HA-SCOM-002 · Twelve steps · Seventeen prompts
Practical Guide to Source Control Using GitHub
Your first website project with Codex: from a project folder to verified handover
By Henok Ahmed
Version 1.0.0
Document HA-SCOM-002 · First prepared 25 October 2026 · Edited 28 October 2026
Status: Published with author approval on 28 October 2026.
Use with the Operating Model and Template Pack, Version 1.0.0. The model defines requirements; this guide explains how to apply them. License: CC BY 4.0, with the exclusions stated below.
Before you begin
This guide is for a computing graduate with little practical website-delivery experience who wants to offer informational or marketing websites as a freelancer or small business. You may be working outside Addis Ababa, with variable connection, power or equipment access. Choose arrangements that fit your circumstances rather than copying somebody else's infrastructure.
You will complete one bounded source-control cycle: inspect the starting work, establish a repository, make a change through a branch and PR, check it, identify a candidate, record the actual decision, apply the authorized result, verify and hand over. This is source-control practice alongside website development; it is not an instruction to generate a whole website from one prompt.
First use an authorized test installation and fictional content. A practice result is DONE—PRACTICE, not a delivered client project. Move to client delivery only when you can demonstrate the checks and work within the approved scope. No duration, earnings or customer-acquisition result is promised.
Read the WordPress foundation guide for client preparation. Build the site using the 24 Working Hours runbook and its prompt deck. Source control begins with its baseline/source decisions and continues through the build; do not wait until the last day to discover that none of the accepted work was recorded. The 24 hours remains a planning allowance, not a tested guarantee, and this GitHub exercise adds no time promise.
Prerequisites and working routes
You need an authorized project folder, a GitHub account you control or are authorized to use, a safe WordPress test destination, usable source/export files, approved or fictional contacts, and an agreed recovery method before affected deployment. Use the actual service to sign in. Do not send passwords or tokens in prompts.
Choose one route and record it:
- Local Git with Codex assistance: when installed Git and authorized tools are available, inspect local changes, make commits, synchronize with GitHub and use its PR workflow. Codex must report actual capabilities first.
- Browser with prepared text files: use GitHub's supported browser controls to create branches, upload/edit reviewed source, commit and open PRs. Keep your local copies reconciled manually. This route does not give your local folder Git history or offline commit capability.
You do not need GitHub CLI, Desktop and several editors at once. If local Git is missing, arrange one supported installation/setup route or use the browser route for a text-based practice exercise. The assistant should give one precise manual action when it cannot act. More complex source/build work may require local tooling before it is suitable for client delivery.
GitHub's official Hello World exercise is a useful browser rehearsal. Finish it in a separate disposable learning repository if repository, branch, commit and PR still feel abstract. Do not use the client's project to learn destructive history commands.
The fictional project
Mekdes, a computing graduate in Debre Markos, is learning to deliver websites. Sara Bekele, in Bahir Dar, is the fictional owner of LumenBridge Advisory. These are the same fictional teaching characters used in the foundation guide. This exercise narrows the practice scope to four pages: Home, About, Services and Contact. A real client's scope would be separately agreed.
The website explains advisory services and offers an authorized email/phone contact route. It has no payment, account, booking or private upload. The illustrative domain lumenbridge.example is not a deployment address. You must supply your actual permitted test URL; never publish fabricated business details as real.
Use a repository name such as lumenbridge-website-practice under your approved account. This is a proposed name, not a repository this document has created. No example URL, record, approval, result or commit label below is evidence that a real action occurred.
The running change is modest: clarify what a digital-strategy service includes and make the contact action understandable. One useful change is enough to learn the complete cycle.
How to use the prompts
Work in the same project/chat where practical. Read each task and check before pasting its prompt. Fill missing operational facts in the project brief when asked. The assistant must inspect actual state, keep changes bounded, and distinguish its own checks from your reported observations. If tools are unavailable, perform the specified manual action and return the observed result; do not accept a claim that a push or save happened without evidence.
The prompts are original teaching instructions, not tested scripts or promises of available tools. Their structure follows OpenAI's prompting guidance: outcome, context, constraints and verification. They do not override permissions or replace competent review.
Step 1 — Open the correct project and inspect capabilities
You do: select a dedicated folder. Keep client projects separate. Open the actual folder rather than assuming the chat can see your whole computer.
Paste into Codex:
Help me establish source control for a bounded WordPress website project.
I have a computing background but little delivery experience.
Inspect only my selected project folder and available capabilities.
Report the absolute path, existing files, Git repository/branch/remotes
if present, installed Git availability, GitHub access and WordPress access.
Do not expose credential values or inspect unrelated folders.
Do not initialize, overwrite, install, upload or change WordPress yet.
Recommend one available working route: local Git or browser with prepared files.
Explain what I must do manually and how we will verify it.
Check: you can identify the folder and explain the chosen route. Existing unrelated work is a stop condition, not an invitation to reset it.
Common error: “not a git repository” may simply mean this folder has not been initialized. It does not justify running initialization in an unverified parent directory. “Permission denied” requires correcting authorized access, not bypassing it.
Step 2 — Record scope, ownership and the profile
You do: supply mode (practice/client), real test URL, approved account/repository owner, permitted contact route and who approves changes. For practice, you act as the exercise owner; do not manufacture Sara's real-world approval.
Paste:
Prepare docs/project-brief.md and the combined Foundation project record
from T01/T02 in the companion Template Pack, using my supplied facts.
Record mode, four-page scope/exclusions, audience, test destination,
source/editor approach, repository ownership/visibility, contact permissions,
service dependencies, recovery owner and actual approval authority.
Assess consequences before recommending Foundation; explain any Controlled trigger.
Mark missing information unresolved. Do not invent URLs, contacts or approval.
Keep the fictional exercise visibly fictional and off real public service routes.
Do not create a remote repository or change WordPress from this prompt.
Check: the selected profile has reasons. The client or practice owner has actually accepted the adoption decision. T01 adoption does not authorize public launch.
Common error: copying every role into the freelancer's name erases the client's authority. Identify which decisions the client retains. If scope requires Controlled and you lack suitable support, resolve that gap before proceeding.
Step 3 — Capture the actual website baseline and source boundary
You do: confirm your editing approach and coordinate the relevant backup/restoration arrangement. For an existing site, preserve the current authored content before changing it. Use a permitted test environment for recovery practice.
Paste:
Read the brief and adopted profile. Inspect the current authorized website
and local source, or request precise manual observations where access is absent.
Record current pages/URLs, environment-specific IDs, editor/builder, relevant
versions and what changed since any existing export.
Prepare a tracked-source inventory using the supported export/editing route.
Record separately managed database, media, plugins, hosting, domain and email.
Keep private records, credentials and database dumps out of source.
Create docs/baseline.md with capture date, source origin, omissions and gaps.
Reference the real backup/recovery arrangement; do not claim it exists or works
without evidence. Do not change the live site.
Check: you can point to captured source and explain what restoring it would not restore. A builder export has a documented editing/import route. Reopen the saved export to check that it contains the expected content.
Common error: an empty content folder is not a baseline. A screenshot is useful visual evidence, but it is not editable source. GitHub cannot reconstruct untracked WordPress data automatically; see WordPress Backups.
Step 4 — Establish the repository and the first setup branch
You do: confirm the approved owner, name and visibility. Complete account setup or service prompts through GitHub itself. If asked to act manually, return the observed repository URL; do not paste credentials.
Paste:
Use the adopted ownership and visibility decision to establish or connect
the project repository through available authorized tools.
Inspect an existing repository before writing. Do not overwrite unknown history,
change visibility, invent remote success or add unapproved collaborators.
For a new empty project, use a minimal README bootstrap commit if necessary
so a setup/source-baseline branch can be created from main; record that bootstrap.
For local Git, inspect the folder first and preserve any existing work.
If repository creation or authentication is unavailable, give me the exact
manual action and wait for the resulting URL/state before dependent work.
Done when we have verified repository identity and the intended setup branch.
Check: the remote points to the intended account/project; the branch is setup/source-baseline or the actual agreed equivalent. A new repository's placeholder README is not yet the website baseline.
Common error: pushing to somebody else's repository or to an old remote. Stop and verify identity. If the repository already contains a README/history, reconcile with it; do not force a first push over it.
Step 5 — Import only reviewed source and test one harmless PR
You do: inspect filenames and the actual diff. Use the model's repository layout as guidance, not as a demand to create empty automation files. Start with useful instructions, source and records.
Paste:
On the verified setup branch, prepare a reviewed baseline from our source inventory.
Create useful README, CONTRIBUTING, SECURITY, CHANGELOG and ignore rules appropriate
to this project; add only actual source, authorized assets and needed records.
Explain how source returns to WordPress and what is excluded.
Inspect staged/proposed changes before committing; report suspicious files
without printing secrets. Do not use broad upload commands over unreviewed folders.
Commit the bounded import and open a setup PR to the declared main branch.
Record checks/review and observed branch/check enforcement in T03.
Test configured restrictions with a harmless test change; do not weaken them.
Do not merge until required review and actual adoption conditions are satisfied.
Check: open the PR, read its changed-files view, and locate an expected service paragraph. Verify the setup source and instructions are usable. After its actual review, merge under the adopted route and record the accepted commit in docs/baseline.md.
Common error: adding .gitignore after a sensitive file was already committed. Ignore rules do not remove tracked history; handle any exposure through the authorized incident route. See GitHub ignoring files.
If enforcement is unavailable: record required versus observed behavior and any owner-approved Foundation compensation. Never make the repository public just to change feature availability. A missing rule remains a gap; it is not “enforced manually.” Consult current ruleset documentation.
Step 6 — Define the first client-sized change
You do: discuss the outcome in plain language. The fictional request is clearer service deliverables and a visible Contact action, without adding a form or changing business claims.
Paste:
Create T04 as an issue or traceable local record for this one change:
clarify service deliverables and make the Contact action understandable.
Use only owner-approved facts and the existing approved contact route.
Record scope/exclusions, affected pages, factual/link/mobile/keyboard checks,
saved-output verification and how to recover the previous content.
Distinguish an included correction from a new feature request.
Do not change WordPress or create unrelated features yet.
Check: you can decide whether the change is done by reading its criteria. Record the actual issue identifier; the example LB-12 is only a teaching label.
Common error: “make it professional” is not an acceptance test. Replace it with observable wording, link and layout criteria. New booking functionality belongs in separate discovery/scope, not in a hidden expansion of this change.
Step 7 — Implement on a focused branch
You do: confirm that the accepted baseline and intended base branch are current. For local work, inspect remote differences before integrating them.
Paste:
Inspect current work, branch and remote state without discarding changes.
From the declared accepted baseline, create a focused branch for our actual issue,
for example feature/12-service-copy only if that identifier is correct.
Implement the agreed content/style correction in tracked source and the authorized
test editing route. Preserve existing URLs and unrelated work.
If the WordPress editor is the source of the change, export/reconcile its accepted
result before committing so the repository reflects actual saved content.
Show the diff and explain the effect. Do not deploy or publish to production.
Check: the diff contains only intended changes. Reopen the saved test page; do not rely only on the editor canvas.
Common error: edits exist only in WordPress or only in a detached local export. Record which is authoritative and reconcile deliberately. A different page ID on another installation is not the same target.
Step 8 — Verify the change and open its PR
You do: inspect desktop and narrow-screen layouts, use the keyboard, and check the actual email/phone destination. If a form is separately included, verify actual synthetic message receipt; it is outside this practice change.
Paste:
Review the exact changed source and saved test page against T04.
Run relevant existing checks; add a meaningful check only for a real failure mode.
Record PASS, FAIL or UNTESTED with version, environment, verifier and evidence.
Check factual wording, links, actual viewport width, keyboard/focus and saved output.
Open/update the PR with T05, issue link, final diff summary, actual checks,
omitted checks, recovery implications and disclosed self-review if applicable.
Do not report comprehensive accessibility or independent human review from AI review.
Keep production unchanged.
Check: you can open the PR, identify changed lines and explain what a passing automated check did. A screenshot corresponds to the current revision.
Common error: green CI despite an obscured Contact action. Record the manual criterion as FAIL, correct the defect and repeat affected checks. An unavailable browser test is UNTESTED, not an assumed pass.
Step 9 — Accept source and identify a candidate
You do: resolve review findings. For Foundation, disclose self-review; for Controlled, obtain the required independent review. Client acceptance of claims remains distinct from technical review.
Paste:
Reconcile PR findings with the actual head revision and evidence.
List unresolved findings and requirements; do not merge while material conditions fail.
When the required actual change-acceptance decision exists, integrate using our
adopted strategy and record the resulting source identity.
Prepare T06 for a candidate such as v0.1.0-rc.1 from exact reviewed source.
Record the full SHA, artifact location/checksum when applicable, checks and limits.
Mark a GitHub candidate release as prerelease if using Releases.
Do not move an old tag, deploy or infer client launch approval.
Check: source, candidate artifact and evidence refer to the same work. If an artifact was built from the pre-merge source, verify that it matches the accepted revision or rebuild/retest from the accepted commit. Do not assume merge identity and branch identity are interchangeable.
Common error: client requests changed wording after testing. Record the change, renew affected checks and identify a new candidate. Keep the rejected candidate in history rather than replacing it silently.
Step 10 — Obtain the actual owner's delivery decision
You do: review or send the real owner the candidate through the agreed review route. Do not send private project material to others without authority. Practice approval is your explicitly labeled exercise decision.
Paste:
Prepare T07 for the exact candidate and intended destination.
Separate content/change acceptance, release readiness, deployment authorization
and public launch. Foundation may record them together, but make each explicit.
Show the owner changed pages, preview, contact route, known limits, operator,
recovery reference and conditions in plain language; link the technical identity.
Do not record approval until the authorized person actually supplies it.
Do not deploy from this request.
When the real decision arrives, record its actual approver, date/time/timezone, candidate, destination, scope, conditions and evidence. If the decision is deferred or rejected, preserve that result.
Check: the decision covers the next action and all material conditions are satisfied. “Approved design” alone is not necessarily authorization to change a live website.
Common error: assuming client silence means consent. Wait for the required decision. A changed candidate or destination requires reconsidering affected authorization.
Step 11 — Apply the authorized version and verify the destination
You do: confirm the operator, recovery availability and window. Practice stays at its appropriate test destination and access boundary. Client mode uses the approved production target.
Paste:
Read the actual T07 decision and T06 candidate before writing.
Verify source/artifact identity, exact target URL/page IDs, current destination state,
backup/restoration availability, operator and approved rollback triggers.
Stop if conditions are missing or unexplained live drift exists.
Apply only the authorized content through our documented WordPress route.
Give one precise manual action when tools cannot act; verify the reported result.
Reopen saved pages and verify navigation, contact destinations, mobile behavior,
required public/signed-out access or restricted practice access, and actual version.
Complete T08 with what was applied and what operates afterward.
On failure use the authorized recovery route and verify restoration.
Do not change DNS, buy services or alter unrelated plugins from this prompt.
Check: the saved destination actually matches the approved result. A successful merge, upload or save message alone does not pass this check. Record an access verification blocked by authentication/CAPTCHA as unperformed; have the authorized person complete the appropriate manual check.
Common error: restoring an earlier Git commit and declaring the live site restored. Apply the relevant content/artifact or operator restoration procedure, then verify the actual website and data affected. See T08/T09.
Step 12 — Hand over, reconcile and close
You do: have the intended maintainer perform an agreed routine edit. Confirm ownership, renewal and support arrangements. Do not assume a GitHub invitation establishes complete client ownership of every service.
Paste:
Create the T08 handover supplement and T12 learning review for the actual delivery.
Record operating version/source, WordPress editing instructions, repository and
service owners, renewals, recovery references, support boundaries and open work.
Include no credentials. Reconcile any final WordPress edits with tracked source.
Separate defects from future work in T11 and assign owners.
Record actual learning/work/waiting time without claiming a guaranteed duration.
Use DONE—PRACTICE only for a verified exercise with the intended access boundary.
Use DONE—CLIENT only when authorized deployment, checks and agreed handover are complete.
Otherwise state IN PROGRESS or BLOCKED with the exact unresolved criterion.
Check: you can trace issue → PR → reviewed source → candidate → real decision → deployed/test result → handover. The owner knows what support continues and who handles restoration. If a required criterion remains untested, the handover states the limit; a material blocker prevents completion.
Common error: treating the source repository as the entire deliverable or promising indefinite support. Confirm the actual agreement, and distinguish delivered work from separately priced maintenance or later scope.
Recovery prompts for common interruptions
Use the relevant prompt with actual, sanitized error details. These prompts diagnose and preserve work; they are not permission to discard history or change security.
A. Push rejected or local/remote histories differ
Inspect the reported synchronization failure and both source identities.
Preserve uncommitted and unpushed work. Explain whether the remote advanced,
the destination is wrong or histories differ. Do not force push or reset away work.
Propose the smallest normal reconciliation and resolve conflicts against the
accepted intent. Repeat affected checks before updating the PR.
B. A merge conflict appears
Explain each conflict in terms of the two intended changes, not just markers.
Preserve both versions until the correct outcome is established.
Resolve only conflicts whose intent is supported; ask for the needed choice otherwise.
Inspect the final diff, check for leftover markers and retest affected behavior.
Do not use blanket accept-ours/accept-theirs choices over unknown work.
C. A credential or private file may have been included
Do not print, upload or repeat the suspected sensitive content.
Identify affected file/commit locations by safe references and stop further distribution.
Prepare the authorized revocation/rotation and exposure-assessment actions.
Explain what adding an ignore rule or deleting the latest line cannot repair.
Keep incident evidence restricted and request the necessary operator involvement.
D. Internet or power interrupted the session
Read the saved project brief, baseline and progress/decision records.
Inspect actual local work and, when connected, the remote and WordPress state.
Identify uncommitted/unpushed work, remote changes and the last verified step.
Do not infer that an interrupted push, merge or save succeeded.
Resume the next bounded action without overwriting unknown work.
E. GitHub access or an automated check is unavailable
Record the exact unavailable capability/check and its consequence.
Preserve prepared work and provide the smallest supported manual action.
Do not claim enforcement, account access or a passed test without observation.
Use the adopted profile's gap/exception process; do not lower the profile or
publish a private repository merely to obtain a feature.
Worked Case A — A first Foundation delivery
This narrative is fictional, including results, identifiers and decisions. It shows how records connect; it is not evidence to copy into your project.
Mekdes completes a restricted four-page LumenBridge practice site, then agrees the bounded service-copy change with the fictional client. Sara retains owner decisions; Mekdes combines implementation, maintenance and operation with disclosed self-review. Codex helps inspect source and prepare checks. It is not the client's representative.
T01/T02 select Foundation because consequences are bounded, no restricted records or application transactions are involved, applicable obligations permit it and recovery is understood. T03 records an illustrative account where desired PR enforcement is unavailable. An explicitly accepted manual review arrangement remains a visible limitation, not enforced protection.
The baseline has an actual full SHA in the fictional project's own record. This narrative uses symbolic labels LB-C1, LB-C2 and LB-C3 only for readability. They are not real SHAs.
| Event | Fictional result | Record |
|---|---|---|
| Issue LB-12 | Clarify deliverables and Contact action | T04 |
| First PR candidate, LB-C1 | Source checks pass; narrow-screen Contact action is hidden | T05/T06: FAIL |
| Correction, LB-C2 | Layout and keyboard checks pass; Sara requests one wording clarification | T11; candidate superseded |
| Revised candidate, LB-C3 | Approved facts and saved output checked; contact destinations correct | New T06 and review evidence |
| Recovery check | Prior authored content restored in the safe test site, then candidate reapplied | T09; scope and limits stated |
| Decision | Sara explicitly accepts the named candidate, target and public claims; authorizes deployment/launch | T07; fictional decision only |
| Delivery | Mekdes checks expected current state, applies candidate and verifies the actual destination | T08 |
| Handover | Sara performs an agreed edit; source is reconciled and owners/support recorded | T08 handover supplement |
| Learning | A mobile criterion is strengthened; booking is recorded as future discovery | T12/T11 |
An acceptance record cannot say “all contact tests passed” if only the email link was checked. If a form is later added, actual receipt and failure behavior become new acceptance work. If a required signed-out check is blocked, the record remains UNTESTED until a person actually performs it.
The candidate becomes stable v0.1.0 only when the adopted readiness and acceptance conditions are met; deployment evidence separately identifies what runs. A later correction follows a new issue/fix branch and becomes v0.1.1. The record does not imply unmeasured gains in enquiries or earnings.
Worked Case B — Recognize a project that needs Controlled
Rivergate Support is a separate fictional customer portal with restricted service records. It is an escalation illustration, not the next exercise for a learner who has only completed a marketing site.
Daniel authors a change, Amina independently reviews, Ruth is the service owner/production approver, and Elias operates deployment/recovery. T02 selects Controlled because unauthorized access and service failure could materially affect customers. They use restricted staging with synthetic data and verify required PR/check/review enforcement. Short-lived branches into main remain suitable; extra branches alone would add no assurance.
Issue RIV-41 adds a request-status view. A negative authorization test reveals access to another customer's record. Required CI fails and the PR is blocked. Daniel corrects the authorization path, runs relevant regression checks and receives Amina's review of the changed revision. A merge identifies accepted source, not deployment.
The next candidate has a full source SHA, artifact checksum and configuration record in T06. T09 exercises the applicable rollback. For this application-only change, the fictional owner chooses 30 minutes as the recovery-time objective with no planned data loss; those are not defaults for other projects. The simulated rehearsal takes 12 minutes. Ruth explicitly authorizes the named artifact and destination with rollback triggers.
Production smoke checks reveal an environment-specific configuration failure. Elias restores the prior artifact/configuration under the preauthorized recovery route and verifies service. The simulated restoration takes 14 minutes. The failed candidate stays in history; no tag is moved and the incident is not hidden by successful source tests.
The team finds evidence of a missing configuration-contract check, prepares a new reviewed candidate, repeats affected checks and obtains fresh authorization. The second simulated deployment succeeds. T12 assigns improvements and retests them.
A beginner should recognize the need for suitable experience, independent people and stronger operational evidence here. They should narrow, partner with a competent team or decline unsupported work rather than treating the portal as a few more WordPress pages. These timings and outcomes are fiction, not performance evidence.
Completion and the next project
Complete the learner demonstration in the Template Pack. Explain a diff, a source/site mismatch, a rejected candidate and the applicable recovery method in your own words. If you cannot explain them, repeat the exercise before selling that capability.
For the next client, reuse the workflow and authorized generic patterns, not private assets or records. Assess the new project independently. Keep actual effort, rework and feedback so later estimates become informed by your own evidence. Source control helps make delivery accountable; it is not a guarantee of demand, income or defect-free work.
References, license and publication record
All prompts and both cases are instructional examples. They have been edited for consistency, not executed as a timed learner study. Product references were consulted 28 October 2026 and linked at relevant passages. Interfaces, access and plan availability must be checked in the actual project.
Edition 1.0.0-draft.4 replaces the primarily narrative draft.3 with a twelve-step first-project path, bounded Codex prompts, common failures, a connected Foundation example and a separate Controlled escalation example. That draft remained unpublished. This release is compatible with the model and templates at Version 1.0.0.
Copyright © 2026 Henok Ahmed. The original educational text, tables, prompts, checklists and text templates in this publication family are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Give appropriate credit, link to the license and indicate changes, without implying endorsement. This grant excludes underlying project software, private files, third-party materials, trademarks, credentials and personal data. Those materials retain their respective rights.
Published with Henok Ahmed’s approval of edition 1.0.0-draft.4 as Version 1.0.0 on 28 October 2026. All prompts, fictional labels and instructional guidance are unchanged. The instructional prompts have not been validated through a timed learner trial. Publication does not authorize client deployment or the overall website’s stable public launch.
Source: https://henokahmed.com/source-control-practical-guide/
Publication qualification: instructional prompts have not been validated through a timed learner trial. Fictional examples and results remain illustrative.