Do not turn an unconfirmed statement into an established fact.

  4. Motivational Interviewing

    04
    Select when
    The task helps a person examine ambivalence, priorities, readiness, or their own reasons for change.
    Role
    Use reflective questions to surface the user's reasons, obstacles, and next choice without replacing their agency.
    Evidence
    Readiness summary, change reasons, barriers, and chosen next step.
    Write into the file
    Use reflective questions to surface the user's reasons, obstacles, and next choice without replacing their agency. Require the work to produce or verify: Readiness summary, change reasons, barriers, and chosen next step.
    Watch
    Do not choose the person's goal or claim readiness they did not express.

  5. SPIN Selling

    05
    Select when
    The assignment must move from the current situation to a problem, its implications, and a defined need.
    Role
    Structure discovery so the final recommendation is tied to an established need.
    Evidence
    Situation brief, problem statement, implication analysis, and need-based recommendation.
    Write into the file
    Structure discovery so the final recommendation is tied to an established need. Require the work to produce or verify: Situation brief, problem statement, implication analysis, and need-based recommendation.
    Watch
    Do not invent a need or implication that the discovery record does not support.

Reasoning and decision framing

5 methods

  1. Reflexive Control

    01
    Select when
    The assignment examines how available information, assumptions, or choice architecture shapes a decision.
    Role
    Map the decision environment and test which supplied conditions change the selected course of action.
    Evidence
    Decision map, assumption ledger, and option comparison.
    Write into the file
    Map the decision environment and test which supplied conditions change the selected course of action. Require the work to produce or verify: Decision map, assumption ledger, and option comparison.
    Watch
    Label assumptions and do not present a constructed frame as an observed fact.

  2. 48 Laws of Power

    02
    Select when
    The assignment centers on leverage, positioning, incentives, reputation, or competing interests.
    Role
    Analyze the power relationship and make the consequences of each position visible.
    Evidence
    Stakeholder map, leverage analysis, and decision memo.
    Write into the file
    Analyze the power relationship and make the consequences of each position visible. Require the work to produce or verify: Stakeholder map, leverage analysis, and decision memo.
    Watch
    Do not replace evidence with dramatic claims about influence or control.

  3. Double Bind

    03
    Select when
    The user is facing two conflicting demands, constraints, or outcomes that appear impossible to satisfy together.
    Role
    Expose the conflict, test the hidden assumptions, and create options outside the false binary.
    Evidence
    Constraint map, assumption test, and alternative decision paths.
    Write into the file
    Expose the conflict, test the hidden assumptions, and create options outside the false binary. Require the work to produce or verify: Constraint map, assumption test, and alternative decision paths.
    Watch
    Do not force a binary choice when the task permits another option.

  4. Cialdini's 7 Principles

    04
    Select when
    The user needs to understand or design how common influence principles affect an audience decision.
    Role
    Map each relevant influence principle to the audience, message, and observable response.
    Evidence
    Influence map, message options, and audience response checks.
    Write into the file
    Map each relevant influence principle to the audience, message, and observable response. Require the work to produce or verify: Influence map, message options, and audience response checks.
    Watch
    Use only principles that match the stated audience and evidence.

  5. Aristotle Rhetoric

    05
    Select when
    The assignment needs a balanced argument that accounts for credibility, audience concerns, and reasoning.
    Role
    Structure the finished case across credibility, emotion, and logic while keeping factual claims sourced.
    Evidence
    Audience-aware argument, evidence plan, and final communication.
    Write into the file
    Structure the finished case across credibility, emotion, and logic while keeping factual claims sourced. Require the work to produce or verify: Audience-aware argument, evidence plan, and final communication.
    Watch
    Do not let emotional framing outrun the evidence or the requested tone.

Conversation and sequence

5 methods

  1. The Echo Chamber

    01
    Select when
    The work depends on carrying the model's earlier language or accepted details into a later turn.
    Role
    Track what the model introduced and test how that material affects the next instruction.
    Evidence
    Turn ledger, quoted carryover, and replay transcript.
    Write into the file
    Track what the model introduced and test how that material affects the next instruction. Require the work to produce or verify: Turn ledger, quoted carryover, and replay transcript.
    Watch
    Keep model-generated details separate from facts supplied by the user.

  2. Manipulation Matrix

    02
    Select when
    Several framing variables must be tested separately before they are combined into one sequence.
    Role
    Isolate each variable, compare the result, and retain only changes that produce a repeatable difference.
    Evidence
    Comparison matrix, selected variables, and ordered runbook.
    Write into the file
    Isolate each variable, compare the result, and retain only changes that produce a repeatable difference. Require the work to produce or verify: Comparison matrix, selected variables, and ordered runbook.
    Watch
    Do not combine variables before their individual effects have been recorded.

  3. Neuro-Linguistic Programming Milton Model

    03
    Select when
    The task depends on language flexibility, interpretation, tone, or how phrasing changes the reader's response.
    Role
    Shape wording and progression while preserving the fixed meaning in the task brief.
    Evidence
    Language map, phrasing variants, and final authored version.
    Write into the file
    Shape wording and progression while preserving the fixed meaning in the task brief. Require the work to produce or verify: Language map, phrasing variants, and final authored version.
    Watch
    Check every revision against the intent contract so ambiguity does not change the objective.

  4. Deny, Attack, Reverse Victim and Offender

    04
    Select when
    The assignment examines reversal, blame shifting, denial, or conflicting accounts of responsibility.
    Role
    Separate each claim, response, and reversal so the sequence can be inspected without collapsing the parties together.
    Evidence
    Claim timeline, responsibility map, and contradiction record.
    Write into the file
    Separate each claim, response, and reversal so the sequence can be inspected without collapsing the parties together. Require the work to produce or verify: Claim timeline, responsibility map, and contradiction record.
    Watch
    Do not treat an accusation or reversal as proof without supporting evidence.

  5. Reid Technique

    05
    Select when
    The assignment requires a structured review of statements, inconsistencies, and decision points across an interview or dialogue.
    Role
    Organize the conversation into explicit stages and compare statements with the evidence record.
    Evidence
    Interview sequence, statement comparison, and unresolved discrepancy list.
    Write into the file
    Organize the conversation into explicit stages and compare statements with the evidence record. Require the work to produce or verify: Interview sequence, statement comparison, and unresolved discrepancy list.
    Watch
    Do not treat pressure, agreement, or a confession as proof of accuracy.

Identity, trust, and pressure analysis

4 methods

  1. Social Engineering

    01
    Select when
    The work examines identity, trust, authority, pretext, or human decision points.
    Role
    Map the actors, trust signals, requested actions, and checks that influence the decision.
    Evidence
    Actor map, trust path, and interaction analysis.
    Write into the file
    Map the actors, trust signals, requested actions, and checks that influence the decision. Require the work to produce or verify: Actor map, trust path, and interaction analysis.
    Watch
    Keep claimed identities and permissions marked as claims until they are verified.

  2. Behavior, Information, Thought, Emotion Model

    02
    Select when
    The assignment needs a four-part analysis of behavior, information, thought, and emotion.
    Role
    Separate observations into the four dimensions and show where they interact.
    Evidence
    Four-domain assessment and cross-domain relationship map.
    Write into the file
    Separate observations into the four dimensions and show where they interact. Require the work to produce or verify: Four-domain assessment and cross-domain relationship map.
    Watch
    Do not diagnose a person or group from incomplete observations.

  3. Money, Ideology, Compromise, Ego

    03
    Select when
    The task examines incentives or motivations that may be financial, ideological, compromising, or status-driven.
    Role
    Classify possible incentives and identify the evidence needed to support or reject each one.
    Evidence
    Motivation matrix, evidence gaps, and ranked hypotheses.
    Write into the file
    Classify possible incentives and identify the evidence needed to support or reject each one. Require the work to produce or verify: Motivation matrix, evidence gaps, and ranked hypotheses.
    Watch
    Treat motivations as hypotheses unless the user supplied supporting evidence.

  4. Dark Psychology

    04
    Select when
    The assignment examines coercion, manipulation, emotional pressure, or exploitative interpersonal patterns.
    Role
    Identify the pattern, its observable signals, and the decision or behavior it is trying to change.
    Evidence
    Pattern analysis, evidence table, and impact summary.
    Write into the file
    Identify the pattern, its observable signals, and the decision or behavior it is trying to change. Require the work to produce or verify: Pattern analysis, evidence table, and impact summary.
    Watch
    Use observable behavior and avoid unsupported clinical or personality labels.

Library methods

2 methods

  1. The Lazarus Library

    01
    Select when
    The assignment depends on a stable identity, source library, or long-form context that must remain consistent across the run.
    Role
    Establish the context frame and the source material the agent must carry into later work.
    Evidence
    Context brief, source map, and staged task plan.
    Write into the file
    Establish the context frame and the source material the agent must carry into later work. Require the work to produce or verify: Context brief, source map, and staged task plan.
    Watch
    Keep the requested deliverable separate from the identity or library frame.

  2. Direct Library Method

    02
    Select when
    The user needs a direct result, but the model also needs a precise body of reference material to produce it correctly.
    Role
    Connect a clear request to an organized reference library without burying the requested action.
    Evidence
    Direct task brief with a linked reference set and completion checks.
    Write into the file
    Connect a clear request to an organized reference library without burying the requested action. Require the work to produce or verify: Direct task brief with a linked reference set and completion checks.
    Watch
    Do not let supporting context become a substitute for the actual request.

04 / Reuse and controlled blends

Reuse the route. Refresh the assignment.

A well-calibrated setup saves time because the operating behavior, file routing, commands, output shape, and verification checks can survive into similar work. Only the facts of the new run should change. When the job changes enough to need different progression, evidence, or decision rules, replace those behaviors instead of stretching the old route past its fit.

Stable
Proven project guidance
Keep the operating behavior, repository rules, and stage logic only while they continue to improve this type of work.
Reusable
Task-family route
Reuse the file path, operating pattern, output shape, commands, and checks for work with the same execution needs.
Replace
Run-specific facts
Refresh the objective, files, audience, constraints, known facts, exclusions, and acceptance evidence every time.

Controlled experiment

Blend behaviors by purpose, not by label.

One behavior controls the main route. A supporting behavior may address one named gap. Never paste entire methods or their names into the file and expect the agent to resolve the conflicts.

  1. 01Run the assignment with one operating behavior and save the result.
  2. 02Name one gap the first run did not handle well enough.
  3. 03Add one supporting behavior for that gap only, then rerun the same request and checks.
  4. 04Keep the blend only if a named success check improves without changing the original objective.

Three useful experiments

Primary

Premise testing

Supporting

Unknowns ranking

The main behavior tests the central claim. The supporting behavior ranks only the unknowns that would materially change the conclusion.

Keep it only if: Compare premise coverage and unresolved unknowns against a primary-only run.

Primary

Need-led discovery

Supporting

Readiness checks

The main behavior connects the current situation to the stated problem and need. The supporting behavior records readiness and the person's chosen next step.

Keep it only if: Confirm that every recommendation traces to an expressed need and no readiness claim was invented.

Primary

Source grounding

Supporting

Audience shaping

The main behavior grounds the result in the approved reference set. The supporting behavior shapes the final argument for the named audience.

Keep it only if: Check source coverage separately from audience fit, reasoning, and final delivery.

Ways to adapt the method

Match the instruction pattern to the work.

01

Single behavior profile

Give one operating pattern one defined job in the project instructions. This is the cleanest starting point and the easiest result to evaluate.

02

Behavior comparison

Run the same request twice with a different operating behavior each time. Keep the model, files, tools, and success checks fixed, then compare the results.

03

Staged handoff

Use one behavior for discovery and another for delivery. State the boundary between them so the delivery stage cannot rewrite the findings from discovery.

04

Nested repository scope

Keep the shared objective at the root and add narrow instructions closer to the affected folder. This works well when research, code, documentation, and tests need different completion rules.

05

Reusable command or skill

Turn a proven operating route into a tool-native command or skill. Keep the reusable stages concise and supply the current objective each time.

06

Evidence replay

Save the finished brief, operating notes, changed files, and checks. Start a fresh session and ask another agent to reproduce the result from those files alone.

05 / Final field check

Prove the behaviors improved the result.

A source method is not useful because its name appears in a file. It is useful when the task stays fixed, its strongest behaviors shape the work, and the workspace contains evidence that the requested result was completed.

  1. 01The instruction file describes the useful behaviors in plain language and contains no framework names or copied turn scripts.
  2. 02The live prompt activates the same operating posture, progression, decision gates, and evidence standard for the current assignment without repeating a canned script.
  3. 03The agent can state the user's objective without changing its posture or requested action.
  4. 04Every instruction changes a decision, action, evidence requirement, recovery step, or completion check; decorative method language is absent.
  5. 05The result exists in the requested file or format, not only in chat.
  6. 06Commands and tests are reported with their actual outcomes.
  7. 07A fresh session can follow the project instructions without relying on an old conversation.

Platform sources

Verify the file behavior before you standardize it.

Coding tools change. These first-party references document the instruction surfaces named in this guide. Check the source again when a tool changes how files are discovered or scoped.

01OpenAI: AGENTS.md repository guidance02Anthropic: Claude Code project memory03Kimi: agents and instruction files04Cursor: repository rules05GitHub: Copilot repository instructions06Windsurf: workspace rules and AGENTS.md07Aider: convention files

Next route

Continue with the adjacent problem.

Prompt engineering Prompt optimization AI red teaming