indelible
Learning that survives the cold test.
indelible is a skill for Claude, ChatGPT Work and Codex that runs your self-study like a strict, organised tutor. It interviews you once, plans backwards from your deadline, puts sessions in your calendar, teaches from sheets you read and then close, drills you on paper, re-tests everything cold two days later, and keeps an honest record of what you actually know.
Status: v0.1.7 (early). Originally built for Claude Code, now adapted for ChatGPT Work and Codex. The Python tests cover the shared engine and workspace adapters; live model behaviour on each surface still needs testing.
The method
- Read it, then close it. Theory comes on a sheet you read and put away before the drills. An explanation left in view turns a test into a lookup.
- One thing at a time, then all mixed up. Teaching drills come in blocks of one question type. Anything that measures is unlabelled and interleaved, like the real exam.
- Nothing counts until it survives 48 hours. Every new skill comes back cold two days later, on fresh problems, then again a week later and every 3 weeks until your date. Same-day scores don't count.
- Check backwards, in writing. Beside every answer, write one check that runs the other way: put the answer back into the question, rebuild the total, or test the definition you used against what the question actually says. Re-solving in your head replays the same slip; a written check catches it. On a new topic, the check is the one the theory sheet showed you working. A mock copies the real exam, so it has no check column. Each sheet ends with a single line, Least sure I chose the right idea: ___. There is no confidence flag on every answer: people stop filling it in, and doubt tends to fire on arithmetic while the costly mistakes are confident misreadings of a definition, so the line asks about the idea, not a possible slip.
- Every miss goes in the error log and returns after 1 day, 3 days, 1 week and 3 weeks.
- You do the work. You get the smallest hint that unblocks you and never an answer you could reach yourself. When an answer is wrong, you're shown where, and the fix is yours: you're never sent to hunt for a mistake in something you learned an hour ago. Explaining a concept back is the real test.
Where it works
| Where | In v0.1 |
|---|---|
| Claude Code (terminal, desktop app, IDE extensions) | Supported. v0.1 is built for Claude Code. |
| Cowork | Should work when a folder on your computer is shared with it (your study folder goes inside that folder), but untested in v0.1. |
| ChatGPT Work and Codex with a lasting folder | File mode uses the same Python engine and study folder. New keyed sheets need an isolated agent builder; without it, use your own book's exercises and keep results unverified. |
| Ordinary ChatGPT, claude.ai chat and mobile | Manual mode: follow the portable instructions, save the complete record at the end and bring it to the next chat. Every result is unverified. A temporary sandbox or a chat upload alone does not provide automatic file tracking. |
File mode needs Python 3.9 or later where the assistant runs its commands. The runtime checks capabilities instead of assuming that a product name guarantees storage or tools.
Install in Claude Code
In Claude Code, add the plugin:
claude plugin marketplace add romanprigodskii/indelible
claude plugin install indelible@indelible
Or copy the skills/indelible folder from this repository into ~/.claude/skills/ (on Windows, %USERPROFILE%\.claude\skills\), so that you end up with ~/.claude/skills/indelible/SKILL.md.
Install in ChatGPT Work or Codex
The repository includes a portable plugin.json, an OpenAI marketplace at .agents/plugins/marketplace.json, and skill metadata in agents/openai.yaml. It keeps the Claude plugin packaging too.
Local skill (Codex-compatible hosts). Copy skills/indelible into ~/.agents/skills/indelible (on Windows, %USERPROFILE%\.agents\skills\indelible). Start a new chat in your lasting study folder and say:
Use $indelible to set up my study.
Plugin in the desktop app. Add the repository marketplace, using a Codex CLI that supports plugin marketplaces:
codex plugin marketplace add romanprigodskii/indelible
Alternatively, add your cloned repository with codex plugin marketplace add /absolute/path/to/indelible. Restart the desktop app, open Plugins, choose the Indelible marketplace and install Indelible. In a new ChatGPT Work conversation, select it with @ and ask to set up or start studying. Marketplace availability depends on the client and workspace. This repository is not automatically published to the public plugin directory. See the official packaging guide for the current installation routes.
Existing study folder. Open the same folder in the new host; no data conversion is needed. Ask Indelible to run instructions once, or run:
python3 /absolute/path/to/skills/indelible/scripts/indelible.py --workspace /absolute/path/to/Study instructions
This adds AGENTS.md routing beside the existing CLAUDE.md at the root and in live or paused subjects. Learner notes stay in CLAUDE.md for both assistants. Text outside the managed routing block is preserved, and a changed existing AGENTS.md gets a .bak copy. Do not run concurrent study sessions against the same folder from different assistants.
Use in ordinary ChatGPT
- Open the portable instructions, copy the file into a new chat or your ChatGPT project's instructions, and ask it to run Indelible. You can also attach the file and explicitly ask ChatGPT to follow it.
- Say, for example, "Help me prepare for my statistics final. I have 30 minutes today." It starts the short onboarding and diagnostic in your language.
- At the end, save the complete Markdown record it returns. Paste or attach the latest record when you resume. Ask "export my record" at any time.
Manual mode preserves the tutoring sequence, cold rechecks, mistake reviews and dated next actions. It cannot enforce sealed generated keys, maintain an automatic database, send timer alerts between messages or write your calendar. It uses your external exercises or short probes; all results carry [unverified]. No API key is required. Full automated state-file import/export remains deferred.
Requirements for file mode
- Python 3.9 or later. The scripts use the standard library only, so there is nothing to
pip install. - Optional, for PDF sheets: typst, or Chrome, Chromium or Edge. Without either, sheets come as HTML or Markdown. The skill checks what you have during setup.
One permission rule (optional). A session runs many small script calls. So that Claude Code doesn't ask about each one, you can allow this rule once with /permissions:
Bash(python3 *indelible.py*)
On Windows, the launcher is py -3, so the rule is Bash(py -3 *indelible.py*).
- What the rule allows. It is a text pattern, not a check on which file runs: it allows any
python3command that mentionsindelible.py, including one that runs other code first. The skill itself pre-approves nothing: until you add a rule, Claude Code asks before each script call. - A stricter rule, if you prefer one: in place of
python3 *indelible.py, write the start of the command exactly as Claude Code's permission prompt shows it, with the script's full path (and quotes, if the prompt has them), for exampleBash(python3 /full/path/to/skills/indelible/scripts/indelible.py*). If prompts still appear, compare the rule with the prompt. A plugin install keeps the script in a folder that can change when the plugin updates, so you may need to update the rule then. - You stay in charge. The skill only suggests the rule: you add it yourself, and nothing changes your settings for you. What the script reads and writes is listed under What it runs, writes and sends.
Example prompts
Say what you want in plain words. These examples come from the made-up learners in evals/fixtures/.
| You say | What happens |
|---|---|
| "I need a 7.5 in IELTS Academic by December for my master's offer. Can you set me up and keep me honest?" | Setup. Up to eight short questions, a two-week plan to check and a one-screen summary. Nothing is written until you say yes. Week 1 opens with a diagnostic: nothing is taught until it's marked. |
| "teach me rust I guess. please don't ask me a million questions" | Express setup. Three questions, then a plan; everything else takes a default you can change later. |
| "start ielts" or "What's due today?" | A session. The 2-day recheck first, then mistakes that are due, then new material and drills, then marking and the close. |
| "Here are the photos of this morning's drill sheet." | Marking. The photos are filed first, and only then is the answer key opened. For each miss you say how you got your answer; then you're shown the step where it went wrong, you make the fix, and the miss goes in the error log with a date to come back. |
| "I missed Thursday's session." | Reschedule. The plan is repaired and checked, the 2-day recheck is always rebooked, and you see the changes before anything goes to your calendar. If you say no, the sessions it added are cancelled and it asks what would work instead. |
| "Put my study blocks in my calendar." | Sync. A preview of what will be added, moved or cancelled, written only after your yes, then read back. Without a calendar connector you get an .ics file. |
| "Weekly review" | Review. The week's numbers, each labelled (such as measured or practice), ending in one to three decisions. |
| "Where am I?" | Status. What's due, your week and your numbers. It changes nothing. |
Try it with sample data
examples/sample-workspace/ is a study folder for a made-up learner, ready to use: a graded diagnostic, open mistakes and a planned week. Everything in it is synthetic, the answer keys included.
- Install the plugin (above). The sample isn't part of the installed plugin, so also clone or download this repository.
- Copy
examples/sample-workspacesomewhere else, because a session writes to the folder. - The sample has a fixed clock, Thursday 15 October 2026, 07:00 in Lisbon. In the terminal where you'll start Claude Code, set it first:
export INDELIBLE_NOW=2026-10-15T07:00+01:00(in PowerShell:$env:INDELIBLE_NOW = "2026-10-15T07:00+01:00"). Without it, the dates follow your computer's clock and don't line up with the sample's story. - Open Claude Code in the copy.
- Say "status" for a look that changes nothing, or "what's due?" to start a session.
examples/README.md has the exact commands, what the sample holds and what to expect.
Commands
You don't need to learn these: say what you want in plain words. In Claude Code you can also type /indelible followed by a command (/indelible:indelible when installed as a plugin). In Codex use $indelible; in ChatGPT use the installed plugin's @ menu or the portable instructions above.
| Command | Say something like | What it does |
|---|---|---|
teach | "set me up for…", "add chemistry" | A short interview: your goal, date, level, materials, session length, days and times, and calendar. A second subject gets a shorter re-run. |
session (the default) | "start", "let's go", "what's due", "I have 15 minutes" | A full study session: the 2-day recheck first, then mistakes that are due, then new material and drills, then marking. |
close | "done", "gotta go", "wrap up" | The close checklist: your work filed and marked, mistakes logged, rechecks booked and files updated before you leave. |
status | "where am I", "this week", "how am I doing", "what do you keep?" | A short look at your week, your last recheck and what's solid, in a few plain lines. It changes nothing in your plan or records. |
ask | "what does X mean?", "explain X" | A question about your subject outside a session. Claude asks what you remember first, never discusses a sheet you haven't handed in yet, and notes what it explained or sent you back to read, so your next 2-day recheck still measures what you remember. |
diagnose · mock | "test me properly", "full mock" | A measurement with no teaching: unlabelled, timed and scored. |
plan · reschedule | "plan my week", "I missed Thursday", "sick till Monday" | Builds or repairs your plan and checks it against your sleep and commitments, then shows it to you. Nothing goes to your calendar before your yes; if you say no, the sessions it added are cancelled. |
review | "weekly review", "review my week", "am I on track?" | The weekly review, with whether the plan still fits your date, ending in one to three decisions. |
sync | "put it in my calendar", "fix my calendar" | Shows the calendar changes, makes them only after your yes, then reads them back. |
migrate | "use my existing notes" | Imports a study system you've been running by hand, without losing anything. In v0.1 this is a guided procedure rather than a script command (see What it runs). |
A forget command, which would show exactly what it deletes and delete it only after your yes, is planned for v0.2 and isn't available yet.
How a week looks
An example with a made-up learner: preparing for IELTS Academic 7.5 by 12 December, Portuguese first language, 60-minute sessions before work on Monday, Tuesday and Thursday, and on Saturday mornings.
- Setup, about 8 minutes. She says "set me up for IELTS". Claude asks up to eight short questions, one at a time, each with a default she can take by saying "skip". Then it shows a two-week plan and asks whether the blocks should go in her calendar. Nothing is written until she says yes to a one-screen summary.
- Week 1 opens with the diagnostic: nothing is taught until it's marked. On Monday, part A: a mixed, unlabelled test on paper. A part B follows only for the topics where part A went well, and the results open with what she already knows. Every number carries a label, such as measured or practice. Teaching starts once the diagnostic is marked: with no part B needed, as in the sample, her first new skill comes on Tuesday; with a part B, later in the week.
A typical week from then on:
- Tuesday: a new skill. A theory sheet she reads and then closes. Drills on paper, in blocks of one question type, with a written check beside every answer and Least sure I chose the right idea: ___ at the end. She sends phone photos; they are filed before anything is marked. For each miss she says first how she got her answer; then she is shown the step where it went wrong, and the miss goes into the error log with a date to come back.
- Thursday: the 2-day recheck opens the session. Fresh questions on Tuesday's skill, cold, with nothing reviewed in the 24 hours before. Then the next skill.
- Every session ends with the close, inside the session. Anything she can't finish becomes a dated to-do rather than a vague "later".
- Calendar. Blocks appear in her calendar only after she has seen the changes and said yes. When she writes "I missed Thursday", the plan is repaired, and the 2-day recheck is always rebooked.
- Sunday is her rest day. At the end of the week, a short review ends in one to three decisions.
Your data
- Plain files in a folder you choose. Everything is kept as JSON, JSONL and Markdown files that you can open. Nothing is stored inside the skill's own folder.
- The scripts make no network calls. A browser they start to print PDF sheets runs with its network access blocked. Calendar events are written only after you've seen them and said yes, unless you chose to let the assistant move a missed session within 24 hours. Your conversation, uploaded records and files read by the assistant are processed by the provider of the host you use (OpenAI or Anthropic). For programming subjects, your own build tool may download your project's declared dependencies when your tests run. Details: PRIVACY.md.
- Answer keys stay hidden until you've attempted a sheet. Keys are sealed in a separate folder when a sheet is built, and they open only after your attempt (a photo or your typed answers) is filed.
- Deleting. Your record is plain files in one folder, and deleting that folder removes it all. Deleting just one subject's folder removes most of that subject, but in v0.1 some traces stay in shared files; PRIVACY.md explains. The scripts never delete your data on their own. A
forgetcommand that previews what it removes is planned for v0.2 and isn't available yet. - Optional local history. You can keep the folder under git; the
.gitignoreit comes with keeps answer keys, your answers, photos of your work and the inbox out of it.
What it runs, writes and sends
Everything the plugin's scripts do is readable source in this repository, and the plugin downloads and installs nothing while it runs. Beyond the scripts, Claude runs a few commands that the skill's instructions spell out; they are listed below too. In Claude Code's default permission mode, you are asked before any command that no permission rule allows.
What it runs
- Its own Python scripts:
skills/indelible/scripts/indelible.pyand thelib/folder beside it, standard library only. - Optional programs, only if they are already on your computer:
typst, to make PDF sheets.- Otherwise, an installed Chrome, Chromium or Edge in headless mode, to print a sheet to PDF. It runs with a throwaway profile that is deleted afterwards and a mock keychain, so it never touches your saved passwords or your system keychain. Its network access is blocked: its background services (updates, sync, safe-browsing lists) are switched off, all its traffic goes to a proxy address that doesn't exist, and no host name resolves. The page it prints is a local file with no external fonts, scripts or images. (When run as root on Linux, as in a container, it adds the
--no-sandboxflag that Chromium needs there.) sips(built into macOS) orheif-convert(elsewhere, if installed), to turn iPhone HEIC photos into JPEG.- When
doctorchecks your setup, it makes a one-page test PDF with typst and with the browser, if it finds them, in a temporary folder that it then deletes.
- An isolated helper agent writes each practice sheet and its answer key to files, so the answers never pass through your chat. The host must provide this capability; without it, use external exercises. If you choose local history at setup, the assistant runs
git initin your study folder and never adds a remote. - Programming subjects only: to mark your code, Claude copies your files to a temporary folder, adds the hidden tests and runs them with your own compiler or test tool (for example
cargo test) under a 60-second limit. When the helper agent builds a programming sheet, it runs its own reference solution and tests the same way, in a temporary folder of its own, to check the answer key. Your project is never changed. Your tool may download the dependencies your project declares, as it would if you ran it yourself. migrate(importing a study folder you've been running by hand): v0.1 has no script command for it. Claude runs two short Python programs written out inreferences/migrate.md: one copies the files you name into your study folder and records a SHA-256 hash of each, and one checks the copies against those hashes. To read a mistakes file that also holds the right answers, Claude writes a few lines of Python that print it with the answers replaced by[hidden]. Your original folder is only read, never changed.
What it writes
- Your study folder. Your record lives here, and the scripts refuse to put it inside the skill's own folder. They write files only inside it (the exceptions are listed below): for example, a calendar
.icsfile goes in itsplan/ics/folder, and the scripts refuse any other place. CLAUDE.mdfiles, one at the top of your study folder and one in each subject's folder. Claude Code reads aCLAUDE.mdautomatically, as instructions, whenever it is opened in that folder.- They tell Claude that the folder is a study folder, to use the indelible skill and start with its brief, and to keep out of the sealed answer keys. They also hold your own notes, which Claude adds from what you tell it.
- The scripts create one only where none exists. After that, they change only the part between the
indelible:beginandindelible:endmarkers, each time the summaries are refreshed. If your study folder already had aCLAUDE.md, your text stays as it is and that marked part is added at the end. - The brief reads your notes back from each subject's
CLAUDE.md. The scripts read no otherCLAUDE.md: nothing in~/.claude, and nothing in your other projects. (When you ask to import a study folder you've been running by hand, Claude reads that folder's ownCLAUDE.mdas part of the import.) - The top-level one names the full path of the script, so a Claude session without the skill can still run the brief. That path can go out of date when the plugin updates; the skill itself doesn't depend on it.
AGENTS.mdrouting files in the same folders, for Codex-compatible hosts. They point to the sharedCLAUDE.mdrather than copying learner notes. Only theindelible:agent:begin/indelible:agent:endblock is managed; existing text outside it stays intact.instructionsadds them to older workspaces. Legacy and shadow subjects are skipped.- A one-line pointer file,
~/.indelible/workspace, holding your study folder's path, so Claude can find it from any folder. Setup writes it (indelible.py init --pointer) after you say yes to the setup summary. You can delete it; Claude then finds your study folder when you open Claude in it. - Temporary files in your computer's temp folder:
- the browser's throwaway profile while it prints a sheet, and
doctor's test PDFs, both deleted straight away by the scripts; - for programming subjects, the copy of your code that Claude tests, which Claude deletes once the marks are recorded, and the helper agent's own test folder, which holds its reference solution (part of the answer key) and none of your data. v0.1's instructions don't yet tell it to delete that folder.
- the browser's throwaway profile while it prints a sheet, and
- Python's usual compiled-code cache (
__pycache__) beside the scripts. It holds none of your data. - Files you hand over (photos, PDFs, typed answers, or a project folder for programming practice) are read where they are and copied into your study folder. The originals are never changed or deleted.
- When you say "sent", Claude lists the names of photos and PDFs newer than the sheet in your study folder's inbox, your Downloads folder (where AirDrop puts them) and a photos folder you named, if any. It reads nothing else there, and files them only after you confirm which they are.
What it sends
- Nothing, from the scripts. They make no network requests: there is no
urllib,http,requestsor similar in them. The only use of Python'ssocketmodule is reading your computer's name for the study folder's write lock. A browser they start to print a PDF has its network access blocked (see What it runs). There are no analytics, no telemetry and no update checks, and nothing is sent to the author. - Your calendar, only through a calendar connector you have added to Claude yourself, and only if you choose it at setup. Claude shows you the changes and writes them only after your yes (or, if you choose that option, moves a missed session within 24 hours and tells you afterwards), then reads back what it wrote. With your yes, it may also read your busy times for the next 14 days to find free slots. Without a connector, you get a local
.icsfile to import yourself. - Your conversation with Claude is processed as usual, under your agreement with Anthropic.
What it never does
- It never edits Claude's settings or permissions, and it pre-approves no commands (it declares no
allowed-tools). If Claude Code asks before each script call, the skill suggests the permission rule under Install, for you to add yourself.
Honest limits
- A skill can't physically stop the model from breaking one of its own rules. So the rules that matter most don't rely on instructions alone. Three structural safeguards do the work:
- The only route to a printable sheet runs the sheet checker, which refuses sheets that break the method.
- The only route to an answer key requires your attempt to be filed first.
- The next session start detects anything the previous session skipped, and does it first.
- v0.1 is built for Claude Code (see Where it works). Full claude.ai and mobile support, hooks and two-way calendar sync are planned for v0.2 and aren't available yet. Until then, if you move a block in your calendar app, tell Claude so the plan can be updated.
- Claude writes most practice items, and it can get one wrong. If you think a mark is wrong, say so: the record (your photo and the key) is checked before the mark is kept or changed.
- The principles come from well-studied effects: retrieval practice (Roediger & Karpicke, 2006), spacing (Cepeda et al., 2006) and interleaving (Rohrer & Taylor, 2007); see also Dunlosky et al. (2013). The specific numbers, such as the 48-hour recheck and six questions per block, are defaults that were field-tested, not results of a controlled study. You can change them.
Troubleshooting
- "Python not found" or "too old". The scripts need Python 3.9 or later. Install it from python.org or your package manager, then ask Claude to check your setup again. On Windows, use the Python launcher,
py -3, which the python.org installer adds; ifpython3opens the Microsoft Store or isn't found, that's expected, and the skill usespy -3. Without Python, the skill offers a manual mode, but nothing then enforces its rules. - A permission prompt before every script call. Allow the rule under Install with
/permissions:Bash(python3 *indelible.py*), orBash(py -3 *indelible.py*)on Windows. Read what it allows first; Install also gives a stricter version. - No PDFs.
doctorlists each PDF tool it found and whether its test page worked. Without a working one, sheets come as HTML (read it on screen, or print it from a browser) or Markdown, and everything else works the same. For PDFs, install typst (or Chrome, Chromium or Edge), then ask Claude to check your setup again. - Windows:
doctorwarns that your time zone can't be loaded. Windows has no built-in time zone database, so times follow the computer's own zone instead of the one set for your study. Install the database withpy -3 -m pip install tzdata(you run this yourself; it downloads thetzdatapackage from PyPI). - Your calendar isn't detected. Claude only looks at the connectors you have added. Connect your calendar in Claude's settings (under Connectors), then say "sync my calendar". Or choose the
.icsfile and import it into your calendar app. - A session was left open. Nothing is lost. The next session start closes it first (in 10 minutes at most), logs the close as late, then carries on.
- "No indelible workspace found." Open Claude in your study folder (for example
~/Study) and try again. - Anything else: open an issue in this repository's Issues tab with what you said, what happened, your operating system and Python version, and
doctor's output if it's relevant. The repository is public, so leave out your study files and personal details.
Support and security
- Help and bug reports: the Issues tab of this repository on GitHub.
- Security problems: please report them privately, as described in SECURITY.md.
- Privacy: PRIVACY.md explains what data exists, where it lives and how to delete it.
indelible is an independent open-source project. It is not made, endorsed or supported by Anthropic or OpenAI.
Contributing
- Tests:
python3 -m unittest discover -s tests(Python 3.9+, standard library only). CI runs them on macOS, Linux and Windows. - Privacy rule: every example, fixture, test and issue uses only the synthetic learners A–D in
evals/fixtures/. Never add real learner data, even your own. Once per clone, rungit config core.hooksPath dev/hooks, sodev/privacy_grep.pyruns before every push and stops it on any hit (it needs a local, git-ignored denylist; it also checks commit author names and emails). - Build contract: file formats, commands and checker rules are defined in
dev/CONTRACT.md. Where code and the contract disagree, open an issue. - Trigger evals:
evals/trigger.jsonlists prompts that should and shouldn't start the skill, near misses included (a one-off quiz or explanation with an exam ahead). Running them needs a model, for example the skill-creator skill's description evals, so CI doesn't.tests/test_skill_meta.pychecks that the set is well formed and that the SKILL.md description stays within its 1,024 characters. - Instruction budget: a study session loads SKILL.md and five references (
session-open,session-grade,session-teach,sheetsandclose), and pays for them again after every compaction.tests/test_skill_meta.pycaps them at 110,000 characters together: material only the builder reads goes inassets/prompts/builder.md, and each rule is said once, in the file that owns it, with a pointer from the others.
License
Dual-licensed under MIT or Apache-2.0, at your option.
Designed and field-tested by @romanprigodskii.